DELFICTUS IO — PRISM-AI · CLOSED-LOOP INSTANT-INFERENCE PLATFORM
In development  ·  closed-loop predictor under active build  ·  for the production engine see Live Engine →

PRISM-AI

Zero-shot instant inference.

PRISM-AI is a learned predictor that returns binding-site inference from a single structure in seconds. No engine run is required at inference time. The predictor is bootstrapped from PRISM-4D outputs that have already passed the canonical gating stack. Every disagreement between predictor and engine triggers an authoritative engine run, and that result re-enters the training pool. The demo below operates against an early predictor state. What is in active development is the infrastructure that closes the loop. The engine itself is production today at /live-engine/.

1 · Source of truth
PRISM-4D engine
Production today. Every output passes the canonical gating stack (Therm → Coherence advisory → Localization → Contact-Reorg → Response-Selectivity) before it enters the training pool.
2 · Training pool
Gating-passed outputs only
No DPS, no composite scores, no unvalidated guesses. The pool is grown by replicate consensus across runs, with provenance manifests preserved.
3 · Predictor
PRISM-AI inference
Returns ranked pocket candidates, lining-residue maps, and confidence scores in seconds. No MD trajectory required. No nonequilibrium dynamics. No GPU hours per query.
4 · Disagreement → re-validation
Engine re-run on uncertainty
Low-confidence or out-of-distribution predictions are flagged and queued for an authoritative engine run. The result re-enters the training pool. The loop closes.
How this benefits PRISM-4D workflows
Triage at scale. PRISM-AI screens thousands of structures in the time a single engine run takes. Only the ambiguous or high-value targets get the full validated run.
Continuous validation feedback. Every engine run improves the predictor. Every uncertain prediction earns engine validation. No predictor drift goes uncorrected.
Workflow continuity. Same target taxonomy. Same pocket definitions. Same gating posture. Predictor and engine speak the same language, so a medchem team sees consistent outputs across both surfaces.
Same deployment posture. The closed-loop infrastructure inherits the engine's deployment surface. That covers the consumer-GPU baseline, multi-node scaling, and fully air-gapped cleanroom configurations without compromise.
—— PRISM-AI PLATFORM DEMO

Explore Protein
Binding Sites

Upload a PDB file, enter a PDB ID, or ask a question. PRISM-4D complements existing workflows by revealing binding site candidates from static structures alone.

Upload PDB

Upload a PDB file for binding site analysis. Results typically arrive within 30 seconds.

Fetch from RCSB

Enter a 4-character PDB ID to download and analyze directly from the Protein Data Bank.

Ask PrismAI

Ask about protein biology, drug design, binding sites, or the PRISM-4D platform.

Responses