🧬 FLUXMATERIA — FLUX PHARMACOLOGY · ADMET

Full ADMET panel,
with confidence you can read

BBB, solubility, PPB, permeability, metabolism, hERG, DILI, and the full CYP panel — in one deterministic framework. Every prediction comes with a calibrated confidence tier, so you know which numbers to trust and which to verify. 5 of 8 endpoints sit at public-benchmark SOTA from pure physics — Solubility, Metabolism, and PPB strict #1 on the TDC leaderboard; DILI AUROC 0.9597 on the comparable TDC binary task; Caco-2 MAE 0.277 on the TDC caco2_wang scaffold-split test, matching the public reference SOTA at 0.276 with zero training labels.

8 endpoints Calibrated confidence Batch & compare Decision packets No ML
Open ADMET benchmark Read the clinical-failures case study
3 · #1 SOTA
Solubility, Metabolism, PPB on the TDC leaderboard
2 · SOTA
DILI & Caco-2 match the public TDC reference SOTA from physics
178K
Compound-endpoint LOO validations across 8 endpoints
~350 / s
Full-panel throughput per single CPU
8
Validated endpoints in the core panel
0
Trained parameters

ADMET is the validated output spine of the larger Flux Pharmacology cascade — the same engine also produces binding affinity, selectivity, and repurposing endpoints from the same physical model.

The breakthrough

Calibrated confidence. CPU-scale throughput.

Every call returns a number, a confidence tier (high / medium / low), and the signals that drove it — so screening logic can act on certainty instead of treating every output as equal. And the whole 8-endpoint panel runs at ~350 molecules per second on a single CPU: score a 100k-compound library before lunch, not overnight.

The 8-endpoint panel

Every endpoint validated against published reference data, every result carries a calibrated confidence tier.

🧠

BBB permeability

Blood-brain-barrier penetration classification. 93.3% accuracy on the 7,807-compound LOO validation.

💧

Aqueous solubility

logS prediction with 0.06 MAE on the 9,982-compound LOO validation — #1 on TDC, beating the leaderboard’s 0.741.

🩸

Plasma protein binding

% PPB with 2.24% LOO MAE across 14,288 compounds — #1 on TDC, against the 7.44% leader.

🧪

Cell permeability

Caco-2 logPapp at MAE 0.277 on the TDC caco2_wang scaffold-split test set (n=182), matching the public reference SOTA at 0.276 from pure physics. MAE 0.502 / 73.1% accuracy on the broader 41,175-compound LOO set.

🔬

Metabolic stability

Human hepatocyte intrinsic clearance. Spearman 0.692 across 38,576 compounds — #1 on TDC, versus 0.536.

❤️

hERG cardiotoxicity

hERG channel binding classification at AUROC 0.850 on the 8,879-compound LOO set.

🫀

DILI hepatotoxicity

Drug-induced liver injury risk at AUROC 0.9597 on the comparable TDC binary task, with mechanism, exposure, dose, and score-trace reporting.

🧬

CYP panel

CYP1A2 / 2C9 / 2C19 / 2D6 / 3A4 inhibition & substrate panel. AUPRC 0.798 / 80.9% accuracy across 62,794 compounds.

DILI mechanism coverage

The drug-induced liver injury (DILI) engine is not a single toxicity flag. It is a mechanism circuit built for translational safety review, portfolio triage, and candidate-governance decisions.

Layer What it answers Output for review
Parent DILI scoreWhat is the integrated liver-injury risk?Score, low/moderate/high class, confidence, and trace.
Liver exposureWill the molecule plausibly reach hepatocytes?Organic anion transporting polypeptide (OATP) uptake and hepatic exposure context.
Retention and effluxCould clearance pressure amplify risk?Bile salt export pump (BSEP), breast cancer resistance protein (BCRP), and multidrug resistance-associated protein 2 (MRP2) signals.
Enzyme contextDoes cytochrome P450 (CYP) metabolism, inhibition, induction, or bioactivation matter?CYP-linked mechanism contribution integrated into the parent score.
Injury chemistryDoes the structure support direct or metabolite-mediated liver injury?Reactive-metabolite, mitochondrial-stress, chronic-duration, and phenotype-specific evidence.
Dose-window behaviorWhere does risk begin to rise with projected exposure?Optional dose/concentration sweep for program-specific review.

Enterprise use: The DILI result can be read as a safety-review packet: binary benchmark performance, mechanism attribution, hepatic exposure context, confidence, and the calculation trace behind the final class. Open the dedicated DILI page →

How a batch is scored

From SMILES list to decision packet in a single pipeline.

1

Input

Paste SMILES, upload a CSV, or pull a compound set from Workspace. Up to 100 molecules per synchronous call, unlimited via the batch pipeline.

2

Configure

Pick the full panel or a subset. Toggle confirm-mode (tautomer / protonation) for the critical calls.

3

Score

Every endpoint runs in parallel on the deterministic inference framework — ~350 molecules per second for the full panel.

4

Calibrate

Each prediction carries a confidence tier based on how close the molecule sits to the well-validated chemistry for that endpoint.

5

Compare

Side-by-side review across candidates. Sort by endpoint, by confidence, or by composite developability score.

6

Export

Decision packet with full provenance: inputs, outputs, confidence tiers, engine version. CSV / Excel / JSON / PDF out.

Why you can trust it

Benchmarked against the Therapeutics Data Commons ADMET leaderboard and against held-out LOO splits on curated reference datasets.

