Cytokine-driven immunogenicity prediction: integrating HLA binding and cytokine induction Authors: Alexander Gavrilenko, Maria Sindeeva, Tatiana Shashkova Published: 2026 ABSTRACT AbstractCytokines are central to the immune response, determining whether T-cell activation leads to protective immunity or tolerance. The secretion of key cytokines such as IFN-\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $\gamma$\end{document}, IL-2, IL-4, and IL-10 shapes the immune response, influencing its presence, pathways, and magnitude. While most existing computational methods for predicting peptide immunogenicity typically focus on either cytokine induction or histocompatibility complex binding affinity separately, they fail to capture the full complexity of immune activation. In this study, we introduce a two-stage computational framework that integrates both cytokine induction and HLA binding affinity predictions to provide a more comprehensive prediction of peptide immunogenicity. First, we develop machine learning models EpiLAMA-IL to predict cytokine secretion based on peptide and parent protein sequences. These models outperform current state-of-the-art cytokine release prediction methods, achieving ROC AUC from 0.79 up to 0.92 depending on the task. These cytokine scores are then combined with HLA binding affinity predictions and peptide physico-chemical descriptors to develop a composite immunogenicity score prediction model EpiLAMA-IM. This approach bridges the gap between sequence-based antigen presentation and functional immune outcomes, offering improved accuracy in predicting immune responses and achieves MCC 0.47 on the CD4Episcore dataset. Introduction The ability of peptides to elicit immune responses is a cornerstone of modern immunology, with direct applications in vaccine development, cancer immunotherapy, and immune safety evaluation [1, 2]. Traditionally, peptide immunogenicity has been approximated by its capacity to bind major histocompatibility complex (MHC) molecules and to be recognized by T-cell receptors (TCRs) [3–5]. However, antigen presentation alone does not fully explain the variability in immune outcomes: peptides with similar HLA binding affinities can induce markedly different immune responses—ranging from strong pro-inflammatory activation to immune tolerance [6, 7]. This functional heterogeneity arises largely from the cytokine milieu that shapes T-cell differentiation and effector functions [8–10]. Cytokines play a central role in determining both the magnitude and the quality of immune responses [11]. In MHC class II–restricted pathways, cytokines secreted by CD4\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $^{+}$\end{document} T-helper cells direct the polarization of immune responses into Th1, Th2, Th17, or Treg subsets [9, 12]. For instance, IFN-\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $\gamma$\end{document} and IL-2 drive Th1-type cellular immunity, while IL-10 and TGF-\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $\beta$\end{document} promote regulatory or suppressive effects [13–15]. Consequently, the cytokine induction potential of a peptide provides a functional readout of its immunogenicity, reflecting not only antigen presentation but also the qualitative nature of the immune response it triggers. Because experimental profiling of cytokine secretion is labor-intensive and context-dependent, several in silico approaches have been developed to predict cytokine-inducing peptides using machine learning. Tools such as IL2Pred [16], IL4Pred [17], IL10Pred [18], Meta-IL4 [19], ILeukin10Pred [20], and IFNepitope2 [21] employ supervised learning to identify peptides capable of inducing specific cytokines. All of these models leverage sequence-derived features or motif compositions to discriminate between cytokine-inducing and non-inducing sequences. Although each model focuses on a single cytokine, together they demonstrate that cytokine induction potential is an intrinsic property that can be captured from peptide sequence patterns. At the same time, several methods have been developed to directly predict overall peptide immunogenicity, each focusing on different aspects of this complex notion. For example, VaxiJen [22] predicts peptide’s antigenicity (antigen-antibody binding propensity) based on physico-chemical properties of its amino acid sequence, without considering HLA binding or cytokine release. CD4Episcore [6] integrates predicted MHC class II binding affinity with a separate immunogenicity score trained on cytokine release and tetramer staining assays to identify CD4\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $^{+}$\end{document} T-cell epitopes, and validates it on T-cell proliferation assay datasets. Other methods define immunogenicity as IFN-\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $\gamma$\end{document} release, disregarding other immune response pathways. Deep learning–based methods such as ImmuScope-IM [23] and the approach by Omran et al. [24] learn to model peptide–HLA interactions to predict immunogenicity. Although all of these tools claim to predict immunogenicity, they define it differently, making direct comparison difficult. Additionally, these models typically overlook cytokine-level diversity and the balance between pro- and anti-inflammatory signals, which are essential determinants of functional immunogenicity. In this study, we introduce a two-stage computational framework that explicitly models cytokine induction as an intermediate step in the prediction of overall immunogenicity (Fig. 1). In the first stage, we train individual machine learning models for multiple key cytokines (e.g. IFN-\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $\gamma$\end{document}, IL-2, IL-10, IL-4) using curated datasets of experimentally validated peptide responses from the Immune Epitope Database (IEDB). Each model outputs a probability score representing the likelihood that a peptide induces a given cytokine. Our models demonstrate comparable or superior performance to existing solutions across all benchmarks. Overview of the two-stage EpiLAMA framework: individual machine learning models EpiLAMA-IL* predict cytokine induction probabilities (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $P_{IL2}$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $P_{IL4}$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $P_{IL10}$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $P_{IFN-\gamma }$\end{document}) which are then combined with MHC-II binding affinities and physico-chemical descriptors in a meta-model EpiLAMA-IM, which outputs a composite immunogenicity score. In the second stage, these predicted cytokine scores are combined with HLA binding affinity predictions for multiple MHC-II alleles derived from the MHCNuggets [4] tool, as well as peptide’s physico-chemical properties, to generate a composite immunogenicity score. This meta-model integrates both antigen presentation and functional polarization features, providing a biologically interpretable and scalable estimate of peptide immunogenicity. Our model, EpiLAMA-IM, outperforms alternative solutions and achieves an MCC value of 0.466 on the CD4Episcore test set. We make our models available on GitHub at https://github.com/AIRI-Institute/EpiLAMA. By bridging sequence-based cytokine prediction and global immune outcome modeling, our approach offers a more comprehensive understanding of the determinants of peptide immunogenicity. Such integrative in silico models can support rational vaccine design, improve epitope prioritization, and facilitate immune risk assessment in therapeutic applications. Materials and methods Cytokine datasets All peptide–cytokine interaction records were retrieved from the IEDB (accessed February 2025), the largest public repository of experimentally validated epitopes. We extracted CD4+ T-cell epitopes (i.e. MHC class II–restricted peptides) that had been experimentally tested for their ability to induce secretion of at least one of the cytokines IL-2, IL-4, IL-10, or IFN-\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $\gamma$\end{document} in the host T-cells. Each record in the resulting dataset includes metadata such as the source antigen from which the peptide was derived, the start and end positions of the peptide within the antigen, the host organism in which the immune response was measured, the experimental assay type used to assess cytokine secretion, the measured cytokine production (positive/negative call), and specific MHC allele when reported (e.g. HLA-DRB1*01:01). In this study, we considered human-specific and mixed host datasets. To ensure data consistency and relevance for modeling, the following preprocessing steps were applied: Deduplication. Duplicate entries for the same peptide–cytokine pair were removed to prevent overrepresentation. Deduplication. Duplicate entries for the same peptide–cytokine pair were removed to prevent overrepresentation. Filtering. Only peptides presented by MHC class II alleles (HLA-DP, -DQ, -DR) were retained. Peptides shorter than 13 or longer than 25 amino acids were discarded, reflecting the length range relevant for MHC-II presentation [25]. Only peptides tested in Homo sapiens hosts were included when developing the human-specific dataset. The mixed-host dataset comprised peptides tested in Homo sapiens, Mus musculus (different species), Pan troglodytes, Gallus gallus, and other available in IEDB hosts. Filtering. Only peptides presented by MHC class II alleles (HLA-DP, -DQ, -DR) were