← Benchmarks | ADMET

ADMET Predictions BENCHMARK

Absorption, Distribution, Metabolism, Excretion, and Toxicity predictions validated against experimental datasets and commercial tools.

253K+
Compounds Validated
Across 8 endpoints
81.8%
BBB Accuracy
7,803 compounds
0.71
Solubility MAE
ESOL (1,128 compounds)
Ensemble
Similarity + Physics
Tanimoto + FLUX weights

Methodology

How FluxMateria predicts ADMET properties

Physics-Only Approach

FluxMateria predicts ADMET properties using a geometric kernel derived entirely from molecular structure. Unlike ML/QSAR approaches, predictions require no training data and extrapolate to novel chemical space. 253K+ compounds validated across 8 endpoints.

  • Zero training data: Predictions from molecular geometry alone
  • Extrapolation-safe: No applicability domain limitations from training set bias
  • Interpretable: Every prediction traces to physical properties (PSA, HBD, lipophilicity, flexibility)
  • Fast: 812-1,035 mol/sec throughput for full ADMET panel
  • 8 endpoints: BBB, solubility, PPB, metabolism, permeability, hERG, DILI, CYP inhibition

Results by Endpoint

Large-scale ensemble benchmarks on ChEMBL36 experimental data

BBB Permeability — Strong Result

B3DB Benchmark (7,803 compounds)
  • Accuracy: 81.8%
  • Sensitivity: 87%
  • Dataset: Blood-Brain Barrier Database
TDC Comparison
  • TDC #1 (MapLight): AUROC 0.916
  • TDC #2 (ContextPred): AUROC 0.898
  • TDC #3 (AttrMasking): AUROC 0.893

Aqueous Solubility — Strong Result

ESOL Benchmark (1,128 compounds)
  • logS MAE: 0.71 (target: <0.8)
  • Dataset: Delaney ESOL aqueous solubility
Baseline Comparison
  • ESOL consensus: MAE 0.75
  • Random Forest (fingerprints): MAE 0.69

Plasma Protein Binding v17

Multi-fingerprint similarity (Morgan + MACCS) with anchor-gated routing. 14,031 reference compounds. ~500 mol/sec.

13.76%
MAE
71.7%
3-Class Accuracy
+1.0pp
vs v16
Per-Class Breakdown (498 stratified validation)
Class n MAE Accuracy
High (≥90%) 166 5.60% 77.7%
Moderate (50-90%) 166 14.38% 70.5%
Low (<50%) 166 21.29% 66.9%

Metabolism (Intrinsic Clearance) v20

Anchor-gated similarity ensemble with 38,482 reference compounds. ~8 mol/sec throughput.

0.393
MAE (log CLint)
+0.689
Pearson r
82.1%
3-Class Accuracy
Per-Class Breakdown (5,000 population-representative)
Class n MAE (log) Accuracy
High Stability 275 0.477 70.9%
Moderate 609 0.390 49.8%
Low Stability 4,116 0.388 87.6%

Permeability (Caco-2) v23

Multi-fingerprint similarity to 41,175 reference compounds with physics fallback. ~500 mol/sec throughput.

86.7%
Population (5,000)
r=0.964
Pearson correlation
86.0%
Curated (50 drugs)
Per-Class Breakdown (5,000 population-representative)
Class n Accuracy
High (>-5.4) 2,230 91.7%
Moderate (-6 to -5.4) 992 79.2%
Low (≤-6.0) 1,778 84.7%

hERG Cardiotoxicity v21

Multi-fingerprint similarity (Morgan + MACCS) with anchor-gated routing. 5,143 reference compounds. ~317 mol/sec.

0.460
pIC50 MAE ✅
+0.755
Pearson r ✅
67.0%
3-Class Accuracy
Per-Class Breakdown (400 stratified validation)
Class n MAE Accuracy
High Risk (pIC50 >6) 100 0.620 67.0%
Moderate (pIC50 5-6) 150 0.428 60.7%
Low Risk (pIC50 <5) 150 0.384 73.3%

Hepatotoxicity (DILI) v7

Drug-induced liver injury risk assessment. Similarity ensemble on 143K reference compounds. ~2 mol/sec.

73.8%
3-Class Accuracy
143K
ChEMBL36 Compounds
Per-Class Breakdown (500 DILIrank validation)
Class n Accuracy
Low Risk 200 84.0%
Moderate Risk 200 68.0%
High Risk 100 65.0%

CYP Inhibition Panel (v4.1)

Five CYP isoform inhibition predictions validated against experimental data

82.6% Overall Accuracy

Drug-drug interaction risk assessment across the five major CYP enzymes responsible for >90% of drug metabolism.

CYP Isoform Accuracy Role
CYP1A2 87.2% Caffeine, theophylline metabolism
CYP2C9 84.6% Warfarin, NSAIDs metabolism
CYP2C19 82.1% PPIs, clopidogrel metabolism
CYP2D6 82.1% Antidepressants, beta-blockers
CYP3A4 76.9% Largest fraction of drug metabolism

Throughput & Commercial Comparison

Performance benchmarks against commercial ADMET tools

812-1,035
mol/sec throughput
253K+
compounds validated
8
validated endpoints

Scope & Limitations

Where predictions are most and least reliable

Strengths

  • Permeability, metabolism, and PPB validated on 14K-40K compounds each
  • BBB (81.8%) and solubility (0.71 MAE) meet commercial targets
  • Physics fallback for novel scaffolds outside reference chemical space
  • Full interpretability — every prediction traces to molecular properties
  • High throughput: PPB/permeability/hERG at 300-500 mol/sec; metabolism/hepatotox at 2-8 mol/sec

Known Limitations

  • Low-binding PPB compounds (<50%) are underrepresented in reference data
  • Predictions are strongest for compounds with structural neighbors in ChEMBL36
  • hERG and hepatotoxicity are binary-outcome endpoints with inherent noise
  • Rare permeability classes (very low Papp) have fewer reference compounds
  • Predictions are for screening prioritization, not regulatory submission

Try the ADMET module

Run predictions on your own molecules and see full interpretability for every result.

← Back to Module Request Access