Can AI flag chemicals that bind a hormone-carrying protein? An FDA-led model found useful signal—but only nine external non-binders
A consensus of four molecular models reached 0.430 Matthews correlation on an external set of 164 chemicals. Its 88.4% accuracy is easy to overread: 155 external records were binders, seven came from non-human assays and the model has not replaced laboratory testing or predicted harm in people.
The short version
- What happened
- FDA and University of Maryland researchers trained random-forest, support-vector, multilayer-perceptron and graph-neural-network models on 125 experimentally tested chemicals: 87 binders and 38 non-binders.
- Who it affects
- Chemical-safety scientists could use the model to prioritise scarce laboratory testing, but people are protected only if uncertain and high-consequence candidates still receive experimental assessment.
- What remains unproven
- The development dataset had 125 chemicals and the external dataset had only nine non-binders, making specificity and MCC estimates sensitive to a few cases.
Keep in mind
The development dataset had 125 chemicals and the external dataset had only nine non-binders, making specificity and MCC estimates sensitive to a few cases.
External labels combined different assays, activity types, thresholds, species and experimental conditions rather than one harmonised prospective test.
Seven external binder records came from rat or rabbit assays and all seven were false negatives for the consensus model.
Descriptor dissimilarity did not establish performance on genuinely novel chemical scaffolds, and no prospective screening workflow was tested.
SHBG binding is a mechanistic endpoint, not an adverse health outcome; exposure, potency and other biological pathways were outside the study.
Source analysis · 1 linked source. Primary analysed source date: 5 November 2026.
Read the full report · 7 min The evidence, methods, implications and limitations
The answer: the model could help prioritise tests, but it is not a chemical-safety verdict
An FDA-funded research team built five models to classify whether a chemical binds sex hormone-binding globulin, or SHBG, a blood protein that carries and regulates access to sex hormones. The most balanced model averaged probabilities from four different algorithms. On an external set of 164 chemicals, that consensus produced 89.0% sensitivity, 77.8% specificity and a Matthews correlation coefficient of 0.430. Those figures show useful predictive signal across an independently assembled dataset, especially for deciding which chemicals should move earlier into laboratory testing.
The result is not evidence that the model can declare a chemical safe, predict endocrine disease or replace a binding assay. SHBG interaction is one mechanistic endpoint among many. A molecule predicted to bind may not cause harm at realistic exposure, and a non-binding prediction says nothing about other hormone receptors or toxic effects. The external dataset was also extremely imbalanced: 155 records were labelled binders and only nine non-binders. That makes the headline 88.4% accuracy much less informative than the paired sensitivity, specificity and Matthews correlation reported by the authors.[1]
Why SHBG binding matters—and why experimental screening is difficult
SHBG binds circulating androgens and oestrogens. A chemical that competes for those binding sites could alter the balance between protein-bound and freely available hormone, while SHBG affinity can also affect how a drug or chemical is distributed in the body. That makes the endpoint relevant to endocrine-disruption screening and to development of steroid-like or endocrine-active medicines. It does not establish an adverse outcome by itself: biological effect depends on potency, exposure, metabolism, other targets and the organism involved.
Competitive binding assays require a validated protein source, labelled reference ligand, controlled equilibrium conditions, multiple concentrations and checks for interference, solubility and stability. Those demands limit throughput across large chemical inventories. Computational screening can rank candidates before scarce laboratory resources are used. The practical comparator in this study is therefore not a clinician or a complete toxicology assessment; it is the current need to select chemicals for more definitive experimental work.[1]
The distinction that matters
How to read the external validation
The overall accuracy is dominated by binders, so class-specific results and the denominator matter.
External set
164 chemicals
155 binders, 9 non-binders
Consensus
Predict binders
89.0% sensitivity
Consensus
Reject non-binders
77.8% specificity
Balanced summary
Combine both classes
MCC 0.430
How the five models were built and compared
The development dataset contained 125 chemicals previously tested in a human SHBG competitive-binding assay using pregnancy plasma and radiolabelled dihydrotestosterone: 87 binders and 38 non-binders. Three models—random forest, support-vector machine and multilayer perceptron—used 478 molecular descriptors derived from chemical structures. A graph isomorphism network represented atoms and bonds directly. The fifth model was a consensus that averaged the four individual prediction probabilities.
Model selection used repeated stratified cross-validation rather than one favourable train-test split. Candidate hyperparameters were compared with five iterations of five-fold cross-validation, and final performance was evaluated across 100 repeated five-fold splits. The authors report internal accuracy between 0.811 and 0.855. The consensus reached 0.855 accuracy internally and had the most balanced overall profile. Repeated splitting reduces dependence on one partition, but all internal folds still draw from the same small 125-chemical source dataset.[1]
The external test is valuable—and unusually easy to misread
For external validation, the team curated 164 non-overlapping chemicals from ChEMBL. Labels came from quantitative measurements such as dissociation or inhibition constants and from qualitative activity annotations. Of these records, 157 referred to human SHBG, six to rat and one to rabbit. The dataset contained 155 binders and nine non-binders because activity databases tend to record compounds with measurable interactions. The consensus correctly detected most binders and seven of nine non-binders, producing 89.0% sensitivity, 77.8% specificity and the highest equal-weight consensus MCC at 0.430.
