Hi! 👋🏼 I'm Aaryesh Deshpande

A fine-tuned Transformer model mapping gene expression perturbations to molecular SMILES for targeted cancer drug design; achieved 23% novel scaffold generation with high pathway alignment and QED >0.75.

Built a CNN+Bi-directional GRU classifier on GaMD trajectories to classify protein conformational states. Analyzed inter-subunit pathways using dynamic cross-correlation and estimated transition kinetics with corrected state-transition matrices.

Developed a pseudo-log-likelihood typicality framework for protein language models (ESM2, ProtBERT) combined with TwoNN intrinsic dimensionality to audit evolutionary coverage and detect distributional mismatches in embedding representations.

A dual-encoder DTI model with state-conditioned hypergraph fusion. The system encodes ligands via pharmacophore-aware hypergraphs and protein pockets via state-specific residue graphs (agonist vs antagonist), fusing representations through cross-attention for ligand toxicity and functional state classification.