retained. Peptides shorter than 13 or longer than 25 amino acids were discarded, reflecting the length range relevant for MHC-II presentation [25]. Only peptides tested in Homo sapiens hosts were included when developing the human-specific dataset. The mixed-host dataset comprised peptides tested in Homo sapiens, Mus musculus (different species), Pan troglodytes, Gallus gallus, and other available in IEDB hosts. Conflict resolution. Peptides annotated as both inducers and non-inducers of the same cytokine were excluded. Conflict resolution. Peptides annotated as both inducers and non-inducers of the same cytokine were excluded. Peptide sequence mapping. Full parent antigen sequences were retrieved from UniProt and NCBI databases, and each peptide was verified to match the specified start and end positions within the corresponding protein sequence. Peptide sequence mapping. Full parent antigen sequences were retrieved from UniProt and NCBI databases, and each peptide was verified to match the specified start and end positions within the corresponding protein sequence. The final number of samples for each datasets is presented in Table 1. Table 1Sample distribution for each dataset.NamePositiveNegativeTotalHostIL-2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $_{H}$\end{document}7207861507HumanIL-4\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $_{H}$\end{document}415464880HumanIL-10\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $_{H}$\end{document}2997421042HumanIFN-\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $\gamma _{H}$\end{document}10 88427 11237 997HumanIL-2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $_{M}$\end{document}291934066325MixedIL-4\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $_{M}$\end{document}110813972505MixedIL-10\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $_{M}$\end{document}6549421596MixedIFN-\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $\gamma _{M}$\end{document}15 04640 48155 527MixedProIMM-train\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $_{H}$\end{document}400334187421HumanProIMM-test\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $_{H}$\end{document}18899452834HumanCD4Episcore-train108510 26411 349MixedCD4Episcore-test52316832206Mixed Sample distribution for each dataset. Immunogenicity dataset To characterize the presence or absence of an immune response for each peptide, we chose to use CD4+ cell proliferation assay binary response. While IFN-\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $\gamma$\end{document} is sometimes employed as a measure of immunogenicity [23, 24], not all CD4+ T-cells produce IFN-\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $\gamma$\end{document} as a part of the immune response. While CD4+ cells differentiated into Th1 lineage cells do, adopting IFN-\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $\gamma$\end{document} as a proxy for immunogenicity would miss activation of several other Th cell lineages, such as Th2 or Th17 cells [12]. CD4+ cell proliferation assays, together with cytokine release and tetramer binding assays, have been used in [6] to construct a validation dataset for immunogenicity prediction. However, we aim to avoid mixing multiple types of assays in our datasets to keep predictions easier to interpret: in our work, immunogenicity means CD4+ cell proliferation. We gathered a dataset of 11 225 peptides with experimentally measured cell proliferation assays from the IEDB [26] on 4 September 2025 (exact assay types and search criteria are available in the Supplementary Data). To consider a peptide a positive sample, previous studies [6] have applied study-specific thresholds: some peptides are considered positive if they have at least one positive response, while others may require at least 50 % positive responses. We applied a stricter and more uniform threshold and considered peptides with \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $> 90\%$\end{document} positive responses as positive, and peptides with \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $< 10\%$\end{document} positive responses as negative. We filtered out peptides without information about their parent protein, and retrieved the parent protein sequence for each remaining peptide from UniProt [27] and RefSeq NCBI [28] databases. We additionally excluded any peptides overlapping between our proliferation dataset and the CD4Episcore [6] test dataset to avoid data leakage in future model evaluation. To avoid data leakage between different peptides from the same parent protein, we performed a 65/35 stratified group split of parent proteins into train and test sets. We also employed the CD4Episcore [6] training and test datasets for model training and evaluation. The CD4Episcore training dataset combines experimentally obtained cytokine release assays, positive experimental tetramer staining assays, and synthetic negatives from proteins with positive tetramer staining assays from human, mouse, and macaque hosts. The CD4Episcore test dataset is a combination of experimentally obtained T-cell proliferation, cytokine release, and activation assays from human and mouse hosts. The peptides present in these datasets were matched back to their corresponding IEDB records, and for each peptide, its parent protein sequence was retrieved. Peptides without matches to any IEDB records or without parent protein information were excluded from the datasets. The final dataset statistics for the proliferation dataset (ProIMM\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $_{H}$\end{document}) and the CD4Episcore dataset are presented in Table 1. Peptide representations In this section, we describe all peptide representations that were considered for cytokine release prediction models. Classical sequence descriptors We utilized three classical sequence descriptors: Amino Acid Composition (AAC), Dipeptide Composition (DPC), and Amino Acid Pair Composition (AAP). These are extensively applied in sequence-based proteomics tasks due to their simplicity, interpretability, and effectiveness in capturing sequence-level statistical characteristics [16, 29, 30]: Amino Acid Composition (AAC): A 20D vector capturing the frequency of each standard amino acid in the peptide. The \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $i$\end{document}th component is given by (1)\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} \begin{eqnarray*}& \mathrm{AAC}_{i} = \frac{n_{i}}{N},\end{eqnarray*}\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $n_{i}$\end{document} is the number of residues of type \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $i$\end{document}, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $N$\end{document} is the total peptide length. Amino Acid Composition (AAC): A 20D vector capturing the frequency of each standard amino acid in the peptide. The \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $i$\end{document}th component is given by (1)\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} \begin{eqnarray*}& \mathrm{AAC}_{i} = \frac{n_{i}}{N},\end{eqnarray*}\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $n_{i}$\end{document} is the number of residues of type \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $i$\end{document}, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $N$\end{document} is the total peptide length. Dipeptide Composition (DPC): A 400D vector encoding the relative frequencies of all possible adjacent amino acid pairs. The composition of the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $ij$\end{document}th dipeptide is computed as (2)\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} \begin{eqnarray*}& \mathrm{DPC}_{ij} = \frac{n_{ij}}{N - 1},\end{eqnarray*}\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $n_{ij}$\end{document} is the number of times the dipeptide \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $(i,j)$\end{document} appears, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $N - 1$\end{document} is the total number of adjacent residue pairs. Dipeptide Composition (DPC): A 400D vector encoding the relative frequencies of all possible adjacent amino acid pairs. The composition of the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $ij$\end{document}th dipeptide is computed as (2)\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} \begin{eqnarray*}& \mathrm{DPC}_{ij} = \frac{n_{ij}}{N - 1},\end{eqnarray*}\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $n_{ij}$\end{document} is the number of times the dipeptide \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $(i,j)$\end{document} appears, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $N - 1$\end{document} is the total number of adjacent residue pairs. Amino Acid Pair Antigenicity Index (AAP): This descriptor applies a log-odds scale to distinguish cytokine-inducing peptides from non-inducers, based on dipeptide enrichment: (3)\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} \begin{eqnarray*}& \mathrm{AAP}_{ij} = \left( \frac{n_{ij}}{N - 1} \right) \log \left( \frac{n_{ij}^{+}}{n_{ij}^{-}} \right),\end{eqnarray*}\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $n_{ij}^{+}$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $n_{ij}^{-}$\end{document} are the frequencies of dipeptide \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $(i,j)$\end{document} in the positive and negative training sets, respectively, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $N - 1$\end{document} is the total number of adjacent residue