MindNetQ | Pipeline Gates

Monotone cost escalation. Cheapest first. Every compound earns the right to the next gate. Updated June 20, 2026.

Active Gate System (gpu-batch-rs)

Gate	Name	Pass Condition	Fail Action	Cost	Stage
G0	SMILES Validity	RDKit parse + sanitize succeeds	DROP	$0	CPU
G1	Structural Alerts	No PAINS + no reactive groups + no 30 med-chem patterns (Brenk/Lilly)	DROP	$0	CPU
G2	Drug-Likeness	MW 200–600, QED ≥ 0.3, ≤ 10 rotatable bonds, Lipinski ≤ 1 violation	DROP	$0	CPU
G3	Atom Filter	Only C, N, O, F, P, S, Cl, Br — no exotic atoms	DROP	$0	CPU
G4	Docking (Uni-Dock)	Vina ≤ -6.5 kcal/mol (LE-adjusted)	REJECT	~$0.05	GPU
G5	Pose Consensus	≥ 1 valid pose within pocket, variance < threshold	FLAG	$0	GPU
G6	NPEAD Scoring	advancement_score > threshold (learned, adaptive per target)	REJECT	$0	CPU
G7	Chemistry Validation	No CRITICAL flags, SA < 6.0, no reactivity alerts	GRADE CAP	$0	CPU (nightly)
G8	Evidence Chain	Hash-linked audit trail valid, all prior decisions recorded	HOLD	$0	CPU

Gate

Name

Pass Condition

Fail Action

Cost

Stage

SMILES Validity

RDKit parse + sanitize succeeds

DROP

CPU

Structural Alerts

No PAINS + no reactive groups + no 30 med-chem patterns (Brenk/Lilly)

DROP

CPU

Drug-Likeness

MW 200–600, QED ≥ 0.3, ≤ 10 rotatable bonds, Lipinski ≤ 1 violation

DROP

CPU

Atom Filter

Only C, N, O, F, P, S, Cl, Br — no exotic atoms

DROP

CPU

Docking (Uni-Dock)

Vina ≤ -6.5 kcal/mol (LE-adjusted)

REJECT

~$0.05

GPU

Pose Consensus

≥ 1 valid pose within pocket, variance < threshold

FLAG

GPU

NPEAD Scoring

advancement_score > threshold (learned, adaptive per target)

REJECT

CPU

Chemistry Validation

No CRITICAL flags, SA < 6.0, no reactivity alerts

GRADE CAP

CPU (nightly)

Evidence Chain

Hash-linked audit trail valid, all prior decisions recorded

HOLD

CPU

Chemistry Danger Patterns (30 rules — G1)

#	Pattern	Source	Risk
1-8	Reactive electrophiles (acyl halides, Michael acceptors, epoxides, isocyanates...)	Brenk 2008	Covalent modification
9-16	Aggregators + PAINS (rhodanines, catechols, quinones, hydroxamic acids...)	Lilly 2012	False positives
17-24	Metabolic liabilities (nitros, anilines, hydrazines, alkyl halides...)	ChEMBL curation	Toxicity / instability
25-30	Privileged but risky (thiol warheads, boronic acids, peroxides...)	MindNetQ custom	Context-dependent

Pattern

Source

Risk

1-8

Reactive electrophiles (acyl halides, Michael acceptors, epoxides, isocyanates...)

Brenk 2008

Covalent modification

9-16

Aggregators + PAINS (rhodanines, catechols, quinones, hydroxamic acids...)

Lilly 2012

False positives

17-24

Metabolic liabilities (nitros, anilines, hydrazines, alkyl halides...)

ChEMBL curation

Toxicity / instability

25-30

Privileged but risky (thiol warheads, boronic acids, peroxides...)

