Identifying potential ligand-receptor interactions by integrating LSTM network and the attention mechanism for cell-cell communication prediction Authors: Yingwei Deng, Min Chen, Pengfei Gao, Ruogu Luo, Zejun Li, Yuhua Yao Published: 2026 ABSTRACT BackgroundCell-cell communication (CCC) mediated by ligand-receptor (L-R) interactions is fundamental to deciphering tissue development and disease mechanisms. While single-cell RNA sequencing (scRNA-seq) has advanced this field, existing computational methods for inferring CCC often suffer from limitations such as dependence on static databases and a failure to capture the sequential dependency of amino acids within proteins, which restricts their generalizability and predictive accuracy. Therefore, the primary objective of this study was to develop a robust computational framework capable of identifying potential L-R interactions directly from protein sequence data, thereby overcoming the reliance on static databases and enabling the discovery of novel signaling pairs.MethodsTo achieve this objective, we introduce CellAL, a deep learning-based framework for predicting potential interacting L-R pairs and decoding cellular communication. The CellAL pipeline consists of two main stages: (1) L-R Pair Identification, which extracts sequence features using BioTriangle, selects informative features via XGBoost to reduce dimensionality, and classifies interactions using a Long Short-Term Memory (LSTM) network integrated with an attention mechanism specifically designed to capture long-range sequence dependencies that characterize structural binding affinities; and (2) CCC Inference, which filters identified pairs using scRNA-seq data and quantifies crosstalk intensity through a comprehensive scoring strategy that combines expression thresholding, expression product, and specific expression metrics.ResultsPerformance evaluations on four standard L-R interaction datasets demonstrated that CellAL significantly surpassed classical protein-protein interaction prediction methods and achieved competitive performance against state-of-the-art ensemble models, achieving the highest AUPR values on three datasets. The identified To achieve L-R pairs showed a high degree of overlap with existing databases such as CellChat and Connectome. Furthermore, when applied to human melanoma scRNA-seq data, CellAL successfully inferred critical signaling networks, revealing strong bidirectional crosstalk between melanoma cells and cancer-associated fibroblasts (CAFs), macrophages, and endothelial cells. These findings were consistent with results from three other representative CCC prediction tools.ConclusionsCellAL effectively overcomes the limitations of database dependence by leveraging sequence-level biochemical modeling to predict structural L-R interactions. By integrating deep learning predictions with transcriptomic data, CellAL provides a robust and valuable tool for dissecting complex CCC networks at single-cell resolution, particularly within the tumor microenvironment. Introduction In multicellular organisms, cells coordinate their activities through communication between various cell types. Cell-cell communication (CCC) is fundamental to perform the biological functions of cells, tissues, organs, and systems [1-3]. Single-cell RNA sequencing (scRNA-seq) provides unprecedented resolution and promotes our understanding of intracellular and intercellular heterogeneity [4, 5]. We can accurately infer rare cell subpopulations [6], construct complex gene-gene regulatory networks [7], investigate drug response [8–11], and decipher cellular crosstalk across communicating cell types as well as under diseased and homeostatic conditions based on scRNA-seq data [12]. In vivo experiments have advanced the development of cellular signaling pathway analysis. However, these wet-laboratory methods are time-consuming, laborious, and expensive. Consequently, artificial intelligence techniques have been increasingly accumulated to study CCC from scRNA-seq data. To facilitate this, comprehensive resources such as CellCommuNet have been established, curating hundreds of datasets to explore communication differences between disease and normal states [13]. CCC is driven by various mediators including protein-mediated interactions, metabolite-mediated interactions, and host–microbiome interactions [14]. However, ligand–receptor interactions (LRIs)—where ligands in one cell bind to cognate receptors on other cells—are dominant mediators for cellular communications analysis [15]. Recently, this scope has expanded to include non-peptide ligands (NPLs) modeled by tools like MetaLigand [16], and extracellular vesicle-mediated miRNA communication inferred by miRTalk [17], highlighting the complexity of intercellular dynamics. Recently, many artificial intelligence tools have been designed for CCC inference through LRIs using scRNA-seq data. For example, CellPhoneDB [18] and ICELLNET [19] evaluated the product of Ligand and Receptor (L-R) expressions. The widely used CellChat tool has recently evolved to CellChat v2, incorporating an expanded database, interactive visualization, and enhanced functionalities for comparing communication across different biological conditions [20, 21]. SingleCellSignalR [22] computed a p-value for each LRI through a score null distribution when deciphering statistically significant and cell-type–specific signaling. CellChat [21], CellPhoneDB [18], and ICELLNET [19] recapitulated known LRIs based on the multi subunit structures of L-R pairs. CellEnBoost [23], CellDialog [2] and CellGDnG [24] decoded potential LRIs and further inferred CCC by combining scRNA-seq data. In addition to the above LRI analysis tools, several tools both analyze LRIs and investigate downstream signal pathways. to better understand regulatory mechanisms. For example, to analyze gene expressions from multiple facets, PyMINEr [25] built autocrine and paracrine signaling networks. More recently, CausalCCC was developed to reconstruct causal gene-gene interaction pathways by integrating upstream and downstream signals [26]. Similarly, SigXTalk utilizes hypergraph learning to quantify signal fidelity and specificity, explicitly modeling the crosstalk between pathways that was previously overlooked [27]. To decode the functionally influential CCC, NicheNet [28] utilized downstream signaling changes to assess the activity of the upstream receptors. To depict observed heteromeric molecular complexes, CellChat [21] arranged data source of interactions involved in ligands, receptors and corresponding cofactors. Furthermore, CCC inference based on LRIs can remold cellular structure of tissue by integrating spatial transcriptomics data. CSOmap [29] and SpaOTsc [30] recapitulated cellular spatial information through LRIs. Advanced methods like SpaCcLink now account for downstream intracellular signals to decode spatial communication patterns [31], while STCase employs interpretable graph neural networks to uncover niche-specific CCC events that vary across microenvironments [32]. Addressing patient heterogeneity, scACCorDiON applies optimal transport to analyze sample-specific communication events in large disease cohorts [33]. SpaTalk [34] designed a non-negative linear model and scored the L-R-target signaling network involving in spatially proximal cells by deconvoluting cell-type composition. Additionally, Giladi et al. [35] presented a sequencing-based model PIC-seq to decode cellular dialogues by sequencing physically communicating cells. PIC-seq incorporated cell ranking of physically communicating cells and scRNA-seq and systematically depicted L-R crosstalk between cells. It provides a direct tool to dissect intercellular communication specific pathways. Consequently, a comprehensive knowledge of LRIs can efficiently facilitate the deciphering of cellular crosstalk However, despite the advancements of these sophisticated tools in modeling spatial contexts, downstream pathways, and diverse signaling modalities, most existing LRI prediction methods still suffer from limitations including