What sits behind each prediction on this site, how it was measured, and where it breaks.
Every number the tools on this site produce comes from one of the models below. This page
exists so that a number can be judged rather than simply believed. For each model it states the
training corpus, the evaluation split, the measured score where one exists, and the situations in
which the model should not be relied on.
Two conventions are worth stating up front. First, where no held-out evaluation has been run,
the card says so plainly instead of substituting a training-set score. An unvalidated model is
still useful for ranking and triage, but it should not be quoted. Second, the reported figures are
not all the same kind of measurement. Scaffold-split held-out scores and cross-validation scores
are both reported here, and the cross-validation numbers are systematically more optimistic, so
scores are not comparable between models unless the split matches.
Proprietary models used in collaborations are documented separately and are not listed here.
Molecular property models
These predict something about the compound itself,
its physical properties, its ADMET behaviour, its liability to inhibit a metabolising enzyme.
Models that predict potency at a named protein are listed in the target table below instead.
Blood-brain barrier penetration
bbb v1
Probability that a compound is BBB penetrant, P(BBB+).
EvaluationAUROC 0.960 on the original labels and 0.965 on the cleaned labels, from leakage-free five-seed cross-validation.
Leakage-free means every B3DB and Martins near-duplicate twin, matched on the first 14 characters of the InChIKey, was removed from training. The gap between the two figures is a label-cleaning choice that is worth at most 0.005 and does not reliably improve the model, so 0.96 is the number to quote. The deployed model is then retrained on all 8,033 compounds with no holdout, so real-world performance should be at least this.
Method
Standalone CatBoost on MapLight features. No teacher ensemble.
Training data
The deduplicated CNS-8K corpus, 8,033 unique compounds (5,169 BBB+ and 2,864 BBB-), built from a corrected B3DB plus TDC Martins.
Where not to trust it
Trained on a binary penetrant or non-penetrant label, which compresses a continuous and context-dependent property. It says nothing about unbound brain concentration, efflux liability, or whether enough compound reaches the target to matter. For a CNS program treat it as a triage filter, not an endpoint.
ADMET panel
admet_tdc v1
22 absorption, distribution, metabolism, excretion and toxicity endpoints.
EvaluationNot benchmarked on this site
Trained and scored against the TDC leaderboard tasks. Per-endpoint scores vary widely across the 22 tasks and have not been consolidated into a single figure on this site.
Method
CatBoost with a 200-teacher ensemble, a GIN graph network, and CATEExpert blending.
Training data
The Therapeutics Data Commons ADMET benchmark collection.
Where not to trust it
ADMET endpoints are assay-specific and often species-specific. Clearance, half-life and microsomal stability numbers do not transfer between human and rodent matrices. The 22 tasks differ greatly in training-set size, so the weakest endpoints are considerably less reliable than the strongest and the panel does not currently distinguish them for you.
Cytochrome P450 inhibition
cyp v1
Probability of inhibition for five CYP isoforms.
EvaluationNot benchmarked on this site
Selected as the winner of an internal five-fold out-of-fold PRAUC bake-off. The winning PRAUC value itself was not recorded in the model configuration, so no number is claimed here.
Method
CatBoost on MapLight-2563 features plus a 300-dimension GIN, class-balanced.
Training data
Public CYP inhibition data. Corpus size is not recorded in the model configuration.
Where not to trust it
Predicts inhibition, not induction, and not time-dependent or mechanism-based inactivation, which is often the more serious drug-drug interaction risk. A clean prediction across all five isoforms is not a clean DDI profile.
Physicochemical descriptors
physchem v1
Molecular weight, tPSA, cLogP, QED, hydrogen bond donors and acceptors, rotatable bonds.
EvaluationNot benchmarked on this site
Not applicable. An exact calculation has no held-out error.
Method
RDKit, computed directly from the structure. No model weights and no training.
Training data
Not applicable. These are exact calculations, not predictions.
Where not to trust it
cLogP and tPSA are themselves empirical estimators with published error bars, so treat them as good approximations rather than measured values. QED is a composite desirability score and encodes an opinion about what oral drugs look like, which does not transfer to every modality.
Target potency models
Every protein this site can predict activity against, 798 in all.
They come from two places. 1 are curated models behind the drug design dashboard, built
and tuned for one target each. The rest are the screening fleet: one model per gene trained on
that target's ChEMBL data by a uniform pipeline, evaluated on a prospective split by compound registration order, with the most recent 20 percent held out, which asks the model
to predict compounds registered after the ones it trained on. That is harder than a random split
and the scores are correspondingly lower than random-split numbers published elsewhere.
0.50median Spearman
0.27 to 0.67interquartile range
389at or above 0.50
215below 0.30
The spread is the point. Check the target you care
about before trusting a ranking; quality tracks how much measured data the target has, and that
varies by three orders of magnitude across this collection.
Target
Source
Training compounds
Held-out Spearman
Loading the target table.
Applicability domain
Alongside each prediction the tools compute an applicability-domain score: the
Morgan fingerprint Tanimoto similarity between the query compound and the nearest compounds in that
model's training set. A low score means the molecule sits outside the chemistry the model learned
from, and the prediction should be treated as an extrapolation regardless of how confident it looks.
This matters most for generated or heavily modified structures, which routinely fall outside the
domain while still returning a plausible number.
Known gaps
Stated so they are not mistaken for oversights. Several models carry no held-out
evaluation, and are marked as such above. Reported scores come from different split protocols and are
therefore not directly comparable across models. None of these models has been evaluated on the
Polaris community benchmarks,
which would allow neutral comparison against published methods; doing so is planned. Prospective
validation, meaning experimental confirmation of predictions made before the experiment, is reported
in the group's publications rather than here.
How to cite
If a result from these tools contributed to your work, please
cite the site and record the date, since models are updated in place.
@misc{hubbslab_tools,
author = {Hubbs, Jed L.},
title = {Hubbs Lab computational drug discovery tools},
howpublished = {\url{https://hubbslab.com/methods}},
institution = {Boston Children's Hospital and Harvard Medical School},
note = {Accessed: YYYY-MM-DD},
year = {2026}
}
Generated from the live model registry on 2026-07-28. Regenerate with
python3 lab-api/gen-methods.py after any registry change.