Major version: NPEAD evolves from adaptive scorer → self-directed biological intelligence. Decides what to test next, not just what score to assign.
| Capability | NPEAD v5.2 | NPEAD-DLE v6.0 |
|---|---|---|
| Core question | "How good is this candidate?" | "What should we test next?" |
| Decision model | Weighted linear scorer (28 dims) | EIV-driven action selector + scorer |
| Uncertainty | Implicit (missing fields) | Explicit 7-dim uncertainty vector |
| Next action | Rule-based (hardcoded order) | EIV formula: gain × impact × reliability ÷ cost |
| Kill logic | Manual (human demotes) | Auto-kill: TSGs, weak targets, undruggable |
| Authority gates | Static thresholds | Dynamic per evidence level |
| Learning | MAML seed + FusionWeightLearner | + Meta-learning policy (Phase 6, planned) |
| Self-cleaning | None | Auto-detects ORs, TSGs, pseudogenes |
For each hero, the system ranks every possible action (re-dock, MD, DepMap check, selectivity screen, wet-lab, kill) by expected value. The action with highest EIV wins. If no action exceeds threshold → hold. If kill scores highest → auto-remove.
Is the tumor dependent on this target? Source: DepMap CRISPR score. High uncertainty = never checked.
Does the compound stay in the pocket? Source: MD simulation RMSD. High = no MD run yet.
Does it hit only the target? Source: selectivity counter-screen. High = unknown off-targets.
Is it toxic? Source: toxicity prediction + hERG. High = no safety data.
Do we understand WHY this target matters? Source: FRSI regulatory score. High = no biological rationale.
Can we actually make a drug for this? Source: pocket/structure analysis. High = unknown pocket.
Does the compound bind at all? Source: Vina docking score. High = weak or no docking.
| Action | Cost | Reliability | Authority Req | When Selected |
|---|---|---|---|---|
| depmap_check | Free | 0.90 | 0.0 | Dependency unknown |
| pathway_reversal | Free | 0.60 | 0.0 | FRSI = 0 (no mechanism) |
| redock_v4_neural | $0.50 | 0.75 | 0.3 | Weak Vina, V3 failed |
| redock_expanded | $0.50 | 0.70 | 0.3 | Has ChEMBL inhibitors, weak Vina |
| md_simulation | $2.00 | 0.85 | 0.35 | Good dock, unknown stability |
| selectivity_screen | $1.00 | 0.80 | 0.5 | Strong binder, unknown off-targets |
| binding_assay | $500 | 0.95 | 0.7 | MD passed, no wet-lab yet |
| kill | Free | 1.00 | 0.0 | TSG, undruggable, or EIV=0 for all |
23 known TSGs (TP53, PTEN, CDKN2A, SMAD4, RB1, APC, VHL, BRCA1/2, etc.) → cannot inhibit. Auto-reclassified to synthetic_vulnerability_context.
OR pattern genes (OR2G2, OR6F1, etc.) → 7-TM GPCRs with zero cancer relevance. Blocked at pipeline entry.
DepMap non-essential + no FRSI + weak binding (>-5.5 Vina) → no evidence path justifies compute. Killed.
Druggability = LOW + Vina > -6.0 → pocket too shallow/absent. Killed.
| Metric | Before DLE | After DLE | Change |
|---|---|---|---|
| Heroes with compounds | 49 | 76 | +55% |
| Stuck at quality gates | 44 | 6 | -86% |
| Junk in pipeline | ~15 | 0 | Clean |
| A-tier candidates | ~20 | 26 | +30% |
| Rescue rate (stuck heroes) | 0% | 63% | 27/43 rescued |
How reliably each brain's signal predicts good targets. Teacher-evaluated + outcome-validated.
⚠️ Safety brain (3%) reflects evidence coverage, not toxicity quality. Low score = more hERG/tox/ADMET data needed before wet-lab. This is a governance signal, not a failure.
| # | Dimension | Weight | Profile | Magnitude |
|---|
Which gene families strengthen or weaken predictions:
| Family | Accuracy | Correct | Wrong | Signal | Genes |
|---|
Ranks actions by evidence value. Integrated into nightly orchestrator.
7-dimension uncertainty vector per hero. Written to DynamoDB nightly.
Skips actions that won't change the decision. Saves ~20-30% compute.
Records action→result pairs. Seeds Phase 6 meta-learning.
Dynamic escalation: wet-lab requires MD passed + selectivity ≥ 0.5.
Needs 50+ outcomes. Learns which evidence sequences produce hits per target class.
| Version | Date | Change |
|---|---|---|
| v6.0 (DLE) | 2026-06-27 | Major: EIV calculator, uncertainty map, auto-kill, TSG detection, integrated nightly |
| v5.2 | 2026-06-14 | 28-dim scorer, dual profiles, MAML seed, authority base 0.6 |
| v5.0 | 2026-06-09 | Hard gates, conditional novelty, PPB bounded |
| v4.0 | 2026-05-30 | 23-dim, authority-constrained, learn-to-learn framework |