database dependence, limited generalizability, and lack of sequence-level biochemical modeling. Moreover, standard machine learning models often ignore the sequential dependency of amino acids within proteins. Although ligand-receptor interactions are not temporal events, proteins are composed of amino acids arranged in a specific linear order, which determines their physicochemical properties and binding affinities. To address this, we hypothesize that capturing the long-range dependencies within amino acid sequences is crucial for accurate LRI prediction. Thus, in this manuscript, we propose a computational model called CellAL for predicting potential interacting L-R pairs and further decoding cellular communication. CellAL integrates sequence-derived descriptors, XGBoost feature prioritization, and feature-wise attention for heterogeneous ligand–receptor features to address the aforementioned limitations. After testing the performance of CellAL, we employ it to deduce cellular crosstalk within melanoma. This work has three contributions:A deep learning-based model CellAL is devised to predict potential interacting L-R pairs by integrating L-R feature extraction through BioTriangle, L-R feature selection with XGBoost, and L-R classification using Long Short-Term Memory (LSTM) network with the attention mechanism.Expression thresholding, expression product, and specific expression of L-R pairs are combined to evaluate CCC.Cancer-associated fibroblasts (CAFs) is inferred to communicate with melanoma cells. A deep learning-based model CellAL is devised to predict potential interacting L-R pairs by integrating L-R feature extraction through BioTriangle, L-R feature selection with XGBoost, and L-R classification using Long Short-Term Memory (LSTM) network with the attention mechanism. A deep learning-based model CellAL is devised to predict potential interacting L-R pairs by integrating L-R feature extraction through BioTriangle, L-R feature selection with XGBoost, and L-R classification using Long Short-Term Memory (LSTM) network with the attention mechanism. Expression thresholding, expression product, and specific expression of L-R pairs are combined to evaluate CCC. Expression thresholding, expression product, and specific expression of L-R pairs are combined to evaluate CCC. Cancer-associated fibroblasts (CAFs) is inferred to communicate with melanoma cells. Cancer-associated fibroblasts (CAFs) is inferred to communicate with melanoma cells. Materials and methods Datasets Peng et al. [23] provide four L-R datasets by arranging data from existing public data sources. As shown in Table 1, Datasets 1 (human) and 2 (mouse) were from CellTalkDB [36]. Datasets 3 (mouse) and 4 (human) were from Skelly et al. [37] and Ximerakis et al. [38], respectively. These L-R datasets were preprocessed by deleting repeated LRIs and LRIs without sequence information in the Uni Prot database [39]. To address the class imbalance, we constructed a balanced dataset by randomly under-sampling the non-interacting pairs to match the number of interacting pairs.Table 1The information of four L-R datasetsDatasetLigandsReceptorsLRIsDataset1 (human)8127803,390Dataset2 (mouse)6505882,031Dataset3 (mouse)5745592,006Dataset4 (human)1,1291,3356,585 Table 1The information of four L-R datasetsDatasetLigandsReceptorsLRIsDataset1 (human)8127803,390Dataset2 (mouse)6505882,031Dataset3 (mouse)5745592,006Dataset4 (human)1,1291,3356,585 The information of four L-R datasets CallAL In this manuscript, we present a computational algorithm called CellAL for predicting potential interacting L-R pairs and further inferring intercellular signaling. As shown in Fig. 1, the pipeline begins with (A) sequence and expression data input. CellAL then predicts potential L-R pairs through a (B) deep learning prediction framework, which involves extracting L-R features using BioTriangle [40], reducing the dimension of the extracted features through XGBoost, and predicting L-R pairs with high confidence by designing an LSTM network with the attention mechanism. Next, it performs (C) CCC inference and scoring by filtering known and predicted L-R pairs with scRNA-seq data and quantifying communication strength via expression threshold, product, and specificity. Finally, after (D) final output generation, the workflow concludes with (E) result visualization via network view and heatmap.Fig. 1The pipeline for CellAL based on potential interacting L-R pair prediction and cellular communication inference. A. Sequence and Expression Data Input B. Deep Learning Prediction Framework C. CCC Inference and Scoring D. Final Output Generation E. Result Visualization The pipeline for CellAL based on potential interacting L-R pair prediction and cellular communication inference. A. Sequence and Expression Data Input B. Deep Learning Prediction Framework C. CCC Inference and Scoring D. Final Output Generation E. Result Visualization Potential interacting L-R pair prediction Step 1: L-R feature extraction BioTriangle [40] is a useful web-accessible tool for describing various molecules including proteins. Here, we apply BioTriangle to generate molecular representations for ligands and receptors. As shown in Table 2, we extract 1890 features for each ligand or receptor based on their sequences. And each L-R pair has 3780 features after concatenating their vectors.Table 2The feature number of each ligand or receptorFeaturesNumberAmino acid composition20Dipeptide composition400M-B autocorrelation240Moran autocorrelation240Geary autocorrelation240CTD composition*21CTD transition21CTD distribution105Conjoint triad features343Quasi-sequence order descriptors100Sequence order coupling number60Pseudo amino acid composition 150Pseudo amino acid composition 250Total1890*Where CTD denotes composition (C), transition(T), and distribution (D) Table 2The feature number of each ligand or receptorFeaturesNumberAmino acid composition20Dipeptide composition400M-B autocorrelation240Moran autocorrelation240Geary autocorrelation240CTD composition*21CTD transition21CTD distribution105Conjoint triad features343Quasi-sequence order descriptors100Sequence order coupling number60Pseudo amino acid composition 150Pseudo amino acid composition 250Total1890*Where CTD denotes composition (C), transition(T), and distribution (D) The feature number of each ligand or receptor *Where CTD denotes composition (C), transition(T), and distribution (D) Step 2: L-R feature selection Feature selection is a powerful data preprocessing technique in machine learning. It aims to design an algorithm to select the more informative features and improve the model performance [41]. The obtained L-R features using BioTriangle are high-dimensional and contains much robust information, which could cause the curse of dimension, overfitting, increased computational complexity, and reduced interpretability of the model. XGBoost [42] is a wide-used end-to-end tree growth model. It exhibits good ability when dealing with sparse data. To reduce the L-R feature dimension, we adopt XGBoost for scoring L-R feature importance. By ranking the L-R feature importance and selecting the most informative features, we can reduce the dimension of L-R features, improve the performance, avoid the overfitting, and enhance the interpretability of the model. For an LRI dataset with n L-R pairs \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\cal D} = \left\{ {\left( {{x_i},{y_i}} \right)} \right\}\left( {\left| {\cal D} \right| = n,{x_i} \in {R^m},{y_i} \in \mathbb{R}} \right)$$\end{document}, each pair has 3780 features (i.e., m = 3780). We utilize k additive functions to obtain the model output by Eq. (1): 1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\matrix{ {{{\hat y}_i} = \phi \left( {{x_i}} \right) = \mathop \sum \limits_{k = 1}^K \,{f_k}\left( {{x_i}} \right),{f_k} \in F)} \cr }$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathcal{F} = \left\{ {{\rm{f}}\left( {\bf{x}} \right) = {{\rm{w}}_{{\rm{q}}\left( {\bf{x}} \right)}}} \right\}\left( {{\rm{q}}:{\mathbb{R}^{\rm{m}}} - {\rm{T}},{\rm{w}} \in {\mathbb{R}^{\rm{T}}}} \right)$$\end{document} denotes the space of the regression tree. The L-R features are selected by Eq. (2): 2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\matrix{ {L = {1 \over 2}\left[ {{{{{(\mathop \sum \nolimits_{i \in {I_L}} {\rm{ }}{g_i})}^2}} \over {\mathop \sum \nolimits_{i \in {I_L}} {\rm{ }}{h_i} + \lambda }} + {{{{(\mathop \sum \nolimits_{i \in {I_R}} {\rm{ }}{g_i})}^2}} \over {\mathop \sum \nolimits_{i \in {I_R}} {\rm{ }}{h_i} + \lambda }} - {{{{(\mathop \sum \nolimits_{i \in I} {\rm{ }}{g_i})}^2}} \over {\mathop \sum \nolimits_{i \in I} {\rm{ }}{h_i} + \lambda }}} \right] - \gamma} \cr } $$\end{document} Here, we use the greedy algorithm to solve the model (2). Suppose that \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${g_i} = {\partial _{{{\hat y}^{\left( {t - 1} \right)}}}}l\left( {{y_i},{{\hat y}^{\left( {t - 1} \right)}}} \right)$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${h_i} = \partial _{\hat y\left( {t - 1} \right)}^2l\left( {{y_i},{{\hat y}^{\left( {t - 1} \right)}}} \right)$$\end{document} represent the first and second order derivatives involving the loss function, respectively. