DyAb: sequence-based antibody design and property prediction in a low-data regime Authors: Joshua Yao-Yu Lin, Jennifer L Hofmann, Andrew Leaver-Fay, Wei-Ching Liang, Stefania Vasilaki, Edith Lee, Pedro O Pinheiro, Natasa Tagasovska, James R Kiefer, Yan Wu, Franziska Seeger, Richard Bonneau, Vladimir Gligorijevic, Andrew Watkins, Kyunghyun Cho, Nathan Frey Published: 2026 ABSTRACT ABSTRACTProtein therapeutic design and property prediction are frequently hampered by data scarcity. Here we propose a model, DyAb, that addresses these issues by leveraging a pair-wise representation to predict differences in binding affinity, rather than absolute values. DyAb is built on top of a pre-trained protein language model and achieves a Spearman rank correlation of up to 0.85 on binding affinity prediction across monoclonal antibodies targeting three different antigens (EGFR, IL-6, and an internal target), given as few as 100 training data. We employ DyAb in two design contexts: as a ranking model to score combinations of known mutations, and combined with a genetic algorithm to generate new sequences. Our method consistently generates antibody variants with high binding rates, including designs that improve on the binding affinity of the lead molecule by more than ten-fold. DyAb represents a powerful tool for optimizing antibody binding affinity in low data regimes common in early-stage drug development.KEYWORDS: Affinity maturation, antibody engineering, machine learning Labeled data scarcity is foremost among the challenges for applying machine learning to biologic drug development. While among the best selling pharmaceuticals worldwide,1,2 data generation for biologics is particularly costly and laborious. Such datasets often contain inherent noise and variation across batches3 and require careful quality control. Learning from sparse, noisy datasets is thus an ongoing challenge in early-stage drug discovery and lead optimization.4 Moreover, the ‘best’ assessed candidate at early stages may need to be discarded down the line if a liability is discovered – e.g., chemical or physical instabilities, poor pharmacokinetics, or immunogenicity.4–6 An overarching objective within this paradigm is thus to produce and refine a select cohort of the most promising therapeutic candidates as quickly and cheaply as possible. Distinguishing between differences in measurements, rather than merely focusing on absolute values, offers potential to address data scarcity. For example, Fralish et al. developed DeepDelta,7 a deep learning model that predicts differences in pharmacokinetic properties between pairs of small molecules. By learning on concatenated representations of molecular structure pairs, their method showed good performance on small datasets containing ∼ 200–1000 labeled points. For protein fitness, a pair-wise method can provide multiple benefits, including preventing over-fitting and batch effects by combinatorially expanding the training dataset and potentially improving out-of-distribution performance. Rather than learning on N raw values, a model would instead learn on N2 =N(N−1)/2 pairs. For example, a dataset of 500 points would swell to one containing nearly 125,000 pairs of data. Here, we show that a similar approach can be applied to therapeutic large molecules by leveraging embeddings from language models trained on protein sequence data. Such protein language models (pLMs) have shown success on property prediction and engineering tasks, learning powerful representations of protein sequences.8–14 Additionally, pair-matching strategies have most recently been employed for protein fitness prediction within contrastive learning (ConFit)15 and implicit guidance (PropEn)16 frameworks. Our methodology captures the nuances of protein sequence variation by learning on relative embeddings and property differences, rather than absolute values. We call our model ‘DyAb’ - from Dyad, meaning a pair, and Antibody, the protein class on which we focus here. One of the most important steps in antibody development is in vitro affinity maturation.17 This process can entail screening single point mutants for binding against the antigen of interest, and selecting variants with improved affinity. We employ our model on three such datasets from three different targets (i.e., disease programs), containing as few as 100 labeled points. For each set, we generate antibody variants that express in mammalian cells and bind strongly to the target of interest. Moreover, our method can design mutation combinations and antibody variants with improved affinity on the single point mutants. DyAb thus represents a promising approach for sequence diversification and optimization in the early stages of biologic therapeutic development, where labeled data can