ToxScreen Blog

Computational toxicology, drug discovery, and the science behind calibrated ML ADMET safety screening.

product update admet-ai adme safety

Extended ADMET Panel: 15+ Endpoints in Every Report

June 2026

ADMET-AI computes 49+ molecular properties on every call. ToxScreen previously surfaced only 4 — hERG and the CYP panel. Every report now includes 15+ supplementary endpoints across extended CYP (CYP1A2, CYP2C19), absorption (HIA, Caco-2, P-gp), distribution (BBB, PPB, Vdss), excretion (microsomal clearance, half-life), and safety flags (Ames/ICH M7, DILI with calibration caveat, carcinogenicity, skin sensitization, LD50). All endpoints carry Tanimoto applicability-domain badges — IN DOMAIN / BORDERLINE / OUT OF DOMAIN — so predictions on structurally novel chemistry are labelled for what they are.

Read more →
machine learning calibration novel chemistry

Why Honest-ADMET Wins on Novel Chemistry

June 2026

Most ML ADMET models degrade silently on novel scaffolds — they extrapolate without saying so. Honest-ADMET combines isotonic calibration, Tanimoto applicability-domain abstention, and Wilson confidence intervals to close the hERG AUROC gap from 0.69 (random split) to 0.84 (out-of-domain hold-out) while telling you exactly when to validate at the bench.

Read more →
calibration admet-ai validation

How We Calibrate ADMET-AI for Tox Pre-Screening

June 2026

ToxScreen's tox panel runs ADMET-AI, an open-source graph neural network, wrapped with isotonic probability calibration and Tanimoto applicability-domain abstention — so it can tell you when your chemistry is too novel to trust. We publish the real per-target AUROC numbers (random split, n ≈ 1,800 per target): CYP3A4 0.95, CYP2D6 0.94, and CYP2C9 0.94 all PASS, while hERG is the hard one at 0.69 on a random split (0.71 ADMET-AI baseline, 0.84 ToxScreen calibrated on novel chemistry) — reported honestly rather than dressed up. CYP inhibition is strong; hERG leans on AD-abstention instead of guessing.

Read more →
drug discovery tox screening workflow

Why Computational Toxicology Deserves a Place in Your Screening Cascade

June 2026

A full hERG patch-clamp plus CYP inhibition panel runs into six figures for an early-stage hit series and takes weeks to complete. A calibrated ML ADMET pre-screen costs a fraction of that and returns results in seconds. Beyond cost and speed, computational toxicology addresses the EU 3Rs directive on animal welfare and aligns with the growing consensus against cosmetic animal testing. On accuracy: ToxScreen's CYP endpoints reach AUROC 0.94–0.95, strongly separating inhibitors from non-inhibitors, while hERG is harder and handled with conservative calibration plus abstention. We advocate a “screen before you synthesise” philosophy, positioning computational tox as a triage layer ahead of expensive in-vitro and in-vivo follow-up.

Read more →
natural products coconut campaign longevity

COCONUT R1: Screening Nature's Chemical Space for Longevity Targets

June 2026

The COCONUT database (Collection of Open Natural Products) contains over 400,000 unique natural product structures released under a CC0 public domain license. Using the EstimaBio Boltz-2 pipeline (a structure-based longevity research program, distinct from the ToxScreen tox engine), we screened a top-priority tier of 459–493 compounds per target across an 8-target longevity panel. Hit rates ranged from 3.7% (PARP1) to 37.7% (NLRP3 — see update note in full post). A triple-hit chalcone (ZAVWFOFRECTXEE) was flagged by PAINS filters, reinforcing the importance of substructure alerting. Results are in-silico; wet-lab IC50 confirmation is still pending. Correction (July 2026): the CD38 calibration and selectivity claims in this post were withdrawn after the CD38 reference set was traced to the wrong protein; on the corrected set CD38 fails validation and is BLOCKED. PARP1 is currently the only validated longevity target — see the post's correction note.

Read more →

Get notified when we publish

New calibration data, campaign results, and methodology updates. No spam, no third-party sharing.

Subscribe via email →

Or email info@toxscreen.ai directly.