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${I_L}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${I_R}$$\end{document} separately represent the instance sets of the left and right nodes after splitting, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\lambda $$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\gamma $$\end{document} represent the regularization parameters. The higher the information gain of an LRI feature denotes its more effective and importance for L-R classification. Consequently, each L-R pair has d features after feature selection. Step 3: interacting L-R prediction LSTM is a special recurrent neural network (RNN) model for solving the gradient disappearance or gradient explosion problems in traditional RNN [43]. It introduces three gates (i.e., input gate, forget gate, output gate) and a cell state to effectively regulate the flow of information and enable the network to perform information transmission over long intervals through an adaptive gating mechanism. For an L-R pair \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${x_i} = \left\{ {{x_{i1}},{x_{i2}},...{x_{it}},...,{x_{id}}} \right\}$$\end{document}, at each iteration, LSTM updates the hidden state \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${h_t}$$\end{document} and the storage unit state \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${m_t}$$\end{document} by Eq. (3): 3\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\left[ {\matrix{ {{I_t}} \cr {{F_t}} \cr {Ou{t_t}} \cr {{{\hat c}_t}} } } \right] = \left[ {\matrix{ S \cr S \cr S \cr T \cr } } \right]W \cdot \left[ {\matrix{ {{h_{t - 1}},{\mkern 1mu} {j_t}} \cr } } \right]$$\end{document} Where4\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\matrix{ {{m_t} = F{o_t} \odot {m_{t - 1}} + I{n_t} \odot {{\hat c}_t}} \cr } $$\end{document} 5\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\matrix{ {{h_t} = Ou{t_t} \odot T\left( {{m_t}} \right)} \cr } $$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${j_t}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\hat c_t}$$\end{document} represent the input and the current state of a single storage unit, respectively. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$I{n_t},{\rm{ }}F{o_t}$$\end{document}, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$Ou{t_t}$$\end{document} denote the results in the input gate, the forgotten gate, and the output gate, respectively. Although LSTM demonstrates good accuracy in traditional classification tasks, it has limitations. First, regardless of the length of the input sequence, LSTM encodes the sequence as a fixed-length hidden state vector through a loop connection and the gating mechanism. As a result, sequence information may be lost or the raw data may not be adequately expressed, especially for long or complex input sequences. It makes the model difficult to capture subtle changes and important features in the input data, resulting in poor performance and generalization ability of the model. Moreover, the output in LSTM’s intermediate state is underutilized. The LSTM performance can be improved by increasing the maximum pooling layer, but it may result in a large amount of information loss [44, 45]. The attention mechanism is an important technique in the field of deep learning, which is inspired by the human visual system [46]. This mechanism sets different weights for different components of the input data, enabling the model to prioritize information that is more important to the task. Based on the attention mechanism, the model can better focus attention on information that are most critical to the task and ignore other less important information when processing data [47]. Adding the attention mechanism to LSTM facilitates maintaining information as possible as and makes LSTM more efficient when processing sequence data. Thus, we present LSTM with the attention mechanism to classify L-R pairs. Suppose that \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\rm{V}}$$\end{document} denotes the output in the LSTM encoder and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$W$$\end{document} denotes the linear transformation coefficient of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\rm{V}}$$\end{document}, the self attention model is represented by Eq. (6): Given a variable length vector \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${a_n}$$\end{document} as the weight, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${c_w}$$\end{document} is the weighted feature, we set a location-based function to combine the current input and the existing sample state based on the number of true features and weighted features by Eqs. (7)-(9): 6\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\matrix{ {V{^{\prime}} = L\left( V \right) = W*V} \cr } $$\end{document} 7\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${a_n} = softmax\left( {\left[ {c_1^T*{V^\prime },c_2^T*V\prime , \ldots ,c_j^T*{V^\prime }} \right]} \right)$$\end{document} 8\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\matrix{ {{C_w} = \left( {{c_1}{\rm{*}}{a_n} + {c_2}{\rm{*}}{a_n} + , \ldots , + {c_n}{\rm{*}}{a_n}} \right)} \cr } $$\end{document} 9\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\matrix{ {{{\hat C}_n} = \left[ {\mathbb{C},{\mathbb{C}_w}} \right]} \cr } $$\end{document} Based on \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$p(y|x) = softmax\left( {MLP\left( {\left[ {{{\hat C}_n},{h_n}} \right]} \right)} \right)$$\end{document}, each L-R pair is set to interaction probability. Cellular communication inference Step 1: L-R filtering CCC is commonly mediated by interacting L-R pairs. To test CCC intensity within a tissue, first, we filtrate L-R pairs from known and predicted pairs via a threshold: an L-R pair is viewed as interacting if its linkage probability is larger than the threshold, otherwise, it is no-interacting. Subsequently, the filtrated pairs are further filtered based on scRNA-seq data from GEO [48]: when their ligand and/or receptor has no expression within a tissue, they will be viewed as no-mediating signaling. As a result, we can better focus on biologically relevant L-R pairs and dissect cellular crosstalk accurately. Step 2: \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\bf{\it{LR}}}{{\bf{\it{I}}}_{{\bf{\it{p}}},{\bf{\it{q}}},{\bf{\it{m}}},{\bf{\it{n}}}}}$$\end{document} computation For a filtered L-R pairs \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\left\{ {\left( {{l_1},{r_1}} \right),\left( {{l_2},{r_2}} \right)....