Research Overview
PhD RESEARCHPaTz ā Physics-Aware Transformer with eXplainability for HIV-TB Co-infection Protein Analysis
This research develops a novel deep learning framework ā PaTz ā that combines protein language model embeddings (ESM-2 650M), geometric reasoning (Pairformer), physics-constrained structure generation (score-based diffusion with differentiable energy terms), and explainability (Integrated Gradients + MC-Dropout uncertainty) for the analysis of HIV-TB co-infection proteins. The framework is evaluated on a curated dataset of 2,620 proteins from HIV-1 and M. tuberculosis H37Rv, achieving Fmax=0.3212 on GO function prediction ā competitive for multilabel annotation of pathogen proteomes.
Identify structural and functional commonalities between HIV-1 and M. tuberculosis proteomes to understand co-infection mechanisms at the molecular level.
Integrate differentiable CĪ± energy terms (bond, angle, clash) into a score-based diffusion model for physically-constrained 3D protein structure generation.
Multilabel prediction of 354 Gene Ontology terms (BP, MF, CC) using a specialized FunctionHead. CAFA-style Fmax evaluation protocol.
Per-residue attribution analysis using Captum IG to identify which amino acid positions drive each functional prediction, supporting mechanistic interpretability.
Multi-objective scoring of 20 flavonoid compounds against HIV-TB co-infection targets, combining structural, functional, binding, immune, and XAI scores.