be few and far between. Learning differences between protein pairs in a low-data regime The DyAb model framework was inspired by the types of datasets common in early-stage biologic development – specifically, data from COmprehensive Substitution for Multidimensional Optimization (COSMO) experiments.18 Such experiments mutationally scan residues in antibody complementary-determining regions (CDRs) with all natural amino acids, except cysteine. The resulting dataset contains affinity labels for ∼ 500–1000 point variants around a lead molecule, providing insights into the most important residues for antigen binding. Our model leverages these closely related variant sets to learn relationships between antibody pairs and predict higher edit distance sequences that maintain antigen binding. For example, in this pair-wise framework, the model learns on almost 125,000 pairs of binding data rather than the 500 point mutant data in the COSMO set. DyAb takes pairs of protein sequences as input (Figure 1a). Protein sequences are first fed through a pre-trained protein language model (pLM), from which embeddings are extracted (Figure 1b). Here, we primarily discuss results using AntiBERTy19 and LBSTER,20 pLMs trained exclusively or primarily on antibody sequences. We then compute a ‘relative embedding’ for each unique sequence pair by taking the difference between their embeddings. Please see Supplementary Figure S1-S2 for more detailed descriptions of this framework. A convolutional neural network (CNN) trained on these relative embeddings and binding affinity labels then outputs a prediction of the affinity difference between two sequences (Figure 1c). Thus, DyAb can either be used as a ranking model to score combinations of known mutations, or paired with a sampling algorithm, such as a genetic algorithm (GA), to directly generate new sequences (Figure 1d). Both are explored in this work. Overview of the DyAb workflow for antibody binding affinity prediction. (a) Pairs of closely related protein sequences – here, the variable domain of a therapeutic antibody (blue) – are fed through a (b) Pre-trained language model (LM). The relative embedding between these sequences is then (c) Used as input to a convolutional neural network (CNN) to predict differences in binding affinity. (d) A genetic algorithm (GA) is optionally employed to sample mutation combinations. Created in BioRender. Hofmann, J. (2026) https://BioRender.com/r30u354. We tested DyAb on three antibody-antigen binding affinity datasets of varying size and complexity (Figure 2). Here, binding affinity is defined as the log transform of the equilibrium dissociation constant, pKD=−log10(KD)=−log10(kd/ka), where kd and ka are the measured dissociation and association rate constants, respectively. Each set contains labels for variants around an antibody ‘lead,’ a known binder of a given antigen. Binding affinity differences, ΔpKD, are then computed relative to this lead. The first dataset represents the use case that inspired the model: a COSMO dataset of N=512 point mutants of an internal antibody (anonymized here, termed ‘lead A’). DyAb performs well on a held-out test set of N=77 point variants (N=2926 pairs), with a Pearson correlation coefficient r=0.77 and Spearman rank correlation coefficient ρ=0.80 (Figure 2a, p < 0.001 for both). DyAb model performance on small antibody affinity datasets. Predicted versus measured improvements in affinity, ΔpKD, for variants of three lead antibodies (upper, colors, (a-c)), for the test set held-out during training. Pearson (r) and Spearman (ρ) correlation coefficients are reported for each test set. Here, higher ΔpKD means better for affinity maturation. The starting KD and equivalent pKD are shown for each lead. Created in BioRender. Hofmann, J. (2026) https://biorender.com/r30u354. We also evaluated DyAb on more difficult data landscapes for anti-EGFR and anti-IL-6 leads, given only ∼100 affinity labels for each. These sets included more distant variants of the lead, within 20 edit distances (ED, see Supplementary Figure S4). Such variants might typically be obtained via rounds of in silico or in vitro affinity maturation.17 When trained only on point-variant data from alanine-scanning mutagenesis21 and a handful of higher ED variants, DyAb predicts variant ΔpKD with r=0.75 and ρ=0.75 (Figure 2b; p < 0.001 for both). We observed good performance even in the complete absence of point-variant data, with r=0.84 and ρ=0.84 over the anti-IL-6 variant test set (Figure 2c; p < 0.001 for both). DyAb generates expressing antibody variants with high binding rates We next applied DyAb to generate variants of these three lead antibodies and further optimize antigen binding affinity. First, for the lead A set, we selected all COSMO mutations that individually improved affinity. We then generated all combinations of these mutations with ED = 3–4, and used the trained DyAb-AntiBERTy model (Figure 2a) to score designs by predicted ΔpKD relative to the lead. Starting from this initial population, we employed a genetic algorithm (GA) to select and mutate these sequences to sample the vast design space and iteratively improve the predicted ΔpKD (see Methods). The 47 top-ranked designs produced by this DyAb-GA model were selected for experimental testing (Figure 3a, termed R1). 