\left( {{l_a},{r_a}} \right)} \right\}$$\end{document}, and the gene expression \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${l_p}$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$and$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${r_q}$$\end{document}of the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$p - q$$\end{document} pairs in scRNA-seq data, let \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\matrix{ {{{{{\bar L}_{p,m}}} \mathord{\left/ {\vphantom {{{{\bar L}_{p,m}}} {{{\bar R}_{q,n}}}}} \right. \kern-\nulldelimiterspace} {{{\bar R}_{q,n}}}}} } $$\end{document} denotes the arithmetic mean expression of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$p/q{\rm{ }}$$\end{document} in \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${C_m}/{C_n},{\bar L_p}/{\bar R_q}$$\end{document} denotes the arithmetic mean expression of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$p/q$$\end{document} in all cell types. The crosstalk intensity \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$LR{I_{p,q,m,n}}$$\end{document} between two cell clusters \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${C_m}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${C_n}$$\end{document} mediated by the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$p - q$$\end{document} pair is computed by aggregating expression thresholding, expression product, and specific expression [49]. Expression thresholding: If the expression of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$p\left( {or{\rm{ }}q} \right)$$\end{document} is higher than a known threshold in \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${C_m}\left( {or{\rm{ }}{C_n}} \right)$$\end{document}, it has high expression in the cluster. When both of them are highly expressed, the pair is viewed as a signaling pathway between \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${C_m}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${C_n}$$\end{document}. Thus, their mediated cellular crosstalk score \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$LR{I_{p,q,m,n}}$$\end{document} through expression thresholding is computed by Eq. (10): 10\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\matrix{ {LRI_{p,q,m,n}^t = \left( {{{\bar L}_{p,m}} > {{\bar L}_p} + {\sigma _p}} \right)and\left( {{{\bar R}_{q,n}} > {{\bar R}_q} + {\sigma _q}} \right)} } $$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\sigma _p}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\sigma _q}$$\end{document} are the standard deviations. The \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${C_m} - {C_n}$$\end{document} crosstalk score \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${g_t}\left( {m,n} \right)$$\end{document} is evaluated by summing \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$LRI_{p,q,m,n}^t$$\end{document} by Eq. (11): 11\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\matrix{ {{g_t}\left( {m,n} \right) = \mathop \sum \limits_{p,q = 1}^{{z_t}} \,LRI_{p,q,m,n}^t} \cr } $$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${Z_t}$$\end{document} is the number of L-R pairs mediating \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${C_m} - {C_n}$$\end{document} signaling through expression thresh olding. Expression product: The expression product method evaluates CCC through the product of the L-R ex pressions. The \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$p - {\rm{ }}q$$\end{document} mediated cellular crosstalk score \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$LR{I_{p,q,m,n}}$$\end{document} between \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${C_m}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${C_n}$$\end{document} through expression product is computed by Eq. (12): 12\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\matrix{ {LRI_{p,q,m,n}^p = {{\bar L}_{p,m}} \times {{\bar R}_{q,n}}} }$$\end{document} The \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${C_m} - {C_n}$$\end{document} crosstalk score \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${g_p}\left( {m,n} \right)$$\end{document} is evaluated by summing \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$LRI_{p,q,m,n}^t$$\end{document} by Eq. (13): 13\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\matrix{ {{g_p}\left( {m,n} \right) = \mathop \sum \limits_{p,q = 1}^{{z_p}} \,LRI_{p,q,m,n}^p} \cr } $$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${z_p}$$\end{document} is the number of L-R pairs mediating \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${C_m} - {C_n}$$\end{document} signaling through expression product. Specific expression: The specific expression method helps to uncover the most specific crosstalk across cell types regardless of expression levels. The \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$p - {\rm{ }}q$$\end{document} mediated cellular crosstalk score \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$LRI_{p,q,m,n}^{\rm{ }}$$\end{document} between \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${C_m}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${C_n}$$\end{document} through specific expression is computed by Eq. (14): 14\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\matrix{ {LRI_{p,q,m,n}^s = {{{{\bar L}_{p,m}}} \over {\mathop \sum \nolimits_{k = 1}^b {{\bar L}_{p,k}}}} \times {{{{\bar R}_{q,n}}} \over {\mathop \sum \nolimits_{k = 1}^b {{\bar R}_{q,k}}}}} } $$\end{document} The \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${C_m} - {C_n}$$\end{document} crosstalk score \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${g_s}\left( {m,n} \right)$$\end{document} is evaluated by summing \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$LRI_{p,q,m,n}^t$$\end{document} by Eq. (15): 15\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\matrix{ {{g_s}\left( {m,n} \right) = \mathop \sum \limits_{p,q = 1}^{{z_s}} \,LRI_{p,q,m,n}^s} \cr } $$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${z_s}$$\end{document} is the number of L-R pairs mediating \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${C_m} - {C_n}$$\end{document} signaling through specific expression. Step 3: CCC inference The above three methods evaluate the cellular crosstalk score from three different dimensions. To fully use their advantages, we normalize the computed three scores, and take their mean value as the final crosstalk score by Eq. (16): 16\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\matrix{ {g\left( {m,n} \right) = {1 \over 3}\left( {{g_t} + {g_p} + {g_s}} \right)} \cr } $$\end{document} Results Experimental setup Area Under Receiver Operating Characteristic (ROC) Curve (AUC) and Area Under Precision-Recall (PR) curve (AUPR) are two commonly-used indicators to evaluate the performance of machine learning models [50, 51]. AUC is in the range of [0–1], and the larger the value is, the better the model performance is. AUC with 0.5 denotes that the classification accuracy has no difference with the random guess, and AUC with less than 0.5 denotes the poorer classification accuracy. AUPR is in the range of [0–1], with a value closer to 1 denoting better classification accuracy and 0 denoting the poorer classification accuracy. While AUC is a widely used standard metric for classification performance, it can sometimes be over-optimistic in highly imbalanced datasets. Therefore, we also employed the Area Under the Precision-Recall curve (AUPR), which is recognized as a more informative metric for imbalanced data, to comprehensively evaluate the model performance alongside AUC. So the combination of AUC and AUPR can fully evaluate the model performance in different situations. Using AUC and AUPR as main indictors, we assessed the performance of CellAL and other L-R classification models in addition to precision, recall, accuracy, and F1-score. The CellAL model was compared with four state-of-the-art protein-protein interaction identification models, i.e., ACT-SVM [52] (SVM for simplicity), Decision Tree [53] (DT for simplicity), PPISP-XGBoost [54] (XGBoost for simplicity), and siamese Residual recurrent convolutional neural network (RCNN) [55] (PIPR for simplicity). Table 3 shows their parameters. To ensure the robustness of our model and avoid overfitting, we strictly implemented a rigorous 5-fold cross-validation strategy, repeated 20 times to minimize random bias. In each fold, the dataset was randomly