Structural IntelligencePhase 3 API live · Tier 2 + Tier 3 proxy backendsESM + Boltz-2 swap-in pending compute
One layer deeper than the GNN. Sequence structure. Ligand binding. Zero-shot.
The Discovery module answers what does graph topology suggest. This module answers what does the protein sequence and structure say. Tier 2 ESM-style variant effect scoring for the 32% of Indian variants absent from gnomAD. Tier 3 Boltz-2-style co-folding for undocumented phytochemical–enzyme binding hypotheses. Every prediction is structurally separated from the tiers above and below it — the three predicted tiers never collapse into one.
Framing discipline per PRD §8: this module "meaningfully advances" the Indian pharmacogenomic hypothesis-generation stack. It does not "solve" pharmacogenomics for India. Any external claim respects this line.
⚠️
COMPUTATIONAL PREDICTION — NOT VALIDATED FACT
Every output here is either an ESM-style variant disruption score or a Boltz-2-style binding affinity — both are computational estimates, not literature-sourced or wet-lab-confirmed measurements. No output from this module surfaces through IndoPGx / CardioRisk / AyurBridge / IndoG6PD / IndoAutoImmune / Orchestrator clinical paths. Top-ranked candidates surviving all three computational tiers become the Tier 4 wet-lab validation queue — the specific concrete hypotheses offered to partner labs.
Four-tier confidence architecture
Tier 4 — WET-LAB / LITERATURE VALIDATED
The only tier that surfaces through clinical modules. Written by partner labs (Precigenetics-style) or curated literature ingestion.
:TARGETS · :Variant.effect
Tier 3 — Boltz-2 co-folding + affinity
Structure + binding score. Uses curated PDB / AlphaFold DB structures. Proxy today; real Boltz-2 (open-source June 2025) on PARAM Siddhi / A100 next.
:PREDICTED_BINDING
Tier 2 — ESM zero-shot variant scoring
Sequence-position-based disruption score. Amino acid property distance today; ESM2 masked-LM log-likelihood next.
:Variant.esm_disruption_score
Tier 1 — GNN topology link prediction
Graph co-occurrence heuristic. Discovery module. Cheapest tier — runs across the whole substrate.
:PREDICTED_TARGETS
The three predicted tiers write to distinct graph properties/edges. Clinical modules query only Tier 4. Auditable at /api/structural/tiers.
Full three-tier funnelTIER 1 → TIER 2 → TIER 3 · END-TO-END
Runs all three predictive tiers on the same input with hard boundaries maintained. Emits a Tier 4 wet-lab-candidate flag only when all three tiers agree with high confidence and the pair is not already documented — the concrete hypothesis list you offer a partner lab.