pairs. Amino Acid Pair Antigenicity Index (AAP): This descriptor applies a log-odds scale to distinguish cytokine-inducing peptides from non-inducers, based on dipeptide enrichment: (3)\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} \begin{eqnarray*}& \mathrm{AAP}_{ij} = \left( \frac{n_{ij}}{N - 1} \right) \log \left( \frac{n_{ij}^{+}}{n_{ij}^{-}} \right),\end{eqnarray*}\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $n_{ij}^{+}$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $n_{ij}^{-}$\end{document} are the frequencies of dipeptide \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $(i,j)$\end{document} in the positive and negative training sets, respectively, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $N - 1$\end{document} is the total number of adjacent residue pairs. Protein language model embeddings We obtained peptide embeddings in two ways: (i) by encoding the full parent antigen sequence or (ii) by encoding only the peptide sequence using the ESM C [31] or Ankh [32] protein language models. We extracted the per-residue embeddings for each peptide from each model and took a position-wise average to create a fixed-length representation. For antigens longer than the maximum context of the protein language model (2048 residues for ESM C, 1024 for Ankh), we built a sequence window around each peptide. We extended this window symmetrically upstream and downstream and then trimmed it to the allowed length. When the peptide was close to the N- or C-terminus and a symmetric window did not fit, we used all residues available on the terminal side and truncated only from the opposite side. To provide an external reference point for sequence-based peptide classification, we also considered PepDoRA, a recently proposed deep learning framework for general peptide property prediction [33]. Although originally developed for broad biochemical property prediction (e.g. solubility, aggregation, antimicrobial activity), its architecture (built upon multi-scale convolutional encoders and attention-based sequence summarization) captures local residue patterns and global contextual dependencies that are also relevant for distinguishing cytokine-inducing epitopes. We used PepDoRA as a baseline peptide encoder, generating fixed-length embeddings from raw amino acid sequences, which are subsequently fed into the classifier. This allowed us to evaluate whether a general-purpose peptide representation model can serve as a competitive encoder for cytokine-specific peptide classification tasks. Cytokine induction prediction model We used the LightAutoML [34] framework to train separate binary classifiers for each cytokine. The framework constructs an ensemble of gradient-boosted decision trees (LightGBM, CatBoost) and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $\ell _{2}$\end{document}-regularized logistic regression models, with automated hyperparameter optimization via Optuna (100 trials per task). The final prediction is computed by averaging the out-of-fold outputs of all learners. To address the issue of imbalanced class distributions (e.g. positive samples comprising 30% of the IL-10 dataset), we applied synthetic oversampling techniques to augment only the training sets for each cytokine. Specifically, we used SMOTE [35], which generates synthetic minority-class samples by interpolation within minority-class neighborhoods, and ADASYN [36], which adaptively generates synthetic samples with a greater focus on minority-class instances that are harder to classify, thus enhancing model robustness. To assess model performance, we conducted five-fold stratified cross-validation on both the human dataset and the mixed-host datasets (Table 1), reporting the mean and standard deviation of evaluation metrics. The final models for Epitope prediction using LAMA-based InterLeukin secretion modeling were termed EpiLAMA-IL2, EpiLAMA-IL4, EpiLAMA-IL10, and EpiLAMA-IFN\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $\gamma$\end{document}. Cytokine induction models evaluation To ensure fair benchmarking with previous methods, we additionally utilized mixed-hosts datasets from prior studies [16–18, 20, 21]. For models that do not provide code for retraining we used their datasets to train and test our model. Otherwise, we retrained external models on our datasets. More specifically, for comparison with IL2Pred, IL4Pred, and IL10Pred we used their original mixed-host datasets from the corresponding articles to train and test our models. For IFNepitope2, we used their original human training dataset and an independent dataset 1 released by the authors. Additionally, for all of these datasets we retrieved parent protein sequences for each peptide. The authors of ILeukin10Pred released their dataset as feature matrices rather than raw sequences. To relate these data back to the corresponding IEDB records, we matched the provided dipeptide (”digram”) presence signatures back to candidate peptides. To avoid ambiguous mappings, we retained only one-to-one matches, i.e. if two or more peptides shared an identical binary digram vector, we removed those entries. After this unambiguous mapping and standard processing (see Section Cytokine datasets), the dataset comprised 311 positives and 836 negatives. Models like Meta-IL4, Omran et al. [24] and ImmuScope-IM were retrained on our datasets. However, Omran et al. [24] and ImmuScope-IM models also require information of the presenting MHC molecule as an input, so we modified our human and mixed-hosts IFN-\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $\gamma$\end{document} datasets (Table 1). We retained only entries annotated with HLA alleles supported by ImmuScope-IM for which a pseudo MHC sequence was available. The resulting datasets contained 2661 positive and 1786 negative samples for human-specific dataset, and 3417 positive and 2572 negative for mixed-host dataset. Immunogenicity prediction model We built the immunogenicity score as a meta-score that utilizes predictions of multiple characteristics to predict the overall immunogenicity of a peptide, similar to [6]. Below we provide detailed descriptions of scores that were considered for the final model: Interleukine release scores. These are the scores predicted by our EpiLAMA-IL2, EpiLAMA-IL4, EpiLAMA-IL10, and EpiLAMA-IFN\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $\gamma$\end{document} models, i.e. values between 0 and 1, with higher values corresponding to higher probability of interleukine release. Interleukine release scores. These are the scores predicted by our EpiLAMA-IL2, EpiLAMA-IL4, EpiLAMA-IL10, and EpiLAMA-IFN\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $\gamma$\end{document} models, i.e. values between 0 and 1, with higher values corresponding to higher probability of interleukine release. MHCNuggets binding affinity scores. For each peptide we ran MHCNuggets to obtain predictions of binding affinity values (IC50) for all supported human alleles (see Supplementary Data). Since MHCNuggets predicts IC50 directly, we applied the following transformation to them: \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $score = clip(1 - \mathrm{IC50} / 5000, 0, 1)$\end{document}. This means that IC50 of 5000 or more are considered non-binders and receive a score of 0, and the rest receive a score between 0 and 1, with higher values corresponding to higher binding affinity. MHCNuggets binding affinity scores. For each peptide we ran MHCNuggets to obtain predictions of binding affinity values (IC50) for all supported human alleles (see Supplementary Data). Since MHCNuggets predicts IC50 directly, we applied the following transformation to them: \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $score = clip(1 - \mathrm{IC50} / 5000, 0, 1)$\end{document}. This means that IC50 of 5000 or more are considered non-binders and receive a score of 0, and the rest receive a score between 0 and 1, with higher values corresponding to higher binding affinity. DeepPeptide peptide cleavage probabilities. For each parent protein in the dataset we used DeepPeptide to obtain per-residue predictions of probability of residue belonging to a peptide, a propeptide, or neither. Then, for each peptide in the dataset we calculated mean peptide probability for residues comprising this peptide, as well as mean propeptide probability for five residues preceding the peptide (if present). DeepPeptide peptide cleavage probabilities. For each parent protein in the dataset we used DeepPeptide to obtain per-residue predictions of probability of residue belonging to a peptide, a propeptide, or neither. Then, for each peptide in the dataset we calculated mean peptide probability for residues comprising this peptide, as well as mean propeptide probability for five residues preceding the peptide (if present). Chemical descriptors. We calculated the following physico-chemical peptide sequence features using Biopython [37] and peptides.py [38] Python modules: molecular weight, aromaticity, secondary structure fraction (helix, turn, sheet), isoelectric point, grand average of hydropathy (GRAVY), molar extinction coefficient (reduced), instability index, Shannon entropy, hydrophobic moment, Boman Index, MS-WHIM. Chemical descriptors. We calculated the following physico-chemical peptide sequence features using Biopython [37] and peptides.py [38] Python modules: molecular weight, aromaticity, secondary structure fraction (helix, turn, sheet), isoelectric point, grand average of hydropathy (GRAVY), molar extinction coefficient (reduced), instability