Computational Biology
Physics-informed deep learning for protein stability and enzyme kinetics. Treats proteins as vibrating machines — extracting Vibrational Density of States from Normal Mode Analysis for a novel tri-modal prediction architecture.
Fig. 1 — Molecular structure transformed into a graph representation for GNN processing.
Architecture
Three encoders capture sequence, structure+dynamics, and chemical features. Gated attention fusion combines them for prediction.
Sequence Encoder
Protein language model (Rostlab/prot_t5_xl_uniref50) generates 1024-dim embeddings from amino acid sequences.
Structure & Dynamics
Graph Attention Network encodes protein topology. Vibrational Density of States from Normal Mode Analysis captures dynamics invisible to static models.
Chemical Encoder
Chemical language model (seyonec/ChemBERTa-zinc-base-v1) with Differential Reaction Fingerprints encodes substrates into 512-dim embeddings.
GATv2
Each node aggregates information from neighbors using learned attention weights. After multiple layers, atoms encode their full molecular environment — capturing the chemical context that determines protein behavior.
Fig. 2 — GATv2 message passing with attention-weighted neighbor aggregation.
Framework