partitioned into training and testing sets, ensuring that the testing data was completely unseen during the model training process to prevent data leakage.Table 3Parameter settingsMethodParameter settings.SVMC = 1.0, kernel = rbf, gamma = scale, cache_size = 200, degree = 3, probability = False, verbose = False, decision_function_shape = ovrXGBoostmin_child_weight = 1, max_delta_step = 0, gamma = 1, subsample = 1, colsample_bytree = 1, max_depth = 5, learning_rate = 0.01, n_estimators = 10PIPRn_estimators = 10, depth = 5, split = 5, neighbours = 3DTcriterion=“gini”, min_samples_split = 2, min_samples_leaf = 1CellALlearning_rate = 0.0005, epochs = 400, beta_1 = 0.9, epsilon = 1e-8, optimizer = optimizer, loss=’binary_crossentropy’, batch_size = 64, verbose= Table 3Parameter settingsMethodParameter settings.SVMC = 1.0, kernel = rbf, gamma = scale, cache_size = 200, degree = 3, probability = False, verbose = False, decision_function_shape = ovrXGBoostmin_child_weight = 1, max_delta_step = 0, gamma = 1, subsample = 1, colsample_bytree = 1, max_depth = 5, learning_rate = 0.01, n_estimators = 10PIPRn_estimators = 10, depth = 5, split = 5, neighbours = 3DTcriterion=“gini”, min_samples_split = 2, min_samples_leaf = 1CellALlearning_rate = 0.0005, epochs = 400, beta_1 = 0.9, epsilon = 1e-8, optimizer = optimizer, loss=’binary_crossentropy’, batch_size = 64, verbose= Parameter settings Performance comparison between CellAL with other methods To estimate the performance of CellAL, we compared it with four classical PPI prediction models, i.e., SVM [52], DT [56], XGBoost [54], and PIPR [55], as well as two state-of-the-art methods, CellEnBoost [23] and CellComNet [57]. SVM [52] mapped a protein sequence to digital features and combined these features as input to SVM for predicting PPIs. DT classified protein-protein pairs by recursively constructing the structure of a decision tree. XGBoost [54] extracted characteristics of proteins through feature extraction methods, preprocessed features to obtain more optimized representations, and used synthetic minority oversampling and nuclear principal component analysis to alleviate class imbalances and remove redundant features, finally feeding these features into XGBoost to uncover PPI sites. PIPR [55] combined the Siamese architecture of deep RCNN to incorporate local features and context information for PPI prediction. CellEnBoost [23] utilizes an ensemble approach—comprising LightGBM, AdaBoost, and CNNs—to predict LRIs, supported by three scoring methods for CCC inference. Similarly, CellComNet [57] leverages deep neural networks and heterogeneous Newton boosting for LRI identification, using a joint scoring scheme to assess CCC intensity. Table 4 and Fig 2 present a comprehensive performance comparison of CellAL against four classical methods (SVM, DT, XGBoost, and PIPR) across four datasets. These illustrations evaluate key metrics including precision, recall, accuracy, F1-score, AUC, and AUPR, along with their average performance across the datasets. As visualized in Fig. 2, CellAL consistently achieved the best performance across these indicators compared to the classical models. Furthermore, Fig 3 characterizes the ROC and PR curves, extending the comparison to include advanced ensemble and deep learning-based methods (CellEnBoost and CellComNet). The results demonstrated that CellAL significantly outperformed the four classical methods across all datasets. When compared with the advanced methods (CellEnBoost and CellComNet), CellAL exhibited highly competitive performance. Notably, CellAL achieved the highest AUPR values on Datasets 1, 2, and 3 (0.8579, 0.8244, and 0.8187, respectively) and the highest AUC on Dataset 4 (0.8366). Since AUPR is a critical metric for evaluating performance on imbalanced biological datasets, these results highlight the robustness of CellAL in identifying true interacting pairs with high precision. In summary, CellAL proved to have strong LRI prediction ability, comparable to or surpassing state-of-the-art methods in key performance indicators. Therefore, its predicted LRIs can be reliably used for CCC inference.Table 4Performance comparison between CellAL with SVM, DT, XGBoost, PIPR on four LRI datasetsIndexDatasetSVMDTXGBoostPIPRCellADataset10.5989 ± 0.00320.6676 ± 0.01270.7685 ± 0.00490.7203 ± 0.00780.7883+0.0030Dataset20.6169 ± 0.00350.6444 ± 0.02190.7316 ± 0.00550.6856 ± 0.00900.7427+0.0057PrecisionDataset30.6261 ± 0.00290.6304 ± 0.01970.7433 ± 0.00840.6893 ± 0.01010.7353+0.0070Dataset40.6398 ± 0.00170.6434 ± 0.00880.7606 ± 0.00410.7399 ± 0.00720.7649+0.0047AVE.0.62040.64650.75100.70880.7578Dataset10.5495 ± 0.00410.6929 ± 0.01980.6616 ± 0.00710.7403 ± 0.01400.7688+0.0051Dataset20.5350 ± 0.00390.6639 ± 0.02750.6526 ± 0.00600.7068 ± 0.01880.7396+0.0083recallDataset30.5415 ± 0.00380.6579 ± 0.02500.5941 ± 0.00860.7068 ± 0.01720.7384+0.0086Dataset40.5584 ± 0.00260.6634 ± 0.01240.6150 ± 0.00360.7666 ± 0.01110.7800+0.0047AVE.0.54610.66950.63080.73010.7567Dataset10.5906 ± 0.00260.6738 ± 0.01180.7310 ± 0.00370.7256 ± 0.00550.7810+0.0024Dataset20.6014 ± 0.00280.6487 ± 0.02220.7064 ± 0.00440.6907 ± 0.00610.7414+0.0055AccuracyDataset30.6088 ± 0.00230.6359 ± 0.01930.6942 ± 0.00600.6934 ± 0.00700.7366+0.0046Dataset40.6220 ± 0.00150.6477 ± 0.00800.7106 ± 0.00270.7479 ± 0.00360.7700+0.0029AVE.0.60570.65150.71060.71440.7573Dataset10.5727 ± 0.00280.6798 ± 0.01260.7107 ± 0.00460.7292 ± 0.00610.7783+0.0028Dataset20.5728 ± 0.00330.6538 ± 0.02260.6896 ± 0.00480.6949 ± 0.00790.7408+0.0060F1-scoreDataset30.5804 ± 0.00280.6437 ± 0.01920.6598 ± 0.00710.6969 ± 0.00790.7361+0.0047Dataset40.5962 ± 0.00190.6531 ± 0.00830.6799 ± 0.00300.7524 ± 0.00380.7723+0.0026AVE.0.58050.65760.68500.71840.7569Dataset10.6251 ± 0.00170.6738 ± 0.01180.7944 ± 0.00380.7910 ± 0.00490.8398+0.0034Dataset20.6283 ± 0.00170.6487 ± 0.02220.7753 ± 0.00420.7605 ± 0.00630.8063+0.0048AUCDataset30.6322 ± 0.00090.6359 ± 0.01930.7596 ± 0.00400.7580 ± 0.00600.7960+0.0045Dataset40.6662 ± 0.00070.6477 ± 0.00800.7656 ± 0.00180.8222 ± 0.00360.8366+0.0021AVE.0.63790.65150.77370.78290.8197Dataset10.6354 ± 0.00160.7570 ± 0.00880.8086 ± 0.00350.7796 ± 0.00590.8579+0.0033Dataset20.6257 ± 0.00210.7382 ± 0.01640.7824 ± 0.00480.7515 ± 0.00690.8244+0.0038AUPRDataset30.6452 ± 0.00140.7297 ± 0.01400.7762 ± 0.00430.7494 ± 0.00610.8187+0.0051Dataset40.6727 ± 0.00100.7375 ± 0.00590.7868 ± 0.00180.8176 ± 0.00450.8500+0.0026AVE.0.64480.74060.78850.77450.8378Fig. 2The performance comparison of SVM, DT, XGBoost, PIPR, and CellAL on four datasets (A-D: Dataset 1 to 4)Fig. 3The ROC and PR curves obtained from SVM, DT, XGBoost, PIPR, and CellAL on four LRI datasets. A-D: the ROC curves from Dataset 1 to 4. E-H: the PR curves from Dataset 1 to 4 Table 4Performance comparison between CellAL with SVM, DT, XGBoost, PIPR on four LRI datasetsIndexDatasetSVMDTXGBoostPIPRCellADataset10.5989 ± 0.00320.6676 ± 0.01270.7685 ± 0.00490.7203 ± 0.00780.7883+0.0030Dataset20.6169 ± 0.00350.6444 ± 0.02190.7316 ± 0.00550.6856 ± 0.00900.7427+0.0057PrecisionDataset30.6261 ± 0.00290.6304 ± 0.01970.7433 ± 0.00840.6893 ± 0.01010.7353+0.0070Dataset40.6398 ± 0.00170.6434 ± 0.00880.7606 ± 0.00410.7399 ± 0.00720.7649+0.0047AVE.0.62040.64650.75100.70880.7578Dataset10.5495 ± 0.00410.6929 ± 0.01980.6616 ± 0.00710.7403 ± 0.01400.7688+0.0051Dataset20.5350 ± 0.00390.6639 ± 0.02750.6526 ± 0.00600.7068 ± 0.01880.7396+0.0083recallDataset30.5415 ± 0.00380.6579 ± 0.02500.5941 ± 0.00860.7068 ± 0.01720.7384+0.0086Dataset40.5584 ± 0.00260.6634 ± 0.01240.6150 ± 0.00360.7666 ± 0.01110.7800+0.0047AVE.0.54610.66950.63080.73010.7567Dataset10.5906 ± 0.00260.6738 ± 0.01180.7310 ± 0.00370.7256 ± 0.00550.7810+0.0024Dataset20.6014 ± 0.00280.6487 ± 0.02220.7064 ± 0.00440.6907 ± 0.00610.7414+0.0055AccuracyDataset30.6088 ± 0.00230.6359 ± 0.01930.6942 ± 0.00600.6934 ± 0.00700.7366+0.0046Dataset40.6220 ± 0.00150.6477 ± 0.00800.7106 ± 0.00270.7479 ± 0.00360.7700+0.0029AVE.0.60570.65150.71060.71440.7573Dataset10.5727 ± 0.00280.6798 ± 0.01260.7107 ± 0.00460.7292 ± 0.00610.7783+0.0028Dataset20.5728 ± 0.00330.6538 ± 0.02260.6896 ± 0.00480.6949 ± 