PaTz โ€” Physics-Aware Transformer with eXplainability
โ— API Offlinepatz_env

Research Framework

PaTz

Physics-Aware Transformer with eXplainability โ€” end-to-end research pipeline for HIV-TB co-infection protein analysis.

Data Layer
  • ยทUniProtKB sequences
  • ยทPDB 3D structures
  • ยทGene Ontology database
  • ยทMMseqs2 sequence clustering
Embedding Layer
  • ยทESM-2 650M (Meta AI)
  • ยทPer-residue 1280-dim embeddings
  • ยทFrozen weights โ€” no fine-tuning
Model Layer
  • ยทPairformer (4 blocks, d=256/64)
  • ยทScore-based Diffusion (T=1000)
  • ยทFunctionHead (354 GO terms)
  • ยทPhysics Energy Terms (Cฮฑ)
Training Layer
  • ยทFP16 + gradient accumulation
  • ยทAdamW + cosine LR scheduler
  • ยทMLflow experiment tracking
  • ยทEarly stopping (patience=10)
Explainability Layer
  • ยทIntegrated Gradients (Captum)
  • ยทMC-Dropout uncertainty (N=20)
  • ยทPer-residue attribution heatmaps
Application Layer
  • ยทDrug candidate screening (20 compounds)
  • ยทMulti-objective scoring J
  • ยทFastAPI REST + WebSocket backend
  • ยทNext.js research platform (this UI)