When the seven non-human records were removed, the consensus reached 92.36% accuracy, 93.24% sensitivity, unchanged 77.78% specificity and an MCC of 0.5313. But all seven removed chemicals were binders that the model had missed, so this improvement also shows a species-transfer weakness. With only nine non-binders, each individual mistake changes specificity by more than eleven percentage points. The authors accordingly say specificity and Matthews correlation for non-binder detection are less stable than binder classification.[1]
What confidence scores and molecular diversity add
The study calculated prediction confidence and found that higher-confidence groups were generally more accurate. This can support triage: a laboratory might prioritise both likely binders and uncertain chemicals rather than treating every output equally. Confidence is still model-specific and was assessed on the same external records used for validation. Before operational use, a regulator or laboratory would need a predetermined action rule, calibration monitoring and evidence that confidence remains informative in the chemical inventory it actually screens.
The authors also compared molecular-descriptor similarity between development and external chemicals. Low average similarity suggests the external set was not merely a near-copy of the training set. But descriptor-profile dissimilarity is not the same as demonstrating new molecular scaffolds, and the paper calls for further scaffold-based validation. The external labels also combine different assay types, measurement thresholds, species and experimental conditions. Apparent model error can therefore reflect both genuine generalisation failure and inconsistency in the reference labels.[1]
What would be needed before regulatory or drug-development use
A stronger next test would prospectively select a balanced and chemically diverse panel, generate all reference labels with one prespecified human SHBG assay and keep the laboratory team blind to model predictions. Reporting should include confidence intervals, calibration, performance on novel scaffolds and the number of chemicals sent to follow-up testing. Comparison with existing quantitative structure–activity tools and simple similarity baselines would show whether the consensus adds enough value to justify maintenance and governance.
For chemical safety, researchers would then need to connect binding predictions with concentration, exposure and downstream biological effects. For medicines, developers would need to test whether prioritisation improves lead selection or dose interpretation without missing relevant compounds. None of those real-world outcomes was measured here. The defensible conclusion is that four complementary molecular models produced a useful research-screening signal on a difficult external dataset, with clear imbalance and species limitations—not that AI has solved endocrine-disruption assessment.[1]
Keep following the evidence
Join The Impact Brief founding list for selected AI reports, their effect on people and the evidence limits. Read a sample first.
Continue the story
Read next
Continue exploring machine learning: compare the finding and its evidence limits.
Can a six-protein AI score rule out serious fatty liver disease?
A multicohort study reports AUCs of 0.79–0.83 for a six-protein MASH score and a 0.99 negative predictive value in a low-prevalence UK cohort. Different reference standards, small biopsy cohorts and unavailable routine assays keep it from clinical readiness.
Read the next reportExplore the full source and evidence record
What this means for people
- Chemical-safety scientists could use the model to prioritise scarce laboratory testing, but people are protected only if uncertain and high-consequence candidates still receive experimental assessment.
- Drug developers may gain an early signal about SHBG interaction, while exposure, potency and clinical relevance still require laboratory and pharmacological evidence.
- Communities concerned about endocrine-active chemicals should not interpret a model prediction as evidence of harm or safety; the study tested molecular binding labels, not population health outcomes.
Global context
The model was developed by US FDA and university researchers using public chemical resources, but its intended screening problem is global. Laboratories differ in assay capacity, chemical inventories and regulatory requirements. Open scripts and supplementary data support replication, yet a model validated on historical records should be locally tested before it influences regulatory prioritisation or drug-development decisions.
What the evidence does not yet show
- The development dataset had 125 chemicals and the external dataset had only nine non-binders, making specificity and MCC estimates sensitive to a few cases.
- External labels combined different assays, activity types, thresholds, species and experimental conditions rather than one harmonised prospective test.
- Seven external binder records came from rat or rabbit assays and all seven were false negatives for the consensus model.
- Descriptor dissimilarity did not establish performance on genuinely novel chemical scaffolds, and no prospective screening workflow was tested.
- SHBG binding is a mechanistic endpoint, not an adverse health outcome; exposure, potency and other biological pathways were outside the study.
What to watch next
- Prospective validation on a balanced panel tested with one human SHBG assay.
- Performance and calibration on genuinely novel chemical scaffolds.
- Head-to-head comparison with simpler similarity and established QSAR baselines.
- Whether model confidence improves laboratory prioritisation without increasing missed binders.
- Evidence connecting predicted SHBG binding with exposure and downstream biological outcomes.
Living evidence record
Impact record IAI-0IQWIJD
Evidence stage
Studied
Confidence
Supported
Reporting basis
Source analysis
Independent or research support
Present
Record status
Monitoring
Last checked
6 November 2026
Source trail
1 direct source across 1 source type.
People impact
Documented in this record.
Uncertainty
Limits and next checks are explicit.
Stages describe the evidence available—not whether a technology is good or bad. See the public method.
Single-source reporting disclosure
This record analyses one direct source. It can establish what Frontiers in Chemistry published or reported, but it is not independent corroboration of every performance claim or predicted outcome. The confidence label will change only when broader evidence is added.
Evidence trail
Sources used for this report
Links checked 6 November 2026
Research question, topics and regions
Machine-learning classification of whether chemicals bind sex hormone-binding globulin using a 125-chemical training dataset and a 164-chemical external ChEMBL dataset
Topics: Chemical safety · Endocrine disruption · FDA research · Machine learning · Drug discovery · External validation
Regions: United States · North America · Global
This report is labelled source analysis. We summarise and analyse source material in our own words; company statements remain attributed claims until independently supported. Translated summaries preserve the meaning of the original source and link back to it. Read our editorial standards.
Reader commentary
What does this mean where you are?
Tell us what you have seen at work, in your community or in everyday life. A question is welcome too. No account is needed; we review notes before publishing them.
Explore commentary across the portal →Published reader notes
0No published reader notes yet. You can start the evidence-led discussion above.
Prefer a private correction or response? Contact the newsroom.