Back to benchmarks | ADMET

CYP3A4 Inducer Prediction

A public benchmark card for the canonical FluxMateria v25 CYP3A4 inducer route: on-the-fly Flux physics scoring, external holdout validation, and explicit limitations.

0.9350
primary external BA
0.7778
primary precision
0.8750
primary recall
0.9950
primary specificity
This benchmark reports external-holdout performance for the canonical v25 physics route. It is not a clinical-grade DDI validation or a regulatory submission benchmark.

Benchmark Results

External holdout validation for the canonical v25 physics route.

Model Balanced acc Accuracy Precision Recall Specificity F1 TP / FP / FN / TN
v25 primary external holdout 0.9350 0.9926 0.7778 0.8750 0.9950 0.8235 7 / 2 / 1 / 395
v25 all scored rows 0.9652 0.9956 0.8750 0.9333 0.9970 0.9032 14 / 2 / 1 / 669

The primary row excludes compounds overlapping the current reference snapshot and reports only primary-metric-eligible cases. The all-scored row is shown as supporting context, not as the headline claim.

What the model predicts

CYP3A4 induction is treated as a mechanism-aware risk classification problem.

The v25 module does not treat CYP3A4 induction as a single catalytic-site binding score. It computes ligand Flux state, PXR/CAR/direct-induction compatibility, exposure support, hard-negative pressure, and binder-not-agonist pressure before assigning the final risk tier.

Mechanism branches

PXR, CAR, and direct induction evidence are kept as separate signals before final calibration.

Local chemistry

Flux molecular-state descriptors are computed directly from structure and projected onto fixed receptor and process-compatibility terms.

ADMET context

Exposure and chemistry-context terms constrain the mechanism score so binder-like molecules are not automatically treated as inducers.

Dataset and validation

Dataset

  • 686 scored external-holdout rows
  • 405 primary-metric-eligible rows after overlap filtering
  • 14 positives and 672 negatives across all scored rows
  • FDA/DailyMed and source-backed clinical-label evidence

Production rule

  • v25_physics is the default production route
  • the route scores from molecular structure at query time
  • risk tiers separate high-confidence positives from review cases
  • benchmark labels are used for validation scoring

Scope and limitations

Supported claim

FluxMateria v25 is a production CYP3A4 inducer prioritization model that reaches 0.9350 balanced accuracy on the primary external holdout for the canonical Flux physics route.

Not claimed

  • external SOTA on CYP3A4 induction
  • clinical-grade DDI prediction
  • regulatory validation
  • direct CYP3A4 catalytic-site binding prediction

Next validation steps

The next meaningful improvements are broader validation and evidence-depth improvements.

  • larger external holdout from sources outside the current reference snapshot
  • richer potency, efficacy, cytotoxicity, and curve-class fields
  • prospective partner or wet-lab validation on newly selected molecules
  • separate public benchmarks for PXR, CAR, and final CYP3A4 induction
  • calibration curves and class-specific error analysis by scaffold family
Download summary JSON ADMET benchmark

Use CYP3A4 induction as a risk signal

The v25 physics route is designed for prioritization and triage. For workflows where missing an inducer is more costly than reviewing extra positives, use the risk tier and review flags alongside the binary call.

ADMET module Request access

Benchmark basis

The canonical CYP3A4 inducer route is the v25 on-the-fly Flux physics scorer. Reference labels are used to score benchmark performance, not to produce predictions.

Flux Physics