85% of this design set successfully expressed in mammalian cells and bound to the target antigen, an improved binding rate to that of the COSMO point mutants (59%). DyAb performance on the regression task for design sets is shown in Supplementary Figure S3. Of all DyAb-designed binders against target A, 84% improved on the parent affinity of 76 nM, with the strongest binder reaching 15 nM. DyAb designs expressing antibody candidates against multiple antigens with high binding rates. (a-c) (left) affinity improvements relative to the starting lead molecule, ΔpKD, for the training datasets of point mutations (light gray, general ‘point mut.’ or alanine-only ‘ala. scan’) and higher edit-distance variants (black), and DyAb-generated designs (colors). R1 models used AntiBERTy embeddings19 and R2 used LBSTER embeddings.20 (right) binding rates for antibodies in the training set and DyAb designs. (e, g) sequence analysis of the heavy chain CDRs for the highest-affinity DyAb designs against EGFR and IL-6. Mutations are colored by amino acid character: aliphatic (pink), polar (red), negative (grey), positive (black), and other (gold). Arrows denote sites of all H-CDR mutations found in DyAb-designed binders. (f, h) structural analysis of DyAb designs and their starting leads. Anti-EGFR structures were solved experimentally (PDB entries 9MU1 for the lead and 9MSW for aEGFR-R1-1), whereas fv structures for the anti-IL-6 lead and R2-1 design were predicted via ABodyBuilder2.22 mutations in the CDRs and framework regions (spheres) are labeled. Here, higher means better for affinity maturation. Structures were visualized with PyMol.23 created in BioRender. Hofmann, J. (2026) https://BioRender.com/r30u354. We followed a similar procedure for designing anti-EGFR variants. First, we selected all mutations in the training set that improved affinity individually (in the alanine-scan) or within a higher ED variant. Instead of employing a GA, we then exhaustively generated and scored all combinations of these mutations between edit distances of 3 and 11; exhaustive enumeration was feasible here because fewer affinity-improving mutations were identified relative to the COSMO set, and was preferred to ensure complete coverage of the combinatorial design space. Designs with a DyAb-predicted ΔpKD>0 were then submitted to another ranking oracle20,24 that selected a final 44 sequences for experimental testing. In the first round, 89% of designs expressed and bound EGFR (Figure 3b, R1). 79% of these binders had stronger measured affinities than that of the starting point lead candidate (3.0 nM), including eleven designs with at least a ten-fold improvement. While the strongest binders had similar affinities to high-ED variants in the training set (∼ 100 pM), DyAb produced binders at a much higher rate, on par with that of Alanine point mutants. We then undertook a second round of design that incorporated these data back into the training set. A single sequence was selected for experimental testing (R2, Figure 3b). This design expressed, bound EGFR, and further improved affinity to 66 pM, exhibiting a near 50-fold improvement. Finally, we tested if DyAb can generate antibodies with favorable properties given data for only ∼ 100 variants of an anti-IL-6 lead. As before, we selected and exhaustively combined all mutations from affinity-improving variants in the training set up an ED of 9, and scored the generated sequences with a DyAb-LBSTER model (Figure 2c). A subset of 8 designs with predicted improved affinity ΔpKD>0 were then selected24 for experimental testing (Figure 3c, R2). All generated sequences were successfully expressed, bound IL-6, and improved affinity relative to the lead (1.4 nM). Four designs increased affinity by more than 3-fold, surpassing all variants in the training set. Structural analysis of top DyAb designs While DyAb is a sequence-based model, analysis of top designs suggests putative structural mechanisms underlying the observed affinity improvement. We determined crystal structures of the corresponding Fabs of the anti-EGFR lead and R1-1 design to 2.4Å and 2.1Å resolution, respectively, by molecular replacement and refinement (see Methods; electron density maps are shown in Supplementary