MindNetQ custom

Context-dependent

NPEAD Gate Learning (Adaptive — G6)

Gate thresholds are not fixed. NPEAD Phase 9 learns from outcomes:

Docking success → lower threshold (easier to pass for similar targets)
Repeated failures → raise threshold (tighter filter)
Gate outcomes written to brain-bus: learning:gate_outcome:{gene}
Current weights: ligand_efficiency (0.235), binding_affinity (0.185), admet (0.15)

Capsule Influence on Gates

Capsule Signal	Gate Effect
Chemistry (Best Vina ≤ -8.0)	Generator tightens → fewer candidates reach G4, but higher pass rate
Safety (all clear)	G1 thresholds maintained — current patterns work
Evidence (weak)	Generator explores more → more diversity at G2, lower average but wider search
Transcriptomics (many dysregulated)	Prioritization signal — more GPU budget allocated to these targets

Capsule Signal

Gate Effect

Chemistry (Best Vina ≤ -8.0)

Generator tightens → fewer candidates reach G4, but higher pass rate

Safety (all clear)

G1 thresholds maintained — current patterns work

Evidence (weak)

Generator explores more → more diversity at G2, lower average but wider search

Transcriptomics (many dysregulated)

Prioritization signal — more GPU budget allocated to these targets

Cost Escalation Path

CPU ($0) G0 → G1 → G2 → G3 → G6 → G7 → G8 (kills ~80% of candidates) │ ▼ GPU (~$0.05/mol) G4 → G5 (kills ~50% of survivors) │ ▼ Quantum ($330+) VQE binding confirmation (only A+ heroes, PI-approved) │ ▼ Wet-Lab ($15K+) Human gate + budget + CRO (final survivors only)

Grade	Criteria	Meaning
A+	Vina ≤ -8.5, validated, all gates pass	Preclinical candidate
A	Vina ≤ -7.5, validated	Strong hit
B+	Vina ≤ -6.5, docked	Promising — needs optimization
B	Vina > -6.5, docked	Weak — scaffold hop needed
C	All molecules failed gates	Target needs different approach

Grade

Criteria

Meaning

A+

Vina ≤ -8.5, validated, all gates pass

Preclinical candidate

Vina ≤ -7.5, validated

Strong hit

B+

Vina ≤ -6.5, docked

Promising — needs optimization

Vina > -6.5, docked

Weak — scaffold hop needed

All molecules failed gates

Target needs different approach

Governance Rules

Rule	Enforcement
Monotone cost escalation	Cheapest gates always first — never spend GPU on what fails CPU
No grade inflation	next-action-rs checks for pipeline bugs before any downgrade; CDK9 stays A (PI decision)
FRSI is advisory	WEAK_FRSI does not block generation — it's informational
Per-gene dedup	Max 3 GPU jobs per gene (watchdog enforces)
No on-demand GPU	All Spot — no spending without PI approval
Evidence chain	Every gate decision SHA-256 hash-linked, tamper-evident
Quantum gate = human-approved	VQE only fires after explicit PI approval + budget
No downgrade without bug check	System checks for error/timeout/OOM/spot before any grade reduction

Rule

Enforcement

Monotone cost escalation

Cheapest gates always first — never spend GPU on what fails CPU

No grade inflation

next-action-rs checks for pipeline bugs before any downgrade; CDK9 stays A (PI decision)

FRSI is advisory

WEAK_FRSI does not block generation — it's informational

Per-gene dedup

Max 3 GPU jobs per gene (watchdog enforces)

No on-demand GPU

All Spot — no spending without PI approval

Evidence chain

Every gate decision SHA-256 hash-linked, tamper-evident

Quantum gate = human-approved

VQE only fires after explicit PI approval + budget

No downgrade without bug check

System checks for error/timeout/OOM/spot before any grade reduction

🚦 Pipeline Gates — Quality Control

Active Gate System (gpu-batch-rs)

Chemistry Danger Patterns (30 rules — G1)

NPEAD Gate Learning (Adaptive — G6)

Capsule Influence on Gates

Cost Escalation Path

Grade Assignment (post-gate)

Governance Rules

Live Gate Performance