0.00790.7408+0.0060F1-scoreDataset30.5804 ± 0.00280.6437 ± 0.01920.6598 ± 0.00710.6969 ± 0.00790.7361+0.0047Dataset40.5962 ± 0.00190.6531 ± 0.00830.6799 ± 0.00300.7524 ± 0.00380.7723+0.0026AVE.0.58050.65760.68500.71840.7569Dataset10.6251 ± 0.00170.6738 ± 0.01180.7944 ± 0.00380.7910 ± 0.00490.8398+0.0034Dataset20.6283 ± 0.00170.6487 ± 0.02220.7753 ± 0.00420.7605 ± 0.00630.8063+0.0048AUCDataset30.6322 ± 0.00090.6359 ± 0.01930.7596 ± 0.00400.7580 ± 0.00600.7960+0.0045Dataset40.6662 ± 0.00070.6477 ± 0.00800.7656 ± 0.00180.8222 ± 0.00360.8366+0.0021AVE.0.63790.65150.77370.78290.8197Dataset10.6354 ± 0.00160.7570 ± 0.00880.8086 ± 0.00350.7796 ± 0.00590.8579+0.0033Dataset20.6257 ± 0.00210.7382 ± 0.01640.7824 ± 0.00480.7515 ± 0.00690.8244+0.0038AUPRDataset30.6452 ± 0.00140.7297 ± 0.01400.7762 ± 0.00430.7494 ± 0.00610.8187+0.0051Dataset40.6727 ± 0.00100.7375 ± 0.00590.7868 ± 0.00180.8176 ± 0.00450.8500+0.0026AVE.0.64480.74060.78850.77450.8378 Performance comparison between CellAL with SVM, DT, XGBoost, PIPR on four LRI datasets The performance comparison of SVM, DT, XGBoost, PIPR, and CellAL on four datasets (A-D: Dataset 1 to 4) The ROC and PR curves obtained from SVM, DT, XGBoost, PIPR, and CellAL on four LRI datasets. A-D: the ROC curves from Dataset 1 to 4. E-H: the PR curves from Dataset 1 to 4 The selection of L-R feature dimension To study the effect of the L-R feature dimension parameter d on LRI classification performance, we calculated the performance of CellAL with dimension parameter d in the range [200–800] on four datasets. As shown in Table 5 and Fig. 4, the performance of CellAL showed an upward trend from 200 to 600, and almost stabilized after reaching 800. when d is set to 600, it computed better performance in individual cases. When d is set to 800, its performance was best in most situations. Considering the factors such as computing resources, computing time and computing complexity, d was finally set to 800.Table 5Performance of CellAL with dimension of range from 200 to 800 on four datasetsindex200300400500600800Dataset1precision0.7607 ± 0.00550.7741 ± 0.00480.7840 ± 0.00460.7875 ± 0.00520.7859 ± 0.00380.7883 ± 0.0030recall0.7550 ± 0.00670.7575 ± 0.00790.7604 ± 0.00590.7648 ± 0.00520.7676 ± 0.00620.7688 ± 0.0051accuracy0.7585 ± 0.00380.7680 ± 0.00380.7754 ± 0.00390.7791 ± 0.00320.7791 ± 0.00330.7810 ± 0.0024F1 score0.7576 ± 0.00380.7655 ± 0.00430.7719 ± 0.00420.7758 ± 0.00300.7765 ± 0.00370.7783 ± 0.0028AUC0.8165 ± 0.00440.8252 ± 0.00360.8329 ± 0.00330.8360 ± 0.00230.8346 ± 0.00370.8398 ± 0.0034AUPR0.8318 ± 0.00380.8421 ± 0.00260.8508 ± 0.00310.8541 ± 0.00210.8531 ± 0.00310.8579 ± 0.0033Dataset2precision0.7235 ± 0.00640.7264 ± 0.00630.7297 ± 0.00580.7363 ± 0.00710.7360 ± 0.00520.7427 ± 0.0057recall0.7242 ± 0.00900.7289 ± 0.00820.7274 ± 0.01040.7336 ± 0.00610.7352 ± 0.01140.7396 ± 0.0083accuracy0.7234 ± 0.00580.7268 ± 0.00520.7288 ± 0.00490.7352 ± 0.00570.7355 ± 0.00550.7414 ± 0.0055F1 score0.7236 ± 0.00610.7273 ± 0.00540.7283 ± 0.00570.7347 ± 0.00540.7353 ± 0.00680.7408 ± 0.0060AUC0.7797 ± 0.00460.7848 ± 0.00520.7883 ± 0.00430.7970 ± 0.00460.7989 ± 0.00430.8063 ± 0.0048AUPR0.7986 ± 0.00420.8037 ± 0.00440.8066 ± 0.00370.8146 ± 0.00430.8169 ± 0.00390.8244 ± 0.0038Dataset3precision0.7161 ± 0.00700.7241 ± 0.00680.7313 ± 0.00680.7344 ± 0.00710.7372 ± 0.00610.7353 ± 0.0070recall0.7208 ± 0.00730.7328 ± 0.00690.7344 ± 0.00640.7342 ± 0.00790.7377 ± 0.00700.7384 ± 0.0086accuracy0.7173 ± 0.00590.7265 ± 0.00510.7321 ± 0.00510.7340 ± 0.00500.7371 ± 0.00530.7361 ± 0.0047F1 score0.7182 ± 0.00560.7281 ± 0.00460.7326 ± 0.00470.7340 ± 0.00480.7372 ± 0.00540.7366 ± 0.0046AUC0.7687 ± 0.00590.7797 ± 0.00460.7881 ± 0.00480.7922 ± 0.00450.7963 ± 0.00480.7960 ± 0.0045AUPR0.7935 ± 0.00490.8036 ± 0.00400.8108 ± 0.00440.8146 ± 0.00420.8184 ± 0.00500.8187 ± 0.0051Dataset4precision0.7372 ± 0.00610.7515 ± 0.00410.7600 ± 0.00390.7611 ± 0.00460.7668 ± 0.00460.7649 ± 0.0047recall0.7377 ± 0.00700.7549 ± 0.00530.7660 ± 0.00470.7788 ± 0.00460.7770 ± 0.00530.7800 ± 0.0047accuracy0.7371 ± 0.00530.7525 ± 0.00310.7618 ± 0.00250.7670 ± 0.00250.7702 ± 0.00220.7700 ± 0.0029F1 score0.7372 ± 0.00540.7530 ± 0.00320.7628 ± 0.00240.7697 ± 0.00200.7717 ± 0.00200.7723 ± 0.0026AUC0.7963 ± 0.00480.8150 ± 0.00260.8238 ± 0.00190.8329 ± 0.00260.8338 ± 0.00290.8366 ± 0.0021AUPR0.8184 ± 0.00500.8292 ± 0.00300.8395 ± 0.00280.8459 ± 0.00390.8498 ± 0.00140.8500 ± 0.0026Fig. 4Performance of CellAL under different dimensions. (A) AUC and (B) AUPR Table 5Performance of CellAL with dimension of range from 200 to 800 on four datasetsindex200300400500600800Dataset1precision0.7607 ± 0.00550.7741 ± 0.00480.7840 ± 0.00460.7875 ± 0.00520.7859 ± 0.00380.7883 ± 0.0030recall0.7550 ± 0.00670.7575 ± 0.00790.7604 ± 0.00590.7648 ± 0.00520.7676 ± 0.00620.7688 ± 0.0051accuracy0.7585 ± 0.00380.7680 ± 0.00380.7754 ± 0.00390.7791 ± 0.00320.7791 ± 0.00330.7810 ± 0.0024F1 score0.7576 ± 0.00380.7655 ± 0.00430.7719 ± 0.00420.7758 ± 0.00300.7765 ± 0.00370.7783 ± 0.0028AUC0.8165 ± 0.00440.8252 ± 0.00360.8329 ± 0.00330.8360 ± 0.00230.8346 ± 0.00370.8398 ± 0.0034AUPR0.8318 ± 0.00380.8421 ± 0.00260.8508 ± 0.00310.8541 ± 0.00210.8531 ± 0.00310.8579 ± 0.0033Dataset2precision0.7235 ± 0.00640.7264 ± 0.00630.7297 ± 0.00580.7363 ± 0.00710.7360 ± 0.00520.7427 ± 0.0057recall0.7242 ± 0.00900.7289 ± 0.00820.7274 ± 0.01040.7336 ± 0.00610.7352 ± 0.01140.7396 ± 0.0083accuracy0.7234 ± 0.00580.7268 ± 0.00520.7288 ± 0.00490.7352 ± 0.00570.7355 ± 0.00550.7414 ± 0.0055F1 score0.7236 ± 0.00610.7273 ± 0.00540.7283 ± 0.00570.7347 ± 0.00540.7353 ± 0.00680.7408 ± 0.0060AUC0.7797 ± 0.00460.7848 ± 0.00520.7883 ± 0.00430.7970 ± 0.00460.7989 ± 0.00430.8063 ± 0.0048AUPR0.7986 ± 0.00420.8037 ± 0.00440.8066 ± 0.00370.8146 ± 0.00430.8169 ± 0.00390.8244 ± 0.0038Dataset3precision0.7161 ± 0.00700.7241 ± 0.00680.7313 ± 0.00680.7344 ± 0.00710.7372 ± 0.00610.7353 ± 0.0070recall0.7208 ± 0.00730.7328 ± 0.00690.7344 ± 0.00640.7342 ± 0.00790.7377 ± 0.00700.7384 ± 0.0086accuracy0.7173 ± 0.00590.7265 ± 0.00510.7321 ± 0.00510.7340 ± 0.00500.7371 ± 0.00530.7361 ± 0.0047F1 score0.7182 ± 0.00560.7281 ± 0.00460.7326 ± 0.00470.7340 ± 0.00480.7372 ± 0.00540.7366 ± 0.0046AUC0.7687 ± 0.00590.7797 ± 0.00460.7881 ± 0.00480.7922 ± 0.00450.7963 ± 0.00480.7960 ± 0.0045AUPR0.7935 ± 0.00490.8036 ± 0.00400.8108 ± 0.00440.8146 ± 0.00420.8184 ± 0.00500.8187 ± 0.0051Dataset4precision0.7372 ± 0.00610.7515 ± 0.00410.7600 ± 0.00390.7611 ± 0.00460.7668 ± 0.00460.7649 ± 0.0047recall0.7377 ± 0.00700.7549 ± 0.00530.7660 ± 0.00470.7788 ± 0.00460.7770 ± 0.00530.7800 ± 0.0047accuracy0.7371 ± 0.00530.7525 ± 0.00310.7618 ± 0.00250.7670 ± 0.00250.7702 ± 0.00220.7700 ± 0.0029F1 score0.7372 ± 0.00540.7530 ± 0.00320.7628 ± 0.00240.7697 ± 0.00200.7717 ± 0.00200.7723 ± 0.0026AUC0.7963 ± 0.00480.8150 ± 0.00260.8238 ± 0.00190.8329 ± 0.00260.8338 ± 0.00290.8366 ± 0.0021AUPR0.8184 ± 0.00500.8292 ± 0.00300.8395 ± 0.00280.8459 ± 0.00390.8498 ± 0.00140.8500 ± 0.0026 Performance of CellAL with dimension of range from 200 to 800 on four datasets Performance of CellAL under different dimensions. (A) AUC and (B) AUPR LRI validation In CellAL, L-R pairs with an interacting probability higher than 0.999 were viewed as potential interactions. Ultimately, 82,421, 64,081, 19,602, and 233,757 potential LRIs were predicted using CellAL on four datasets, respectively. Next, we analyzed the overlapping L-R pairs between known and predicted LRIs with existing four classical LRI databases, i.e., CellChat [21], Connectome [58], CytoTalk [59], and NATMI [60], respectively. Table 6 and Fig. 5 list the number of LRIs that overlap between CellAL and each of the above databases. The results showed that there were 596, 536, 436, and 1overlapping LRIs between CellAL and CellChat, 2325, 1334, 1957, and 7 overlapping LRIs between CellAL and Connectome, 1789, 1172, 1533, and 0 overlapping LRIs between CellAL and CytoTalk on the four datasets, respectively. Since NATMI only provides human LRIs, the overlap analysis for Datasets 2 and 3 (mouse) was excluded. Consequently, we identified 1877 and 4 overlapping LRIs between CellAL and NATMI on Datasets 1 and 4, respectively. In