About
Honest computational toxicology
ToxScreen is a product of EstimaBio — an AI-driven drug discovery platform focused on longevity therapeutics from natural products. ToxScreen delivers calibrated machine-learning ADMET predictions that flag hERG and CYP liabilities — and tell you when your chemistry is too novel to trust.
The story
EstimaBio runs two distinct programs. Our longevity research program searches natural-product space for novel compounds that bind longevity targets — but only targets that clear an adversarial validation scorecard (discrimination, negative-control, blinded holdout) can emit a lead, and every novel hit must prove selectivity against a panel of unrelated proteins, not merely bind. Most candidate targets fail that bar; we say which, and why. ToxScreen is a separate product: a calibrated machine-learning ADMET panel that flags hERG and CYP liabilities. They are different engines for different jobs — we do not run tox predictions through the longevity pipeline.
ToxScreen runs honest-ADMET: predictions come from ADMET-AI (an open-source chemprop message-passing graph neural network for ADMET), wrapped with our own layers — isotonic probability calibration so a score reads as a real probability, and a Tanimoto applicability-domain (AD) abstention step that flags low-confidence inputs and abstains when a compound falls outside the model's training distribution. RDKit physicochemical descriptors (MW, LogP, TPSA, PAINS) run as a pre-flight sanity layer. Inference is deterministic and CPU-based — sub-10-second results per compound.
What makes ToxScreen different
- Calibrated probabilities, not raw scores (isotonic calibration)
- Applicability-domain abstention — it tells you when it doesn't know
- Built on ADMET-AI; CYP inhibition is strong in-distribution (AUROC 0.94–0.95)
- Honest about hERG — 0.69 random split, 0.84 calibrated pipeline on de-leaked novel chemistry
- Deterministic CPU inference — sub-10-second results, API / SDK / CLI
- Zero mock data — no fabricated scores, ever
Separately, our longevity research program screens natural products against a structure-based target panel. That work is computational / in-silico; its calibration and limitations are documented in our lab notebook and are not part of the ToxScreen tox panel.
We publish our methodology transparently. Every threshold, every reference, and every known limitation is documented on the Methodology page.
Technology
The stack behind every report
ADMET-AI
ToxScreen's prediction engine. ADMET-AI (Swanson et al., Bioinformatics 2024) is an open-source chemprop message-passing graph neural network for ADMET. It produces the hERG and CYP inhibition probabilities in every result — trained on molecular structure, not hand-picked descriptors.
Calibration & applicability domain
Raw outputs pass through isotonic calibration so a score reads as a true probability, then a Tanimoto AD check classifies each compound IN DOMAIN, BORDERLINE, or OUT OF DOMAIN. ToxScreen abstains rather than extrapolate. Aligns with OECD QSAR Principle 3 and EMA 2024 NAMs guidance.
RDKit pre-flight
Before scoring, RDKit computes physicochemical descriptors — MW, LogP, TPSA, HBD/HBA, QED, PAINS matching, Lipinski violations — as a sanity layer that catches malformed or red-flagged structures.
Composite Tox Index (CTI)
A single weighted score [0, 1] aggregating the calibrated per-endpoint liabilities. Weights reflect clinical priority: hERG 0.35, CYP3A4 0.25, CYP2D6 0.20, CYP2C9 0.20. Five risk bands (NEGLIGIBLE → CRITICAL) drive the traffic-light verdict.
Extended ADMET panel
Every report includes 15+ supplementary ADMET-AI endpoints: CYP1A2/2C19, Ames (ICH M7), Caco-2, P-gp, oral bioavailability, BBB, PPB, Vdss, clearance, half-life, carcinogenicity, skin sensitization, LD50 and DILI (with explicit calibration caveats).
Deterministic & fast
Inference is deterministic and CPU-based — the same compound always returns the same result, in roughly one to seven seconds. No queue, no random seeds. Available via web app, REST API, Python SDK, and CLI.
Validation
We report the numbers as they are
| Endpoint | PDB | AUROC (in-dist.) | n (pos + neg) | Status | Novel-chemistry AUROC |
|---|---|---|---|---|---|
| hERG | 7CN1 | 0.69 | 875 + 925 | FAIL (random split) | 0.84 (Karim, de-leaked) |
| CYP3A4 | 6MA7 | 0.95 | 727 + 1073 | PASS (in-dist.) | 0.47 (ChEMBL, n=701, ~chance) |
| CYP2D6 | 4WNT | 0.94 | 381 + 1419 | PASS (in-dist.) | inconclusive (99% single-class) |
| CYP2C9 | 1OG5 | 0.94 | 671 + 1129 | PASS (in-dist.) | 0.54 (ChEMBL, n=237, ~chance) |
In-distribution vs. novel chemistry — and we say which is which. CYP3A4 / CYP2D6 / CYP2C9 reach AUROC 0.94–0.95 in-distribution (scaffold split of the same TDC data ADMET-AI trained on). An independent ChEMBL cross-dataset test found CYP3A4 and CYP2C9 AUROC statistically indistinguishable from chance (0.47, 0.54); CYP2D6's test was inconclusive. hERG is a weak 0.69 on a random split, but the calibrated ToxScreen pipeline reaches 0.84 on de-leaked novel chemistry (ADMET-AI baseline: 0.71) — currently the only endpoint with a validated novel-chemistry number. See the Methodology page for full data.
Read the lab notebook
Every threshold, every reference compound, every calibration curve, and every known limitation is published openly. We do not hide behind black boxes.
Read the full scientific digest →Methodology, calibration data, known limitations, references, privacy & legal terms.
Important disclaimers
Research use only. ToxScreen is a computational pre-screening tool. Predictions are not a substitute for in vitro or in vivo toxicology studies. Decisions affecting human or animal exposure must be supported by experimentally validated assays.
Not a regulatory submission tool. ToxScreen output cannot replace IND-enabling studies, GLP toxicology packages, or any FDA / EMA / PMDA-mandated assay.
No medical advice. ToxScreen does not provide diagnosis, treatment recommendations, or clinical guidance.