index, Shannon entropy, hydrophobic moment, Boman Index, MS-WHIM. As with the cytokine release prediction models, we trained a LightAutoML ensemble of models on a binary classification task. We split the training dataset into five cross-validation protein-level folds for LightAutoML’s ensemble training, and report evaluation metrics on a held-out test set. The final model for Epitope prediction based on LAMA-assisted Immunogenicity Modeling was named EpiLAMA-IM. Evaluation metrics We used Precision, Recall, F1-score, Matthews Correlation Coefficient (MCC), Area Under the ROC Curve (AUC), as well as Area Under the PR Curve (PR AUC) to comprehensively evaluate model performance. Precision reflects the proportion of true positives among all predicted positives, crucial for reducing false positives in experimental validation. Recall measures the ability to detect actual positives, important for avoiding missed true epitopes. F1 score reflects model performance in both precision and recall. MCC provides a balanced metric well-suited for imbalanced classes by considering all confusion matrix components. AUC summarizes the model’s overall ability to discriminate between classes independently of classification thresholds. Similarly to AUC, PR AUC reflects a model’s overall ability to discriminate between classes, but is better suited for tasks with significant class imbalance. Results Sequence variability across cytokine classes To quantify the variability of epitope sequences associated with different cytokine classes, we computed three complementary metrics: Dipeptide density, measuring the number of distinct dipeptides observed in each epitope (maximum possible = 400). Dipeptide density, measuring the number of distinct dipeptides observed in each epitope (maximum possible = 400). Shannon entropy, capturing the diversity of amino acid compositions within sequences. Shannon entropy, capturing the diversity of amino acid compositions within sequences. Edit distance, which computes the average number of edit operations (insertions, deletions, or substitutions) required to transform one epitope into another—reflecting pairwise sequence dissimilarity. Edit distance, which computes the average number of edit operations (insertions, deletions, or substitutions) required to transform one epitope into another—reflecting pairwise sequence dissimilarity. These metrics collectively characterize the internal sequence heterogeneity of each cytokine-inducing epitope group. Among the three cytokine classes, IL-4–inducing epitopes exhibit the greatest sequence diversity, as reflected by their higher average dipeptide density, Shannon entropy, and edit distance (see Table 2). The variability of these metrics within each cytokine class is comparable, while the mean values for IL-4 are clearly higher. This indicates that sequences IL-4 epitopes are intrinsically more diverse, and that this effect is driven by a shift in the mean rather than by increased dispersion. We hypothesize that this indicates that IL-4 prediction is a more challenging task than other types of cytokines. Table 2Sequence variability metrics for cytokine-inducing epitopes.CytokineDensityEntropyEdit distanceIL-2 \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $14.25 \pm 2.30$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $3.78 \pm 0.29$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $14.83 \pm 2.09$\end{document} IL-4 \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $\mathbf{15.96} \pm 2.74$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $\mathbf{3.94} \pm 0.28$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $\mathbf{16.82} \pm 2.09$\end{document} IL-10 \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $14.07 \pm 2.10$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $3.77 \pm 0.24$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $14.55 \pm 1.96$\end{document} IFN-\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $\gamma$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $14.22 \pm 2.15$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $3.78 \pm 0.24$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $14.00 \pm 1.78$\end{document} Sequence variability metrics for cytokine-inducing epitopes. Classification of cytokine-inducing peptides We systematically evaluated two categories of feature representations for the classification of cytokine-inducing peptides: embeddings from pretrained language models and classical sequence-based descriptors. To explore the utility of learned representations, we considered three ways of peptide embedding extraction: (i) ESM C [31] and Ankh [32] embeddings of peptides within antigenic context, i.e. peptide residue representations extracted from whole antigen embeddings. This approach retains global structural and contextual cues, and we call it Antigen extracted. (ii) Despite ESM C and Ankh being pretrained on full-length proteins, we obtained these models’ embeddings of peptide-only sequences, discarding the surrounding antigenic context. This approach retains local structural information, but does not preserve any global structural or contextual cues. We called this approach Peptide extracted. (iii) We obtained embeddings of peptide sequences without antigenic context using a peptide-specific PepDoRa model. These embeddings were fed into the classifier models and evaluated on corresponding human data sets (Table 3). For each cytokine, the Antigen extracted configuration consistently outperforms both peptide-only approaches in all evaluated metrics (Table 3). The improvement is especially pronounced for IL-10 and IFN-\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $\gamma$\end{document}, where the MCC is higher by roughly 0.21–0.38 compared with peptide-level representations, and ROC AUC exceeds 0.94, reflecting a more confident and accurate classification of peptides. For IL-2 and IL-4 the gains are more moderate but still substantial, driven by simultaneous increases in both precision and recall. PepDoRA, despite being peptide-specific model, remains close to or slightly below the ESM C Peptide extracted model for all tasks, yet in all cases it falls far behind the Antigen extracted ESM C models. Table 3Ablation study results on cytokine datasets (IL-2, IL-4, IL-10, and IFN-\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $\gamma$\end{document}) for Antigen extracted (Source: Antigen) and Peptide extracted (Source: Peptide) models. Values are reported as mean \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $\pm$\end{document} standard deviation.CytokineEncoderSourcePrecisionRecallF1AUCMCCIL-2ESM CPeptide \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $0.609 \pm 0.042$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $0.408 \pm 0.077$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $0.484 \pm 0.056$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $0.672 \pm 0.025$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $0.217 \pm 0.059$\end{document} ESM CAntigen \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $\mathbf{0.788} \pm 0.015$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $\mathbf{0.830} \pm 0.014$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $\mathbf{0.808} \pm 0.008$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $\mathbf{0.863} \pm 0.014$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $\mathbf{0.596} \pm 0.028$\end{document} AnkhPeptide \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $0.596 \pm 0.055$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $0.404 \pm 0.061$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $0.479 \pm 0.052$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $0.630 \pm 0.028$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $0.185 \pm 0.070$\end{document} AnkhAntigen \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $0.781 \pm 0.022$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $0.799 \pm 0.017$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $0.790 \pm 0.015$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $0.845 \pm 0.013$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $0.564 \pm 0.036$\end{document} PepDoRAPeptide \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $0.584 \pm 0.046$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $0.378 \pm 0.060$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $0.455 \pm 0.038$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $0.632 \pm 0.014$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $0.179 \pm 0.035$\end{document} IL-4ESM CPeptide \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $0.650 \pm 0.048$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $0.624 \pm 0.079$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $0.636 \pm 0.061$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $0.740 \pm 0.031$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $0.339 \pm 0.076$\end{document} ESM CAntigen \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $\mathbf{0.762} \pm 0.079$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $\mathbf{0.824} \pm 0.030$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $\mathbf{0.789} \pm 0.038$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $\mathbf{0.843} \pm 0.023$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $\mathbf{0.545} \pm 0.067$\end{document} AnkhPeptide \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $0.668 \pm 0.040$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $0.645 \pm 0.038$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $0.656 \pm 0.037$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $0.760 \pm 0.041$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $0.358 \pm 0.070$\end{document} AnkhAntigen \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $0.735 \pm 0.083$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $0.798 \pm 0.048$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $0.761 \pm 0.031$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $0.829 \pm 0.021$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $0.486 \pm 0.051$\end{document} PepDoRAPeptide \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $0.654 \pm 0.053$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $0.602 \pm 0.067$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $0.626 \pm 0.053$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $0.737 \pm 0.037$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $0.332 \pm 0.078$\end{document} IL-10ESM CPeptide \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $0.705 \pm 0.055$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $0.326 \pm 0.062$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $0.443 \pm 0.064$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $0.834 \pm 0.020$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $0.364 \pm 0.065$\end{document} ESM CAntigen \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $0.780 \pm 0.046$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $\mathbf{0.798} \pm 0.070$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $\mathbf{0.786} \pm 0.024$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $\mathbf{0.935} \pm 0.014$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $\mathbf{0.696} \pm 0.028$\end{document} AnkhPeptide \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $0.645 \pm 0.176$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $0.094 \pm 0.044$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $0.163 \pm 0.070$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $0.740 \pm 0.029$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $0.161 \pm 0.085$\end{document} AnkhAntigen \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $\mathbf{0.801} \pm 0.066$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $0.710 \pm 0.083$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $0.749 \pm 0.042$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $0.933 \pm 0.018$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $0.656 \pm 0.055$\end{document} PepDoRAPeptide \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $0.674 \pm 0.130$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $0.272 \pm 0.150$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $0.371 \pm 0.194$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $0.790 \pm 0.038$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $0.313 \pm 0.159$\end{document} IFN-\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $\gamma$\end{document}ESM CPeptide \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $0.714 \pm 0.009$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $0.474 \pm 0.010$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $0.570 \pm 0.007$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $0.847 \pm 0.004$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $0.467 \pm 0.007$\end{document} ESM CAntigen \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $\mathbf{0.771} \pm 0.008$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $\mathbf{0.756} \pm 0.004$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $\mathbf{0.763} \pm 0.005$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $\mathbf{0.936} \pm 0.002$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $\mathbf{0.677} \pm 0.006$\end{document} AnkhPeptide \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $0.658 \pm 0.013$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $0.314 \pm 0.010$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $0.424 \pm 0.010$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $0.790 \pm 0.005$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $0.331 \pm 0.010$\end{document} AnkhAntigen \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $0.738 \pm 0.007$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $0.662 \pm 0.009$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $0.698 \pm 0.008$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $0.912 \pm 0.002$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $0.596 \pm 0.011$\end{document} PepDoRAPeptide \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $0.709 \pm 0.011$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $0.507 \pm 0.008$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $0.591 \pm 0.005$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $0.843 \pm 0.005$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $0.484 \pm 0.006$\end{document} Ablation study results on cytokine datasets (IL-2, IL-4, IL-10, and IFN-\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $\gamma$\end{document}) for Antigen extracted (Source: Antigen) and Peptide extracted (Source: Peptide) models. Values are reported as mean \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $\pm$\end{document} standard deviation. Together, these results indicate that the dominant performance factor is not the choice of peptide encoder per se, but the inclusion of antigen-level context, which enables the model to capture broader structural and contextual dependencies beyond the isolated peptide. To study how context information influences prediction of peptides from the same proteins, we split the data (i) randomly, (ii) based on peptide sequence, and (iii) based on antigen sequence. We assessed the model performance by splitting the data based on antigen sequence similarity at thresholds of 95%, 90%, and 85% for both antigen extracted and peptide extracted representations. The results do not show significant difference in performance metrics between options (i) and (ii) (see Supplementary Table S1). However, there is a significant drop in metrics values and increased standard deviation when comparing option (iii) to option (ii) for both Antigen extracted and Peptide extracted models across all similarity thresholds. The performance drop when grouping data by antigen similarity points to a more fundamental challenge than data leakage. Specifically, it demonstrates that the reduction in performance is not attributable to the use of the whole antigen as an embedding source, but rather stems from the low peptide diversity inherent in splits defined by protein homology. This lower diversity leads to models overfitting on similar peptides and showing lower metrics on test set and high standard deviation between dataset folds. Based on these observations, and to ensure robust and generalizable performance in future work, we have opted to employ the random splitting strategy. This approach maximizes peptide diversity in the training set and mitigates the risk of overfitting induced by protein-level similarities. Given our focus on predictions in human hosts, we investigated the efficacy of training models on mixed-host and human-only datasets, while evaluating on human-only test sets. The results in Supplementary Table 2 reveal that all models trained on human-specific data outperform those trained on mixed host data. While the observed improvements are not significant for IL-2, IL-4, and IL-10, the improvement for IFN-\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $\gamma$\end{document} is significant (MCC \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $0.627 \pm 0.003$\end{document} versus \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $0.677 \pm 0.006$\end{document}). We believe that this is due to the fact that including nonhuman data introduces challenges such as distribution shifts and potential label heterogeneity, which affect model accuracy on human test sets. To highlight the advantages of using pLM peptide representations we conducted experiments using only classical sequence descriptors, only ESM C peptide embeddings, and their combinations, as shown in Fig. 2 and Supplementary Table 3. For IL-2, IL-4, and IL-10, the combination of ESM C embeddings and either of the classical sequence descriptors does not significantly improve the performance, therefore we choose to keep only ESM C embeddings in the final models for these tasks. For IFN-\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $\gamma$\end{document}, however, the best performing combination is ESM C and all classical descriptors together, achieving an MCC of 0.717 \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $\pm$\end{document} 0.005. Comparison of Matthews correlation coefficient across cytokine datasets and feature representations (AAC—Amino Acid Composition, DPC—Dipeptide Composition, AAP—and Amino Acid Pair Composition). To mitigate the imbalance between positive and negative samples, we employed synthetic oversampling techniques. We evaluated effects of applying SMOTE to IL-2 and IL-4 datasets, and both SMOTE and ADASYN to IL-10 and IFN-\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $\gamma$\end{document}. For IL-10 synthetic augmentation with SMOTE yields the largest gains, likely due to the highest class imbalance in this dataset. For IFN-\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $\gamma$\end{document}, IL-2 and IL-4 synthetic oversampling shows no performance improvement (Supplementary Table 3). Our ablation study concluded that the optimal architecture is cytokine- and data-specific: ESM-C performs