ToxScreen

Before you wet-lab, run the diagnosis.

A computational-first screening read for protein-binding operations: is the target trustworthy, are your compounds real binders or PAINS artifacts, and which assay should you actually spend money on?

The adversarial-7-gate philosophy

Anyone can train a model to fit its own training set. The hard part is proving a prediction survives adversarial, independent, public scrutiny. Every target on our panel must pass seven gates before it is allowed to emit a single lead:

1. Receptor provenance

The structure is the right protein, right chain, right species — not a homologous bait-and-switch.

2. Calibration

Spearman ρ ≥ 0.70 against measured affinities on ≥ 8 independent points, with the confidence interval open.

3. Discrimination

AUC-ROC ≥ 0.80 and it must beat a simple ECFP4 memorization baseline on decoys.

4. Neg-control FPR

Bland metabolites and safe drugs must not trigger as "binders" (FPR ≤ 10%). This is the anti-NLRP3 gate.

5. Enrichment

Real actives must rank above decoys at the head of the list (EF1%), not just beat random in aggregate.

6. Blinded holdout

A set of powerful actives held out of every calibration step is scored blind and must clear the threshold.

7. Replication

Scores are reproducible: replicate calls agree within tolerance and a checkpointed bundle exists.

Only when all seven gates PASS is a target GREEN. Unknown is treated exactly like Fail for lead-emission: no number is published that has not earned its gate. This is the discipline that catches the errors below — and it is the same discipline we apply to a client's target before they spend wet-lab money.

The honest-failure library

Our credibility is not that we never failed — it is that we publish the failures, the root causes, and the fixes. If you are going to trust a number with CRO budget, you should know exactly where the method has already broken.

  • CYP over-claim, retracted. Our published CYP3A4/2D6/2C9 AUROC 0.94–0.95 was in-distribution only. An independent ChEMBL cross-dataset test found CYP3A4 0.47 and CYP2C9 0.54 — statistically indistinguishable from chance — and CYP2D6 inconclusive. We published the correction and now label the CYP panel as research-grade. Do not use it to make spend decisions on novel chemistry.
  • Ensemble trap (AU21). For months, "ensemble=3" measurements were actually single samples because the backend silently ignored the sample count. We found it, fixed it in 17 files, added 9 regression tests, and re-labeled every historical threshold as single-sample rather than quietly pretending it was an ensemble.
  • Contaminated knock-out panel, retracted. A pooled n=89 "PASS" was retracted when the panel was found contaminated, and the class-label rule was changed so nothing is inferred from a name. The revised n=46 result is INCONCLUSIVE (underpowered) — we say so rather than dressing it up.
  • PAINS-heavy "novel" chemotypes. 19 of 23 supposedly-novel scaffolds tripped PAINS/pan-assay alerts. We now flag PAINS before any lead claim, and we do not sell "novel" claims we do not have.
  • Single-sample thresholds labelled as ensembles. Every counted threshold is now honestly labelled single-sample until re-derived with true replicates.

The pattern that produces these is a single one: a threshold that looks principled but is aimed at a quantity the experiment does not actually use. Our discipline — measure the statistic the experiment uses, open the confidence intervals, require the adversarial gates — is the method, and it is what you are buying in a paid diagnosis.

What the diagnosis returns

Give a target and a compound list; get a plan, not just a number.

  • Target verdict — GREEN / SCREEN_ELIGIBLE / BLOCKED / NOT_IN_PANEL, with the gate evidence exploded per gate.
  • PAINS / BRENK / pan-assay flags — the classic source of false "hits" that waste CRO budget.
  • Calibrated binder expectation — how many of your compounds would clear a decision-grade threshold (where the target is validated).
  • Assay guidance — should you wet-lab this, and which assay class (count-based vs biophysical) is the right first spend.
  • Go / no-go read — an honest recommendation, including "do not spend wet-lab money on this target yet".

The free diagnosis is read-only from our validated panel and ADMET module. The paid tier adds the full campaign: powered affinity calibration, selectivity against anti-targets, pocket-knockout verification, blinded holdout, and a reproducible, versioned report you can hand to an investor. That is the computational due-diligence engagement, delivered with ProjectAlpha (catalyticgpu.com) compute.

Try the free diagnosis Request the paid diagnostic

Why none of this is held back

We publish the methodology and the internals: the gates and their bars, the null models, the abstention cut-offs, the pocket-knockout procedure, and the failures the whole apparatus was built to catch. The method was never the moat. Anyone can read how a pocket-knockout control works — the hard part is running it against your own result and publishing the answer when it comes back inconclusive. What a client buys is execution and judgement under that pressure, not access to a secret. The full method is on the methods page.