Figure S6). DyAb makes four mutations in aEGFR-R1-1 across CDR-H1, CDR-H3, and the framework. Mutating Kabat position V H 97 from Glycine to Aspartic acid – consistent across all top designs (Figure 3e) – alters the conformation of CDR-H3. An additional Proline at V H 98 could further stabilize this binding-able conformation in aEGFR-R1-1 relative to the lead. All three top designs also mutate V H 34 at the base of CDR-H1 from Valine to a longer aliphatic amino acid. In lieu of experimental structures, we compared variable domains predicted by ABodyBuilder222 for the highest-affinity anti-IL-6 design and its lead (Figure 3h). While the same V H S98P mutation is seen in the top two designs, the predicted CDR-H3 conformation remains unchanged. Instead, a mutation in CDR-H2 (N52aH) drives extension of the loop into solution (Figure 3g). Antibody-specific protein language models improve DyAb performance DyAb’s pair-wise framework can be constructed from a variety of language model embeddings. While this work focused on antibody-specific pLMs, many general protein language models exist. Here, we compare DyAb performance head-to-head across all three variant datasets using embeddings from different pLMs: AntiBERTy19 (trained on 500 M natural antibody sequences), LBSTER20 (trained on the same antibody sequences, plus the general protein universe), and ESM-2 models of three different sizes25 (trained on just the general protein universe). We evaluated DyAb models trained using relative embeddings from each pLM on the same train/test sets shown in Figure 2. Performance on the downstream ΔpKD prediction task varies between datasets (Figure 4). In some cases, we observed better performance with a different pLM than was used for design. For example, anti-IL-6 R2 designs were generated using LBSTER embeddings, which are slightly outperformed by AntiBERTy-based models across two of the four evaluation metrics. However, pLMs trained primarily or entirely on antibody repertoires produced more predictive DyAb models across the board versus a general pLM (ESM-2). Ablation of language model embeddings. Evaluation of DyAb model performance on the held-out test set for all three antibody leads, trained with relative embedding from five pLms: AntiBERTy,19 a LBSTER pLM trained on antibody-specific and general data20 (see methods), and three ESM-2 models of varying size.25 Asterisks denote the best-performing model on that dataset for each metric: (a) Pearson r2, (b) Spearman ρ, (c) root mean squared error RMSE, and (d) the area-under-the-curve (AUC) of the receiver operating characteristic (ROC) curve when predicting a binary improved/worsened affinity label with a threshold of ΔpKD=0. Created in BioRender. Hofmann, J. (2026) https://BioRender.com/r30u354. Discussion We developed DyAb, a deep learning model that leverages sequence pairs to predict binding affinity differences in a limited data regime. When applied to antibody design, DyAb efficiently generates variant sequences with improved binding affinity given as few as ∼ 100 labeled training data. Designs express and bind at consistently high rates (>85%), comparable to that of single point mutants. Most DyAb-generated sequences improve upon the affinity of the lead and even upon that of the training set. DyAb thus represents a promising tool for early-stage antibody lead optimization and diversification. We tested DyAb both as a ranking model for scoring combinations of mutations, and paired with a genetic algorithm (GA) for sampling. The main failure mode of GAs is a deviation from ‘natural’ sequences after long optimization trajectories away from the starting sequence, exceeding ∼ 8 edits.24 We avoid this issue by setting a low edit distance design limit (ED = 7), incorporating only mutations found in previously stable sequences, and using pLM likelihoods in the discriminator. DyAb can be readily integrated with other algorithms like Monte Carlo tree search26 or generative methods like PropEn16 to further sample the design space. Future iterations could also incorporate protein structural features by leveraging embeddings from structure-informed models like ESMFold27 or SaProt.28 Moreover, experimental structures for DyAb-designed antibodies in co-complex with antigens could provide additional insight into binding mechanisms and model rankings. We note that the anti-EGFR and anti-IL-6 designs tested in experiment were sub-selected from a larger pool of top designs by another ranking oracle,24 responsible for internal cross-project sequence prioritization. The designs against target A were selected solely based on DyAb scores, serving as an internal control demonstrating DyAb’s standalone design capability (85% binding rate). In all cases, DyAb was responsible for creating