best for IL-2 and IL-4, while its combination with SMOTE is optimal for IL-10, and its integration with all descriptors yields the highest performance for IFN-\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $\gamma$\end{document}. Each of these top-performing models is an ensemble of five instances trained on the human-only dataset. The final models are named correspondingly EpiLAMA-IL2, EpiLAMA-IL4, EpiLAMA-IL10, and EpiLAMA-IFN\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $\gamma$\end{document}. Comparison with existing methods We benchmark our models against widely used cytokine predictors, including IL2Pred [16], Meta-IL4 [19], IL10Pred, ILeukin10Pred [39], ImmuScope-IM [6], Omran et al. [24], and IFNepitope2 [40] (see Table 4). Many of these baselines rely on manually engineered sequence descriptors coupled with traditional classifiers. Such pipelines use fixed physico-chemical feature sets and feature projectors, which can miss higher-order, context-dependent regularities in epitope sequences. Table 4Comparison with existing methods (values marked with \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $^{\dagger }$\end{document} are reported directly in the corresponding publication; “–” indicates that the metric was not reported in that study).CytokineDatasetModelPrecisionRecallF1AUCMCCIL-2[16]IL2Pred–0.770\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $^{\dagger }$\end{document}–0.840\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $^{\dagger }$\end{document}0.510\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $^{\dagger }$\end{document}EpiLAMA-IL2 \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $\mathbf{0.776} \pm 0.023$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $\mathbf{0.819} \pm 0.017$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $\mathbf{0.796} \pm 0.010$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $\mathbf{0.855} \pm 0.010$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $\mathbf{0.564} \pm 0.018$\end{document} IL-4[17]IL4Pred–0.787\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $^{\dagger }$\end{document}––0.510\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $^{\dagger }$\end{document}EpiLAMA-IL4 \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $\mathbf{0.717} \pm 0.038$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $\mathbf{0.833} \pm 0.031$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $\mathbf{0.770} \pm 0.022$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $\mathbf{0.795} \pm 0.012$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $0.433 \pm 0.027$\end{document} IL-4\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $_{M}$\end{document}Meta-IL4 \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $0.665 \pm 0.037$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $0.579 \pm 0.029$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $0.619 \pm 0.031$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $0.668 \pm 0.030$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $0.207 \pm 0.044$\end{document} EpiLAMA-IL4 \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $\mathbf{0.720} \pm 0.036$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $\mathbf{0.844} \pm 0.033$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $\mathbf{0.777} \pm 0.030$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $\mathbf{0.793} \pm 0.020$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $\mathbf{0.448} \pm 0.052$\end{document} IL-4\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $_{H}$\end{document}Meta-IL4 \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $0.759 \pm 0.039$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $0.638 \pm 0.045$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $0.692 \pm 0.035$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $0.746 \pm 0.052$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $0.326 \pm 0.095$\end{document} EpiLAMA-IL4 \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $\mathbf{0.791} \pm 0.020$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $\mathbf{0.879} \pm 0.066$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $\mathbf{0.831} \pm 0.021$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $\mathbf{0.851} \pm 0.031$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $\mathbf{0.559} \pm 0.052$\end{document} IL-10[18]IL10Pred–0.779\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $^{\dagger }$\end{document}––0.540\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $^{\dagger }$\end{document}EpiLAMA-IL10 \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $\mathbf{0.727} \pm 0.028$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $\mathbf{0.796} \pm 0.052$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $\mathbf{0.759} \pm 0.032$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $\mathbf{0.918} \pm 0.025$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $\mathbf{0.641} \pm 0.055$\end{document} [20]ILeukin10Pred \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $0.723 \pm 0.030$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $0.603 \pm 0.080$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $0.655 \pm 0.048$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $0.755 \pm 0.033$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $0.540 \pm 0.048$\end{document} EpiLAMA-IL10 \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $\mathbf{0.744} \pm 0.104$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $\mathbf{0.621} \pm 0.071$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $\mathbf{0.669} \pm 0.018$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $\mathbf{0.906} \pm 0.025$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $\mathbf{0.570} \pm 0.025$\end{document} IFN-\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $\gamma _{M}$\end{document}ImmuScope-IM \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $0.796 \pm 0.022$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $0.771 \pm 0.045$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $0.782 \pm 0.018$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $0.828 \pm 0.013$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $0.505 \pm 0.027$\end{document} Omran et al. [24] \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $0.860 \pm 0.008$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $0.626 \pm 0.019$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $0.725 \pm 0.012$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $0.833 \pm 0.012$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $0.517 \pm 0.020$\end{document} EpiLAMA-IFN\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $\gamma$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $\mathbf{0.824} \pm 0.010$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $\mathbf{0.825} \pm 0.013$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $\mathbf{0.824} \pm 0.010$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $\mathbf{0.886} \pm 0.008$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $\mathbf{0.618} \pm 0.016$\end{document} IFN-\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $\gamma$\end{document}IFN-\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $\gamma _{H}$\end{document}ImmuScope-IM \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $0.801 \pm 0.031$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $0.787 \pm 0.036$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $0.793 \pm 0.015$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $0.824 \pm 0.018$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $0.494 \pm 0.031$\end{document} Omran et al. [24] \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $\mathbf{0.866} \pm 0.027$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $0.696 \pm 0.027$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $0.771 \pm 0.013$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $0.829 \pm 0.023$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $0.525 \pm 0.030$\end{document} EpiLAMA-IFN\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $\gamma$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $0.819 \pm 0.024$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $\mathbf{0.842} \pm 0.014$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $\mathbf{0.830} \pm 0.017$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $\mathbf{0.876} \pm 0.018$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $\mathbf{0.592} \pm 0.036$\end{document} [40]\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $_{H}$\end{document}IFNepitope20.8040.8150.8100.890.61EpiLAMA-IFN\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $\gamma$\end{document} 0.835 0.845 0.840 0.917 0.678 Comparison with existing methods (values marked with \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $^{\dagger }$\end{document} are reported directly in the corresponding publication; “–” indicates that the metric was not reported in that study). In contrast, our approach relies on pLM embeddings computed directly from the amino acid sequences, and applies to them a light-weight classifier ensemble. This embedding-first design yields richer, context-aware representations before the decision layer and, as a result, improves both threshold-independent (AUC) and threshold-dependent (MCC, F1, Precision/Recall) metrics. Within each task, models being compared in Table 4 are trained and tested on the same datasets. Across all tasks, our approach yields best or co-best scores in every evaluation metric. Our approach shows the highest