summary, CellAL has a high overlapping rate with the above four databases except for Dataset 4. In particular, Dataset 1 has particularly obvious overlapping rate between CellAL and other databases. Therefore, we used Dataset 1 for CCC analysis in the tumor microenvironmentTable 6The overlapping LRI number between CellAL and CellChat, Connectome, CytoTalk and NATMI on 4 LRI datasetsDatasetCellChatConnectomeCytoTalkNATMI1596232517891877253613041172/343619571533/417/7Fig. 5The Venn diagram for depicting the overlapping LRIs between CellAL and CellChat, Connectome, CytoTalk, and NATMI on Dataset 1 to 4 Table 6The overlapping LRI number between CellAL and CellChat, Connectome, CytoTalk and NATMI on 4 LRI datasetsDatasetCellChatConnectomeCytoTalkNATMI1596232517891877253613041172/343619571533/417/7 The overlapping LRI number between CellAL and CellChat, Connectome, CytoTalk and NATMI on 4 LRI datasets The Venn diagram for depicting the overlapping LRIs between CellAL and CellChat, Connectome, CytoTalk, and NATMI on Dataset 1 to 4 CCC inference within melanoma Recently, the incidence of melanoma has continuously risen in human with white skin in developed countries. Melanoma accounts for 1.7% of cancer diagnoses worldwide. In the United States, it ranks the fifth among the most common cancers [61, 62]. Since 1975, its incidence in the United States has increased by over 320% [63]. scRNA-seq has achieved unprecedented cellular resolution at the whole genome scale and provides abundant gene expression information, enabling us to explore cellular heterogeneity of tumors [64]. In this section, we used CellAL to systematically analyze cellular crosstalk within human melanoma. To estimate cellular crosstalk intensity within melanoma, we conducted a systematic analysis for LRI-based CCC in the melanoma tissue. First, its scRNA-seq information were achieved from the GEO database (entry code: GSE72056) [65]. We utilized the cell type annotations provided by the original study. To address dropouts, gene expression was averaged per cell type. A ligand-receptor pair was considered potentially interacting only if the cluster-specific expression of both proteins exceeded their respective global mean expression plus one standard deviation (dynamic thresholding), ensuring robustness against low-expression noise. The data include seven cell types: melanoma malignant cells, T cells, B cells, macrophages (Macro), endothelial cells (Endo), CAFs, and natural killer (NK) cells. Next, the high-confident L-R pairs predicted by CellAL and known LRIs on Dataset 1 were filtered in combination with its scRNA-seq data. For an L-R pair, when its ligand and receptor are expressed at the same time in the tissue, the pair was retained, otherwise it was removed. Third, CellAL constructed the cellular crosstalk network involving the seven cell types by combining a comprehensive scoring strategy based on expression threshold, expression product, and expression specificity. The L-R pairs mediating cellular crosstalk were ranked and the top three L-R pairs were obtained. Table 7 shows the ranking of crosstalk between malignant cells and the six cell types calculated by expression threshold, expression product, expression specificity, and the comprehensive scoring strategies in human melanoma, and the top 3 L-R pairs/the number of L-R pairs mediating signaling from melanoma cells to other cell types, and the top 3 LRIs/the number of LRIs mediating signaling from other cell types to melanoma cells. CellAL predicted that CAFs, macrophages, and endothelial cells had the strongest crosstalk with human melanoma cells, which were the same as the results based on expression threshold and expression product. Moreover, crosstalk from melanoma cells to CAFs was mediated by L-R pairs FGF4-CANX, VIM-CANX, and B2M-CANX. Crosstalk from CAFs to melanoma cells was mediated by L-R pairs SHANK1-CANX, B2M-CANX, and VIM-CANX. There is growing evidence that crosstalk between malignant cells and non-malignant cells causes tumor aggression, metastasis, and drug resistance. CAFs play a key role in stromal cells [66]. Bidirectional crosstalk between CAFs and melanoma cells promotes their growth and invasion [67]. Melanoma cells secrete diffusion factors to promote CAF proliferation. CAFs can migrate or invade independently of melanoma, providing an effective way for melanoma diagnosis and treatment [68]. Metastasis, growth and invasion were symbiotic development of CAFs and melanoma [69]. CAFs enhance the resistance to BRAF-targeted therapy through integrin/FAK signaling [70]. The findings suggested that melanoma-CAFS communication may be crucial to the growth, invasion, and progression of melanoma.Table 7Cellular crosstalk prediction results based on four different scoring approaches within melanomaFunctionCell type“outcoming” LRI“incoming” LRIsThresholdingCAFs613Endothelial492Macrophages214B cells100T cells00NK cells00ProductCAFsVIM-CANX, B2M-CANX, SERPINE2-CANXB2M-CANX, VIM-CANX, TIMP1-CANXEndothelialB2M-HLA-F, VIM-CANX, HLA-A-APLP2B2M-CANX, APP-RPSA, VIM-CANXMacrophagesVIM-CD53, B2M-CD53, SERPINE2–CD53B2M-CANX, B2M-ADIPOR1, PSAP-CANXNK cellsVIM-CD53, B2M-CD247, LGALS1-PTPRCB2M-CANX, B2M-ADIPOR1, HLA-C-CANXT cellsLGALS1-PTPRC, VIM-CD53, B2M-HLA-FB2M-CANX, B2M-ADIPOR1, HLA-A-CANXB cellsVIM-CD79A, B2M-CD79A, SERPINE2–CD79AB2M-CANX, B2M-CANX, HLA-C-CANXSpecificT cellsFGF4–CD53, OBP2A-CD53, IL9–IL2RGFGF21-CANX, FGF21–ADIPOR1, INS-CANXB cellsFGF4–CD53, OBP2A-CD79A, OBP2A-CD53NTN3-CANX, NTN3–ADIPOR1, NTN3–CD47MacrophagesFGF4–CD53, OBP2A-CD53, FGF4–TNFRSF1BLACRT-CANX, LACRT-ADIPOR1, LACRT-CD46CAFsFGF4-CANX, OBP2A-CANX, FGF4–LPAR1SHANK1-CANX, SHANK1–ADIPOR1, SHANK1–CD47NK cellsFGF4–CD53, OBP2A-CD53, IL9–CD53SCT-SCTR, FGF4–FGFR3, WNT11–KLRG2EndothelialFGF4–CD53, OBP2A-CD53, FGF4–TNFRSF1BAMH-CANX, AMH-ADIPOR1, FGF6-CANXComprehensive scoringCAFsFGF4-CANX, VIM-CANX, B2M-CANXSHANK1-CANX, B2M-CANX, VIM-CANXMacrophagesFGF4–CD53, VIM-CD53, B2M-CD53LACRT-CANX, B2M-CANX, LACRT-ADIPOR1EndothelialFGF4–CD53, B2M-HLA-F, VIM-CANXAMH-CANX, B2M-CANX, APP-RPSAT cellsFGF4–CD53,LGALS1-PTPRC,VIM-CD53FGF21-CANX, B2M-CANX, FGF21–ADIPOR1B cellsFGF4–CD53, VIM-CD79A, OBP2A-CD79ANTN3-CANX, B2M-CANX, NTN3–ADIPOR1NK cellsFGF4–CD53, VIM-CD53, B2M-CD247CER1-CANX, B2M-CANX, TFF1-CANX Table 7Cellular crosstalk prediction results based on four different scoring approaches within melanomaFunctionCell type“outcoming” LRI“incoming” LRIsThresholdingCAFs613Endothelial492Macrophages214B cells100T cells00NK cells00ProductCAFsVIM-CANX, B2M-CANX, SERPINE2-CANXB2M-CANX, VIM-CANX, TIMP1-CANXEndothelialB2M-HLA-F, VIM-CANX, HLA-A-APLP2B2M-CANX, APP-RPSA, VIM-CANXMacrophagesVIM-CD53, B2M-CD53, SERPINE2–CD53B2M-CANX, B2M-ADIPOR1, PSAP-CANXNK cellsVIM-CD53, B2M-CD247, LGALS1-PTPRCB2M-CANX, B2M-ADIPOR1, HLA-C-CANXT cellsLGALS1-PTPRC, VIM-CD53, B2M-HLA-FB2M-CANX, B2M-ADIPOR1, HLA-A-CANXB cellsVIM-CD79A, B2M-CD79A, SERPINE2–CD79AB2M-CANX, B2M-CANX, HLA-C-CANXSpecificT cellsFGF4–CD53, OBP2A-CD53, IL9–IL2RGFGF21-CANX, FGF21–ADIPOR1, INS-CANXB cellsFGF4–CD53, OBP2A-CD79A, OBP2A-CD53NTN3-CANX, NTN3–ADIPOR1, NTN3–CD47MacrophagesFGF4–CD53, OBP2A-CD53, FGF4–TNFRSF1BLACRT-CANX, LACRT-ADIPOR1, LACRT-CD46CAFsFGF4-CANX, OBP2A-CANX, FGF4–LPAR1SHANK1-CANX, SHANK1–ADIPOR1, SHANK1–CD47NK cellsFGF4–CD53, OBP2A-CD53, IL9–CD53SCT-SCTR, FGF4–FGFR3, WNT11–KLRG2EndothelialFGF4–CD53, OBP2A-CD53, FGF4–TNFRSF1BAMH-CANX, AMH-ADIPOR1, FGF6-CANXComprehensive scoringCAFsFGF4-CANX, VIM-CANX, B2M-CANXSHANK1-CANX, B2M-CANX, VIM-CANXMacrophagesFGF4–CD53, VIM-CD53, B2M-CD53LACRT-CANX, B2M-CANX, LACRT-ADIPOR1EndothelialFGF4–CD53, B2M-HLA-F, VIM-CANXAMH-CANX, B2M-CANX, APP-RPSAT cellsFGF4–CD53,LGALS1-PTPRC,VIM-CD53FGF21-CANX, B2M-CANX, FGF21–ADIPOR1B cellsFGF4–CD53, VIM-CD79A, OBP2A-CD79ANTN3-CANX, B2M-CANX, NTN3–ADIPOR1NK cellsFGF4–CD53, VIM-CD53, B2M-CD247CER1-CANX, B2M-CANX, TFF1-CANX Cellular crosstalk prediction results based on four different scoring approaches within melanoma In vitro study suggested that activated macrophages can kill cancer cells [71]. In many cases, macrophages induce specific normal cells to transform into cancer