ParaEM: antigen-aware sequence-based paratope prediction with expectation-maximization.

motivationAccurate identification of antibody paratopes-the residues contacting the antigen-is vital for understanding antibody-antigen recognition, antibody engineering, and vaccine design. However, structure-based predictors require antibody structural inputs, which are often obtained from experimentally determined antibody-antigen complexes in benchmark settings but are generally unavailable for new sequences. Conversely, sequence-based predictors scale efficiently but remain largely antigen-agnostic, limiting their use of pair-specific antigen context.

resultsWe introduce ParaEM, an antigen-aware sequence-based paratope predictor that operates directly on antibody and antigen sequences. ParaEM leverages pretrained ESM3 embeddings to encode both sequences, augmenting antibody residues with trainable, complementarity-determining regions (CDR)-informed embeddings to introduce a biologically informed inductive bias. It models interactions via a residue-level compatibility scorer that defines latent antigen-anchor distributions. During training, a generalized Expectation-Maximization (EM) framework alternates between estimating label-conditioned posterior responsibilities (E-step) and performing gradient-based parameter updates (M-step). Evaluated across three established benchmarks, ParaEM outperforms existing sequence-based baselines in AUC-PR and achieves performance competitive with structure-based predictors without requiring explicit three-dimensional structural inputs. Ablation studies show consistent improvements from generalized EM optimization and an additional gain from CDR-informed embeddings. Antibody-side and antigen-information controls further show that antibody sequence carries substantial predictive signal, while the matching antigen provides additional predictive information. Together, these results support ParaEM as a scalable framework for incorporating matching-antigen context into sequence-based paratope prediction. AVAILABILITY AND IMPLEMENTATION: ParaEM is available at https://github.com/kimtaegyuu/ParaEM. An archival snapshot of the software corresponding to this study is available on Zenodo at DOI: 10.5281/zenodo.22157100.