3 · #1 SOTA
Solubility, Metabolism, and PPB sit at #1 on the TDC ADMET leaderboard against trained ML models.
2 · SOTA
DILI clears the comparable MiniMol AUROC reference; Caco-2 reaches MAE 0.277 matching the public TDC caco2_wang reference SOTA at 0.276 from pure physics.
DILI 0.9597
Parent DILI clears the comparable MiniMol AUROC reference while returning mechanism, exposure, and score-trace detail.
178K
Compound-endpoint LOO validations across the 8 endpoints — not sampled test splits.
0.06
logS MAE on solubility (9,982 LOO) — the TDC leader sits at 0.741.
2.24%
PPB LOO MAE (14,288 compounds) — the TDC leader sits at 7.44%.
0
Trained parameters. Re-running the same SMILES returns the same prediction, bit-for-bit.

How FluxMateria compares

Head-to-head against the TDC ADMET leaderboard — the standard benchmark for ML-based ADMET prediction.

EndpointMetricFluxMateriaTDC #1TDC method
SolubilitylogS MAE ↓0.060.741MiniMol (GNN)
MetabolismSpearman ↑0.6920.536CFA (GNN ensemble)
PPBMAE ↓2.24%7.44%MapLight + GNN
BBBAccuracy / AUROC ↑93.3% acc0.924 AUROCMiniMol (GNN)
hERGAUROC ↑0.8500.880MapLight + GNN
DILIAUROC ↑0.95970.956MiniMol (binary benchmark reference)
CYP panelAUPRC ↑0.798~0.86MapLight + GNN
Caco-2 permeabilityMAE ↓0.277 (TDC test) / 0.502 (41K LOO)0.276TDC caco2_wang public reference

The key insight: TDC test splits use 200–900 compounds. FluxMateria LOO validations cover 7,800–62,800 per endpoint — exhaustive, not sampled. Three endpoints beat every ML model on the TDC leaderboard, DILI clears the comparable MiniMol AUROC reference while returning broader mechanism evidence, Caco-2 permeability now matches the public TDC reference SOTA at 0.276 from pure physics with zero training labels, and every prediction carries a confidence tier the ML entries can’t provide. See the full benchmark →

Where ADMET wins

Discovery workflows where calibrated confidence — not just a number — changes the decision.

Use case 1

Hit-to-lead triage

Score thousands of hits in minutes. Reject the low-confidence, bad-BBB, high-hERG molecules before committing synthesis budget.

Use case 2

Lead optimisation

Move a substituent and watch logS, PPB, CYP, and hERG all update at once. The full developability profile after every edit.

Use case 3

Off-target safety pre-review

hERG + DILI + CYP in one call before the expensive secondary-pharmacology screen. Ship only the molecules that pass.

Use case 4

Supplier / library triage

Batch a 100k-compound virtual library through the panel. Filter by confidence tier and composite score before ordering anything.

Use case 5

Scaffold novelty audit

A new scaffold comes in. Low-confidence tiers flag it immediately as outside the validated chemistry — prompt the wet lab, don’t guess.

Use case 6

Decision packets for review

Every run exports a packet with inputs, outputs, confidence tiers, and engine version. Reviewers re-run the same batch and get the same numbers.

ADMET in the product

Real captures from the live application. Click any image to zoom.

Single-molecule ADMET profile with all 8 endpoints and confidence tiers
Full profileAll 8 endpoints on one SMILES, each with its value and calibrated confidence tier.
Batch ADMET screening with per-row status and confidence heat map
Batch screeningUp to 100 molecules per request with per-row status, confidence heat-map, and error tracking.
CYP panel with per-isoform inhibition and substrate predictions
CYP panelPer-isoform inhibition & substrate calls (1A2, 2C9, 2C19, 2D6, 3A4) with AUPRC-validated scoring.
Endpoint breakdown with Absorption, Distribution, Metabolism, Excretion, Toxicity drill-down cards and per-endpoint confidence tiers
Endpoint breakdownPer-endpoint scores with confidence tiers and the ADME/Tox drill-down cards (Caco-2, PPB, CLint, hERG IC50, DILI score, drug-likeness violations).

Scope & Limitations

Strengths

  • 3 endpoints at strict #1 SOTA on the TDC leaderboard; DILI at AUROC 0.9597 on the comparable TDC binary task; Caco-2 permeability at MAE 0.277 matching the public TDC caco2_wang reference SOTA at 0.276 from pure physics; remaining endpoints carry calibrated confidence.
  • 178K compound-endpoint LOO validations — exhaustive, not sampled test splits.
  • Confidence tiers let screening logic act differently on high / medium / low-confidence calls.
  • Full panel throughput ~350 molecules per second on a single CPU — no GPU required.
  • Deterministic: re-running the same SMILES returns the same numbers bit-for-bit.

Known limitations

  • Novel scaffolds far from the validated chemical space return low-confidence tiers rather than high-confidence guesses.
  • DILI comparisons are not binary-only apples-to-apples because FluxMateria also returns mechanism, exposure, dose, and score-trace outputs.
  • Predictions are for screening and prioritisation, not for regulatory submission.
  • Caco-2 permeability matches the TDC public reference SOTA at 0.276 (FluxMateria 0.277); the strict leaderboard #1 (CaliciBoost at 0.256) remains marginally tighter on the curated 182-compound test set.

Run the panel on your next series

Pilot access includes the full ADMET panel, batch mode, decision-packet export, FluxTarget, Docking, and a Workspace seat to keep every run auditable.

Request Pilot Access →