Raunak

Coarse composition suffices: tabular in-context learning for multi-activity antimicrobial peptide profiling↗

A. Pal*, Raunak Kumar*, D. Solanki*, P. Pareek, J. Singh, J. Singla (*Equal contribution)

bioRxiv preprint; under review at Bioinformatics (Oxford University Press)

Developed a gradient-free, tabular in-context learning framework using TabPFN and 330 sequence-derived descriptors, completely bypassing computationally heavy 3D structure prediction to profile multi-activity antimicrobial peptides. Achieved SOTA multi-label performance on the ESCAPE benchmark (82k+ peptides, 5 labels), increasing macro AP from 72.1% to 77.8% and outperforming prior deep multimodal baselines across all 5 targets simultaneously with strong out-of-domain generalization.

Bayesian Optimization for Ultra-Large Virtual Screening of EphA2 Ligands

Independently designed and validated a Bayesian optimization framework for virtually screening a 100M small-molecule library, recovering >90% of top-500 candidates while docking ~1% of the library at a throughput of 34.1 ms/molecule (~1M molecules in 25–33 hrs on a single GPU). Developed a hierarchical pipeline integrating ensemble docking and molecular dynamics to refine hits to 5 high-confidence EphA2 binder candidates for experimental validation.

De novo Protein Design

Developed a de novo protein design pipeline for reverse transcriptases targeting prime editing, combining RFdiffusion backbone generation with ProteinMPNN inverse folding for 10,000 candidates. Designed a UMAP-based filtering pipeline and adapted Gemma-4B for protein language modeling via LoRA.