Research & Publications

Proposed a structure-aware LLM framework that is the first to jointly predict antibody binding and neutralization for SARS-CoV-2, consistently outperforming prior classifiers, especially on closely related viral variants. We further showed that refining ESM-2 with target-specific antibody sequences and structural information boosts performance substantially, improving binding AUC from 0.84 to 0.93 and neutralization AUC from 0.88 to 0.95, establishing a strong, generalizable baseline for antibody–antigen interaction modeling.

End-to-end web platform for comparative structural analysis of protein–protein complexes, supporting interface characterization, energetic analysis, and variant-level comparison.

ASPred is a large-language-model-based framework trained on antigen-BCR sequence pairs that accurately predicts antigen-specific B-cell receptors from full repertoires, successfully generalizes to previously unseen antigens, and validates top candidates via molecular dynamics simulations, demonstrating that interaction specificity is learnable directly from sequence data.

Developed Ab-Affinity, a hybrid large-language-model-driven antibody design framework that integrates genetic algorithms and simulated annealing to generate novel SARS-CoV-2-targeting antibodies, achieving >160-fold binding affinity improvements over experimental candidates and outperforming existing computational approaches while maintaining strong biophysical stability.

Designed a lightweight, state-of-the-art breast cancer histopathology classifier that combines self-supervised contrastive learning and an efficient ResNet–Inception architecture, achieving up to 98% accuracy across multiple magnification levels with ~50% fewer parameters than existing leading models—enabling accurate, data-efficient, and scalable diagnostic AI.

Developed and benchmarked machine-learning models for SARS-CoV-2 3CLpro inhibitor prediction, evaluating 27 classifiers and a neural network that achieved 91% accuracy on ChEMBL and 93% on combined ChEMBL–PubChem data, with F1-scores of 93–94% for active and inactive compounds. The study further validated scalability on >100,000 molecules and used SHAP-based interpretability to identify key molecular fingerprints, enabling data-driven and explainable antiviral drug screening.

Proposed a data-efficient deep learning architecture for yoga pose recognition that jointly captures spatial and depth features, outperforming state-of-the-art CNNs (ResNet, Inception, Xception) while achieving 94.9% accuracy and 95.6% precision using limited training data.