relative gains for the hardest task (IL-4) and consistent improvements on benchmarks (IL-2, IL-10, IFN-\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $\gamma$\end{document}). Immunogenicity prediction To further build upon our EpiLAMA-IL models we developed an immunogenicity prediction meta-score that incorporates predictions from different stages of the CD4 immune response, including cytokine release stage. For a peptide to generate an immune response, it first needs to be produced through lysosomal proteolysis of the parent protein [41]. Thus, it may be useful to incorporate scores reflecting a peptide’s ability to be cleaved into the immunogenicity prediction model. Since peptide binding at least one MHC-II allele is a necessary first step in the CD4+ T cell immune response formation [41], we included binding affinity scores for several alleles in our meta-score. An often (but not always) observed part of the immune response is the release of IFN-\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $\gamma$\end{document} and IL-2 after MHC-bound peptides bind TCRs [42, 43], whereas release of IL-4 or IL-10 after TCR binding dampens the immune response [42]. We included cytokine release prediction scores from EpiLAMA-IL* models to take this into account. Finally, the peptide’s inherent physico-chemical properties, such as solubility, stability, secondary structure, hydrophobicity, etc., are important factors that determine its ability to be recognized by the immune system [44], and should also be considered by the immunogenicity model. To demonstrate the importance of combining several scores, we provided metrics for using each individual score as immunogenicity predictor, as well as for models trained on combinations of groups of available scores (e.g. only MHC allele binding scores, or only EpiLAMA-IL scores), in addition to a model that utilizes all available scores. We determined five individual scores with best MCC performance on both train and test sets of the ProIMM\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $_{H}$\end{document} and CD4Episcore datasets (see Supplementary Table S4). Our results show that while individual scores sometimes achieve decent MCC and ROC AUC, performance of these scores is not stable when comparing performance of the same score on train and test subsets of the same datasets, suggesting limited utility of using a single score as a predictor, as opposed to considering their combination. To find the optimal set of scores to include in our model, we trained our model on all possible combinations of four groups of scores (16 models in total): EpiLAMA-IL* scores, MHC allele binding affinity scores, physico-chemical property scores, and DeepPeptide scores. We provided the comparison between all these models in Supplementary Tables S5 and S6. We did not see that a single combination of scores performed best for both ProIMM\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $_{H}$\end{document} and CD4Episcore datasets. This is not unexpected, considering the difference in assay types and hosts present within these datasets. We found that the combination of scores that provides good performance on both of these datasets is EpiLAMA-IL scores, MHC allele binding affinity scores, and physico-chemical scores. Indeed, this combination achieves the best ROC AUC and PR AUC on the proliferation dataset, as well as third best ROC AUC on the CD4Episcore dataset, while also keeping high PR AUC and MCC. We called this model EpiLAMA-IM. We compared EpiLAMA-IM with two other immunogenicity prediction approaches in Table 5 on both ProIMM\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $_{H}$\end{document} and CD4Episcore datasets. For each model in this table we chose an optimal decision threshold to maximize MCC. Table 5Comparison of immunogenicity prediction performance on a held-out proliferation test set and the CD4Episcore test set with the best scores in each column highlighted in bold (metrics for models marked with * are taken from the paper; training datasets marked with ** may have been modified by the authors, as the metrics for these models on the test set do not match the ones reported in the paper).ModelTraining datasetThresholdPrecisionRecallF1ROC AUCMCCPR AUC ProIMM-test\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $_{H}$\end{document} VaxiJen-bacteria[22]0.40.6330.5730.6020.437-0.0880.623VaxiJen-virus[22]0.40.6600.5600.6060.488-0.0170.666VaxiJen-tumor[22]0.50.6790.5490.6070.5190.0290.680VaxiJen-parasite[22]0.50.6450.5090.5690.470-0.0490.659VaxiJen-fungal[22]0.50.6550.5390.5920.488-0.0270.661CD4Episcore (server)[6](train+test)\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $^{**}$\end{document}0.150.6520.8580.7410.5500.0870.670CD4Episcore (latest)[6](train+test)\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $^{**}$\end{document}0.20.6860.7380.7110.5920.1350.719EpiLAMA-IM[6](train+test)0.050.691 0.917 0.788 0.558 0.146 0.710CD4Episcore-IMM (server)[6](train)\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $^{**}$\end{document}0.050.6810.4920.5720.5560.0900.673CD4Episcore-IMM (latest)[6](train)\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $^{**}$\end{document}0.50.7900.0590.1100.5310.0690.680EpiLAMA-IM[6](train)0.100.7310.4930.5890.5910.1250.728EpiLAMA-IMProIMM-train\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $_{H}$\end{document}0.55 0.871 0.1170.207 0.600 0.137 0.756 CD4Episcore-test VaxiJen-bacteria[22]0.40.2270.5720.3250.471-0.0350.233VaxiJen-virus[22]0.40.2260.5320.3180.457-0.0330.219VaxiJen-tumor[22]0.50.2420.4860.3230.4960.0070.235VaxiJen-parasite[22]0.50.2260.5090.3130.476-0.0330.222VaxiJen-fungal[22]0.50.2330.5410.3260.480-0.0150.235CD4Episcore\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $^{*}$\end{document}[6](train+test)---0.725--CD4Episcore (server)[6](train+test)\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $^{**}$\end{document}0.550.6540.2720.3840.6880.3240.496CD4Episcore (latest)[6](train+test)\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $^{**}$\end{document}0.60.8030.3920.5260.7780.4810.635EpiLAMA-IM[6](train+test)0.350.6810.4760.5600.8280.4660.646CD4Episcore-IMM\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $^{*}$\end{document}[6]---0.702--CD4Episcore-IMM (server)[6]\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $^{**}$\end{document}0.10.6920.6580.6750.8690.5770.749CD4Episcore-IMM (latest)[6]\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $^{**}$\end{document}0.35 0.930 0.691 0.793 0.924 0.754 0.879 EpiLAMA-IM[6]0.100.4090.6120.4900.7260.3000.414EpiLAMA-IMProIMM-train\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $_{H}$\end{document}0.500.275 0.858 0.4160.6650.1510.361 Comparison of immunogenicity prediction performance on a held-out proliferation test set and the CD4Episcore test set with the best scores in each column highlighted in bold (metrics for models marked with * are taken from the paper; training datasets marked with ** may have been modified by the authors, as the metrics for these models on the test set do not match the ones reported in the paper). VaxiJen [22] is an online server for prediction of peptide’s probability of being an antigen, i.e. its immunogenicity, using ACC descriptors. While it was not developed to predict immunogenicity of peptides in human cells, this tool can be used to assess a peptide’s general immunogenicity potential, and has been successfully applied [45] for candidate filtering in reverse vaccinology. VaxiJen online server provides separate models for peptides of different origins: viral, fungal, tumoral, bacterial, and parasitic. In Table 5 we provide performance metrics of each of these models separately. CD4Episcore [6] is a tool for the prediction of peptide immunogenicity (validated on a dataset of mostly CD4+ cell proliferation assays) based on IEDB’s 7-allele binding affinity scores and an additional separately-trained immunogenicity score. The immunogenicity score in CD4Episcore (CD4Episcore-IMM) is trained on a dataset that combines cytokine release and tetramer staining assays, to which we refer to as CD4Episcore-train. The final CD4Episcore prediction is a linear combination of this immunogenicity score and the IEDB’s 7-allele binding affinity score, weighted by \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $\alpha$\end{document} – a parameter tuned on the dataset of cytokine release, T-cell proliferation, and activation assays. The metrics for the combined score are also reported on this dataset, which is why we refer to it as CD4Episcore-test, despite it being used to train the final model. CD4Episcore is available as a part of the IEDB suite of epitope prediction tools (the latest version), as well as a standalone server mentioned in the original paper (https://tools.iedb.org/CD4episcore/), and has been successfully applied in vaccine design as well [46]. To ensure a fair comparison with CD4Episcore, we trained three versions of our model: one trained on our proliferation dataset, another trained on the CD4Episcore-train dataset (to be compared with CD4Episcore-IMM), and one trained on the combination of CD4Episcore train and test datasets (to be compared with CD4Episcore). While EpiLAMA-IM trained on ProIMM-train\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $_{H}$\end{document} shows one of the best results in terms of ROC AUC, and MCC on the proliferation test set, it also demonstrates poor transferability to the CD4Episcore dataset, with sharp changes in PR AUC and Precision. At the same time, EpiLAMA-IM trained on