the enriched library containing the improved binders, regardless of what final ranking model was used. We refer the reader to the following studies for a discussion of these broader computational efforts.29,30 We additionally compared the DyAb pairwise framework against a linear baseline using absolute pLM embeddings for direct property prediction (Supplementary Table S2). While the baseline achieves competitive performance on held-out test data from the same distribution, DyAb consistently outperforms the baseline on prospective design-round data, suggesting that the pairwise framework better generalizes to novel sequence contexts. We acknowledge several limitations of the current study. First, DyAb was validated exclusively on antibody binding affinity; additional developability properties critical for therapeutic antibodies – including aggregation propensity, nonspecific binding, thermal and chemical stability, viscosity, and polyreactivity – were not assessed. While all DyAb designs were confirmed to express in mammalian cells, a comprehensive developability profile would be necessary before advancing candidates to later stages of drug development. Second, the pairwise framework assumes variants are closely related to a known lead sequence; performance at very high edit distances from the training distribution is expected to degrade, as observed in other ML-guided protein engineering approaches. Third, while the combinatorial expansion from N to N(N−1)/2 training pairs provides a richer representation of sequence–property relationships, these pairs are not statistically independent observations, and the effective information content is bounded by the original N measurements. Finally, the model was evaluated on only three antibody-antigen systems, and additional validation on diverse antibody targets and non-antibody proteins would strengthen claims of generality. The ability to learn in a low-N regime makes DyAb a promising method for engineering other antibody properties for which data are even scarcer. Chemical and physical stability at high concentrations are particularly important for drug development, but material-intensive to measure.6 In future work, high-throughput proxy assays for these properties – e.g., that measuremelting temperature,31 self-interactions,32 or solubility33 – could be used to generate datasets of ∼ 100 variants on which DyAb can be applied. While we focused on antibody therapeutics in this study, the DyAb framework is in principle applicable to other protein variant datasets, though further validation on non-antibody systems would be needed to confirm this. DyAb is made publicly available as a model within LBSTER,20 an open-source protein language model framework that supports multiple backbone pLMs including AntiBERTy; users can therefore employ DyAb with any supported embedding model, selecting the pLM best suited to their target protein family. Supplementary Material Supplemental Material Acknowledgments The authors thank the entirety of Prescient Design and the Genentech Antibody Engineering department for insightful discussions. We thank Yanshu Dou and Viva Biotech for structural work and the Shanghai Synchrotron Radiation Facility (SSRF) and staff of beamline BL02U1 for crystallographic data collection. We gratefully acknowledge the Prescient Design engineering team and Genentech IT for computing support. J.Y.L., A.L., K.C., and N.C.F. conceived the work; J.Y.L., J.L.H., A.L., S.V., E.L., P.O.P., N.T.,R.B., V.G., F.S., A.W., K.C., and N.C.F. designed computational experiments. W.L. and Y.W. designed and conducted all wet-lab experiments, except the x-ray crystallography which J.R.K. designed and conducted. J.Y.L., J.L.H., A.L., W.L., S.V., E.L., P.O.P., N.T., F.S., R.B., V.G., A.W., K.C., and N.C.F. analyzed the results. J.L.H., J.Y.L., and N.C.F. wrote the initial draft and all authors reviewed the manuscript. Funding Statement Funding for the study was provided by Genentech. Disclosure statement No potential conflict of interest was reported by the author(s). Data and materials availability All data associated with this work are provided, excepting the sequences for the lead A dataset and designs, which are proprietary to Genentech. The lead A dataset is not necessary for reproducing the results and overall conclusions of the paper. Coordinates and amplitudes for the crystal structures have been deposited at the RCSB with accession codes 9MU1 and 9MSW. DyAb is implemented and available for use under an Apache-2.0 license within LBSTER (20) at github.com/prescient-design/lobster. Supplementary material Supplemental data for this article can be accessed online at https://doi.org/10.1080/19420862.2026.2717460 References Associated Data Supplementary Materials Supplemental Material