cells [72]. These differences may be caused by different macrophage phenotypes [73]. Activated macrophages have crucial roles in chronic inflammation. In addition, tumor-associated macrophages are also involved to cancer cell invasion, angiogenesis, metastasis, and adaptive anti-tumor immunosuppression [73]. Mutual growth factor exchange between endothelial cells and malignant cells in anoxic microenvironment determines tumor progression [74]. Under normal physiological conditions, low level of constitutive expression of Vascular Endothelial Growth Factor (VEGF) may regulate skin vascular function. During melanoma occurrence, progression, and spread, VEGF plays a fundamental role in the angiogenic response required for melanoma growth and metastasis. The high permeability of VEGF-induced neovascularization may be an important factor in the melanoma metastasis [57]. The expression of VEGF in tumor infiltrating inflammatory cells suggested that these cells can enhance the angiogenesis and progression of malignant melanoma. Consequently, VEGF was a potential target against anti-angiogenic tumor therapy in melanoma. Comparison with other CCC inference results To further investigate the accuracy of the CellAL model, we compared it with three cellular crosstalk prediction tools, i.e., CellPhoneDB [18], iTALK [75], and CellChat [21], within melanoma tumor microenvironment. CellPhoneDB [18] comprehensively considered the subunit structures of L-R pairs, accurately represented the heteropolymer complex, and used scRNA-seq data to predict crosstalk between two cell types. iTALK [75] used scRNA seq data to characterize and interpret cellular communication signals in multicellular tumor ecosystem. It dissected the complexity, diversity, and dynamics of cellular communication. CellChat [21] enabled CCC quantitative inference and analysis. CellChat classified signaling pathways, and characterized conserved and environment-specific pathways on different datasets. Table 8 shows the cellular communication prediction results from CellAL, CellPhoneDB, iTALK, and CellChat within human melanoma tissue. CellAL predicted the strongest crosstalk between CAFs and melanoma cells, which was the same as those predicted by iTALK and CellChat. Moreover, macrophages and endothelial cells may have signaling pathways with melanoma cells, which was validated by iTALK. In summary, CellAL obtained highly similar cellular crosstalk prediction results to those of CellPhoneDB, iTALK, and CellChat, also indicating the accuracy of CellAL predictions.Table 8Comparison with existing methods in human melanoma tissueRankCellPhoneDBiTALKCellChatCellAL1MacrophagesMacrophagesCAFsCAFs2Endothelial cellsEndothelial cellsMacrophagesMacrophages3CAFsCAFsEndothelial cellsEndothelial cells4T cellsT cellsNK cellsT cells5B cellsB cellsT cellsB cells6NK cellsNK cellsB cellsNK cells Table 8Comparison with existing methods in human melanoma tissueRankCellPhoneDBiTALKCellChatCellAL1MacrophagesMacrophagesCAFsCAFs2Endothelial cellsEndothelial cellsMacrophagesMacrophages3CAFsCAFsEndothelial cellsEndothelial cells4T cellsT cellsNK cellsT cells5B cellsB cellsT cellsB cells6NK cellsNK cellsB cellsNK cells Comparison with existing methods in human melanoma tissue Result visualization To more intuitively display communication between melanoma cells and other cell types, CellAL visualized the predictions from three different views: (i) Network view. Each side of a curve represented a cell type, and different cell types were represented by different colors. Communication between two cell types was depicted by a curve, the width of which is proportional to cellular communication strength. (ii) Heatmap view. Using the sending cell and the receiving cell as the horizontal and vertical coordinates, the darker color of a cell denotes the stronger communication. (iii) The top 3 L-R pairs. The ordinate represents the top L-R pairs mediating communication. The depth of color on the horizontal and vertical coordinates indicates the intensity of L-R-mediated communication, and the darker color denotes the stronger mediating ability. Figure 6A-D shows cellular communication based on four different scoring methods in melanoma, respectively. CAFs, endothelial cells, and macrophages were the top 3 cell types communicating with melanoma cells based on expression threshold, expression product and CellAL. The corresponding top 3 L-R pairs were listed, where VIM-CANX, B2M-CANX and B2M-HLA-Aplay significant roles in crosstalk between melanoma cells and multiple cell types. Since the expression threshold method can only determine whether an L-R pair mediates communication between two cell types, the final crosstalk score can’t be calculated, so its top 3 L-R pairs were not listed.Fig. 6Melanoma-related CCC results based on four different scoring methods Melanoma-related CCC results based on four different scoring methods Figure 7 A-D shows heatmaps of cellular crosstalk in CellAL, CellPhoneDB, iTALK, and CellChat within human melanoma tissue. Each intersection cell indicates the crosstalk communication intensity. The more blue color denotes stronger communication. The results demonstrated that CAFs, macrophages, and endothelial cells strongly communicate with melanoma cells. Figure 7 E, We conducted GO functional enrichment analysis on L-R pairs that mediate cellular communication. The results showed that it was significantly enriched in immune-related pathways, such as “antigen receptor-mediated signaling pathway” and “positive regulation of T cell mediated cytotoxicity”. It is suggested that these molecules may play key roles in T cell activation and immune killing. Moreover, the enrichment of “cell population proliferation” and “glucose homeostasis” further indicated their potential regulatory functions in the processes of immunometabolism and cell expansion [76, 77]. Furthermore, L-R pairs were also enriched in “laminin binding” and “COPII-coated ER to Golgi transport vesicles”, which validated their characteristics of secretory and extra cellular matrix interactions, and further revealed their roles in cell migration and tissue microenvironment [78].Fig. 7Heatmap view of CCCC within human melanoma tissue. (A) CellAL (B) CellPhoneDB (C) iTALK (D) CellChat (E) GO functional enrichment analysis Heatmap view of CCCC within human melanoma tissue. (A) CellAL (B) CellPhoneDB (C) iTALK (D) CellChat (E) GO functional enrichment analysis Conclusion In this manuscript, we introduced CellAL, a deep learning framework integrating LSTM networks with an attention mechanism to predict LRIs from sequence-derived biochemical features. By effectively capturing long-range dependencies, CellAL transcends the reliance on static databases inherent in traditional methods, enabling the discovery of novel interactions. Our extensive evaluations confirm that CellAL not only significantly outperforms classical machine learning models but also achieves superior AUPR scores compared to state-of-the-art methods like CellEnBoost and CellComNet, demonstrating exceptional robustness in handling class-imbalanced biological data. The framework’s practical utility was further established by deciphering complex crosstalk between malignant cells and stromal components (CAFs and macrophages) in human melanoma tissue. In this context, while our aggregated scoring strategy does not quantify statistical probability relative to a random null model as permutation-based methods do, it explicitly prioritizes biological relevance by balancing signal magnitude with specificity. This approach enables the robust ranking of potential interactions, particularly for the novel pairs identified by CellAL. However, the current model’s exclusive reliance on sequence features constitutes a limitation, as it may overlook spatial constraints and conformational dynamics determined by 3D protein structures. Consequently, future work will focus on integrating structural information and multi-omics data to further enhance prediction accuracy and provide deeper biological interpretability for exploring tumor microenvironments