BALM-PPI
Fine-tuned protein language models for binding affinity prediction through metric learning, with residue-level Integrated Gradients explanations.
I design sequence-first machine learning systems for protein and antibody binding affinity prediction, with a focus on models that are accurate, efficient to train, and interpretable enough to support scientific decisions. My recent work adapts ESM-2 with parameter-efficient fine-tuning, then uses attribution methods to test whether the model is learning biophysically meaningful interaction signals.
I am a Research Scholar at Shiv Nadar University, Delhi-NCR, and a Research Ambassador for Springer Nature India. I work across protein language models, antibody-antigen modeling, active learning, and explainable AI for drug discovery.
My active-learning benchmark for antibody-antigen binding affinity has been accepted at ACML 2026 (Main Track): a Tanimoto-kernel Gaussian Process over ESM-2 embeddings reaches 67.3% mean Recall@2% versus 8.9% for random sampling under a fixed 10% budget. Recent work also includes NeurIPS 2026 workshop papers on LLM proposers for molecular search (LEAF) and representation effects in molecular active learning (TabAL), plus GRASP, a graph Transformer for molecular property prediction.
Fine-tuned protein language models for binding affinity prediction through metric learning, with residue-level Integrated Gradients explanations.
A sequence-only, chain-aware architecture that models heavy chain, light chain, and antigen interactions without requiring solved structures.
Gaussian-process active learning across eight AbBiBench assays, with attributions validated against crystal structures.
A 93.5M-parameter graph Transformer pretrained on ZINC20 and adapted on 642 ChEMBL assays, reaching mean MAE 0.374 on 23 OpenADMET endpoints.
Structure-free prediction of absolute ΔG and mutation effects from frozen ESM-2 with an interface memory and Pairformer-style refinement.
See the full publication record and project details on the research page.
I also keep a short lab notebook for research notes, implementation decisions, and work-in-progress ideas.
Python (expert), SQL, C/C++, Java, JavaScript, HTML/CSS, LaTeX
PyTorch, Hugging Face (Transformers / PEFT), Scikit-learn, Keras, CatBoost
ESM-2, RDKit, OpenBabel, ANARCI (antibody numbering), NGL Viewer
LoRA, metric learning, gated cross-attention, Integrated Gradients (XAI), self-supervised learning, ensemble learning, active learning
Git, Docker, GCP, Linux/Bash, VS Code, Weights & Biases
Pandas, NumPy, Matplotlib, SciPy, Captum