the CD4Episcore datasets shows both strong performance on the CD4Episcore test and the best performance on the proliferation test in terms of MCC. Compared with both CD4Episcore-IMM and CD4Episcore, EpiLAMA-IM performs better in terms of ROC AUC, when compared with metrics reported in the original paper for these models. Since metrics for predictions obtained from the CD4Episcore (both server and IEDB tools versions) differ significantly from the ones reported in the paper, these models may have been retrained with new data, and comparison on the same datasets may not be possible. Despite this difference, EpiLAMA-IM model trained on CD4Episcore train+test achieves similar performance to the latest CD4Episcore model on both CD4Episcore and ProIMM-test\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $_{H}$\end{document} test sets. Based on the comparison we conducted, we found that EpiLAMA-IM trained on the CD4Episcore train+test outperforms the original paper version of the CD4Episcore model trained on the same dataset, and performs on par with the latest CD4Episcore model version on both test datasets. Discussion Predicting peptide immunogenicity remains one of the most challenging problems in computational immunology. While MHC binding prediction has reached high levels of accuracy, MHC binding alone does not fully describe the diverse functional outcomes of antigen presentation. Peptides with comparable HLA binding affinities may elicit opposite immune effects—from strong Th1-type activation to immune tolerance. This functional divergence is largely governed by cytokine signaling, which determines the polarization and magnitude of T-cell responses. Therefore, integrating cytokine induction profiles into immunogenicity modeling offers a more mechanistic and biologically interpretable framework for understanding immune activation. In this study, we systematically investigated cytokine-specific sequence properties and demonstrated that their variability has practical consequences for computational prediction. Higher sequence heterogeneity, particularly among IL-4–inducing epitopes, increases model complexity and reduces discriminative power. In contrast, the lower sequence diversity observed for IL-10–inducing epitopes results in better and more stable performance, possibly due to recognition of conserved motifs. These findings highlight that cytokine-specific classification tasks differ substantially and are influenced by the underlying sequence variability of their epitopes. Across all tasks, pLM embeddings substantially outperform hand-crafted descriptors. The ESM C model, in particular, captures higher-order contextual dependencies and achieves superior AUC and MCC values for all cytokines. These results emphasize that contextualized embeddings derived from large-scale protein models can encapsulate the structural and functional information necessary for immunological inference, even without explicit physico-chemical feature design. Ablation experiments confirm that antigen-level embeddings outperform peptide-only representations. Incorporating the surrounding antigenic context allows models to learn biologically meaningful dependencies between local sequence motifs and their structural neighborhood, improving prediction robustness. We hypothesize that this contextual information allows the model to capture the propensity of a certain peptide to be cleaved by the protease from the antigen, thus providing additional information crucial for prediction of cytokine release. This observation aligns with the growing recognition that immune recognition cannot be fully captured by isolated peptide features. We also highlight the importance of tailoring models to the human immune system when predicting cytokine induction. We show that models trained on human host datasets show better performance on human test data compared with training on mixed-host data, despite human-only datasets being smaller in size. We believe training on mixed datasets compromises predictive accuracy due to the inherent differences in immune system proteins responsible for cytokine release that lead to significant epitope variability across species. Benchmarking against established cytokine predictors demonstrates that the EpiLAMA-IL family of models achieves state-of-the-art performance across multiple cytokines. Notably, EpiLAMA-IL2, EpiLAMA-IL10, and EpiLAMA-IFN\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $\gamma$\end{document} surpass existing methods in both AUC and MCC metrics, highlighting the advantage of pLM embedding-based peptide representations in capturing immunologically relevant sequence features across diverse datasets. Finally, integrating cytokine induction scores with HLA binding, as well as physico-chemical descriptors within the EpiLAMA-IM meta-model produces a unified framework for immunogenicity prediction. While some individual scores show good performance on train or test subsets, our model built on combination of these scores shows more stable performance across our test sets, highlighting the need for integration of multiple scores for more reliable immunogenicity prediction. This integrative approach outperforms classical methods, such as VaxiJen and CD4Episcore. Surprisingly, adding DeepPeptide protease cleavage probability scores into the immunogenicity model does not significantly improve the performance. While protease cleavage likelihood is an important factor for determining peptide’s immunogenicity, using these scores alone demonstrates rather poor class separation ability compared with others (see Supplemetary Tables S5 and S6). A likely explanation is the lack of feature variance within our dataset: since most of the peptides present in the datasets are experimentally validated to be processed by proteases, it is hard to learn a discriminatory signal from this information. Therefore, adding them to the current model may introduce unnecessary noise, but could be useful in other datasets. We also find that both models trained on the proliferation dataset and on the mixed-assay CD4Episcore dataset fail to achieve ROC AUC higher than 0.6 on the ProIMM\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $_{H}$\end{document} test set. Although we enforced strict filtering thresholds during dataset construction, the relatively low scores may be the result of data filtering still being not strict enough or due to noise in the experimental data comprising this dataset. Another reason could be that the input features used by EpiLAMA-IM and CD4Episcore (namely, cytokine, and allele scores derived from peptide and parent protein sequences) may still lack key T-cell proliferation predictors, making this task too difficult to solve with these inputs alone. In summary, our findings highlight the potential of transformer-based protein language model embeddings to revolutionize immunogenicity prediction. By combining contextual sequence representations with immune system–specific functional data, this work advances the field toward more mechanistic, interpretable, and generalizable models of immune activation. Conclusion This study introduces a cytokine-informed framework for peptide immunogenicity prediction that bridges antigen presentation and cytokine profile with downstream functional outcomes. By integrating cytokine induction modeling with HLA binding affinity and physico-chemical peptide properties, the proposed EpiLAMA-IM architecture provides a biologically grounded and computationally robust approach to estimate peptide immunogenicity. Our results demonstrate that protein language model embeddings substantially improve prediction accuracy across multiple cytokines and outperform traditional feature-based methods. Incorporating antigen-level context further enhances model generalization, underscoring the importance of contextual sequence information for immune modeling. Overall, EpiLAMA-IM offers a unified and interpretable strategy for predicting functional immune responses, supporting more rational epitope selection in vaccine design, immunotherapy development, and immune safety assessment. Key Points We develop a suite of models EpiLAMA-IL for peptide sequence-based prediction of IL-2, IL-4, Il-10, and IFN-\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $\gamma$\end{document} release that show state-of-the art performance. We develop a suite of models EpiLAMA-IL for peptide sequence-based prediction of IL-2, IL-4, Il-10, and IFN-\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} $\gamma$\end{document} release that show state-of-the art performance. We perform systematic ablation studies and show that using protein language model embeddings as classifier inputs consistently outperforms other hand-crafted or learned peptide features for all cytokine release prediction tasks. We perform systematic ablation studies and show that using protein language model embeddings as classifier inputs consistently outperforms other hand-crafted or learned peptide features for all cytokine release prediction tasks. We develop a model for peptide immunogenicity prediction (EpiLAMA-IM) that incorporates peptide-MHC binding scores, EpiLAMA-IL cytokine release scores, and physico-chemical scores. EpiLAMA-IM achieves the best performance on our immunogenicity benchmark. We develop a model for peptide immunogenicity prediction (EpiLAMA-IM) that incorporates peptide-MHC binding scores, EpiLAMA-IL cytokine release scores, and physico-chemical scores. EpiLAMA-IM achieves the best performance on our immunogenicity benchmark. Supplementary Material