Canonical pipeline from the Agros research vault — generate, score, interpret with integrated gradients, optimize constraints, validate in lab, then iterate.
Generation
BoltzGen proposes 1000+ target-specific candidates with structure-based sampling and diversity optimization.
Scoring
Multi-model evaluation: Boltz-2 binding affinity, physics docking (Vina), GNINA deep learning — rank top ~10% for analysis.
Analysis
Explainable AI attribution per residue across models (Boltz-2, Chai, ESM-2, NetSolP, OpenMM in the full stack). Consensus positions get locked.
Optimization
Constraint-based redesign: lock high-attribution subsets, sample combinations, capture epistasis without over-constraining.
In lab testing
Synthesize top candidates, measure binding and efficacy, fine-tune models with experimental data — lab feedback closes the loop.