An AI Model Flags 72 Recently Approved Drugs for Potential Reproductive Toxicity

A new computational study has identified 72 drugs approved between 2015 and 2024 as potentially carrying a high risk of reproductive toxicity, offering a rapid way to screen medicines for effects that may be missed during conventional safety evaluations. The research, published in Molecular Diversity, combines machine learning with real-world reports from the US Food and Drug Administration’s Adverse Event Reporting System, or FAERS, and adds network toxicology, molecular docking and molecular dynamics simulations to investigate how the suspected risks might arise. The authors stress that the results do not prove that the drugs cause infertility, pregnancy complications or other reproductive harms. Instead, the work presents an integrated early-warning and mechanism-hunting framework designed to help researchers prioritize medicines for closer study.

Reproductive toxicity is particularly difficult to evaluate because it can involve several biological processes unfolding over long periods. A compound may affect sperm production, egg maturation, hormone signaling, implantation, fetal development or the health of offspring, sometimes at exposure levels that vary with dose, timing and individual susceptibility. Traditional testing can require extensive animal studies, long observation periods and large clinical datasets. Some effects may also be difficult to detect in premarket trials, which are typically designed to assess efficacy and common short-term adverse events rather than rare or delayed reproductive outcomes. Computational toxicology aims to narrow that gap by using chemical structures and biological data to identify patterns associated with toxicity before laboratory and epidemiological investigations begin.

The research team, led by Jiawang Yan, Yaofeng Zhou, Yihuan Zhao and Fushan Tang at Zunyi Medical University in China, first developed and compared several machine-learning approaches for classifying reproductive toxicity. The model that performed best was LightGBM, a gradient-boosting algorithm that builds a prediction from many decision trees. Rather than relying on a single tree, the method combines successive models, with each new tree attempting to correct errors made by earlier ones. Chemical information was represented through molecular descriptors and fingerprints, numerical features that encode aspects of a compound’s structure, such as the presence of functional groups, connectivity patterns and substructures. These features allow the algorithm to compare molecules in a mathematically tractable form.

On an independent test set, LightGBM achieved an F1 score of 0.854, a receiver operating characteristic area under the curve of 0.933 and a precision-recall area under the curve of 0.931. Its Matthews correlation coefficient was 0.705. Each metric captures a different aspect of performance. The F1 score balances precision—how many flagged compounds are truly positive—with recall, or how many positive compounds are successfully detected. ROC-AUC measures how well the classifier separates the two classes across changing decision thresholds, while PR-AUC is often especially informative when positive cases are less common. The Matthews correlation coefficient summarizes all four categories of prediction—true and false positives and negatives—into a score that remains useful when the classes are unevenly represented. Ten-fold cross-validation, in which the data are repeatedly divided into training and validation portions, was used to test whether the model generalized beyond a particular split of the dataset.

The researchers then applied the model to drugs approved during the decade from 2015 through 2024. Seventy-two were classified as having a high predicted risk of reproductive toxicity. To compare these computational findings with post-marketing experience, the team examined FAERS, a database containing reports submitted by health-care professionals, patients, manufacturers and others. Fifty-five of the 72 predicted high-risk drugs—76.39 percent—were associated with reports involving reproductive-toxicity-related adverse events. This overlap was presented as evidence of consistency between the model and pharmacovigilance data, but it is not equivalent to a confirmed causal relationship. FAERS reports are spontaneous and can be affected by underreporting, duplicate submissions, publicity, differences in prescribing rates and the absence of a suitable untreated comparison group. A report can signal a potential safety issue without establishing that a drug caused the event.

The study’s next step was to move from prediction to biological interpretation. Network toxicology treats toxicity as the result of interactions among compounds, proteins, genes and signaling pathways rather than as the consequence of a single molecular target. By integrating drug-associated targets with genes linked to reproductive toxicity and examining their connections in protein-interaction networks, the researchers identified 12 key targets. Among the most prominent were ESR1, IGF1 and AKT1. ESR1 encodes estrogen receptor alpha, a hormone-sensitive transcription factor that helps regulate reproductive development, ovarian and uterine function and other processes. IGF1, or insulin-like growth factor 1, participates in cell growth, metabolism and reproductive biology. AKT1 is a central kinase in the PI3K-AKT pathway, which controls survival, proliferation, metabolism and apoptosis. Disrupting any of these systems could plausibly influence reproductive tissues, although a computational connection alone does not demonstrate that a particular drug produces such disruption in humans.

To test whether the compounds could physically interact with the predicted targets, the authors used molecular docking. Docking algorithms generate possible orientations, or poses, for a small molecule inside a protein’s binding site and estimate the strength of the interaction using scoring functions. These calculations take into account factors such as shape complementarity, electrostatic attraction, hydrogen bonds and hydrophobic contacts. The study found that drugs with high predicted reproductive-toxicity risk could bind effectively to multiple toxicity-associated targets. Such multi-target binding is biologically significant as a hypothesis because a drug can produce unintended effects through proteins unrelated to its intended therapeutic target. At the same time, docking scores are not measurements of binding in a living organism. They depend on the quality of the protein structure, the assumed binding site and the limitations of the scoring model.

The investigators therefore examined selected drug–ESR1 interactions using molecular dynamics simulations, which model how atoms move over time under physical force fields. Unlike a single docking snapshot, a simulation can reveal whether a protein–ligand complex remains stable as both partners flex and fluctuate in a virtual aqueous environment. The reported interactions were maintained largely through hydrogen bonding and hydrophobic contacts, and favorable binding free-energy estimates provided additional support for the possibility of stable association. Binding free energy is a thermodynamic quantity intended to approximate how energetically favorable complex formation may be, although calculated values are sensitive to the simulation method, sampling time and assumptions used to estimate solvent effects. These results strengthen the mechanistic hypothesis but cannot substitute for biochemical binding assays, cellular experiments or reproductive studies in animals and humans.

The authors describe the work as a layered strategy in which each analytical method addresses a different limitation. Machine learning provides a high-throughput way to rank compounds from chemical features. FAERS comparison tests whether those rankings resemble patterns seen in post-marketing safety reports. Network toxicology links candidate drugs to biological pathways and identifies proteins that may connect several reproductive outcomes. Docking and molecular dynamics then explore whether selected molecules could interact directly and stably with those proteins. The study’s data, source code and trained models have been made available through a GitHub repository, enabling other researchers to reproduce the workflow and assess its performance on additional compounds. Reproducibility will be important because predictive models can inherit biases from their training data and may perform differently when applied to drugs with chemical structures or mechanisms not well represented in the original dataset.

The findings arrive as regulators and drug developers seek faster ways to detect safety signals while medicines are still being studied or soon after they reach the market. A model that identifies candidates for targeted testing could help allocate laboratory resources, guide the design of reproductive and developmental studies and prompt more detailed review of prescribing information. It could also support investigations of medicines used by people who may become pregnant or who are planning to conceive. But the 72 flagged drugs should not be interpreted as a list of proven hazards, and the 17 predicted high-risk drugs without corresponding FAERS signals should not automatically be considered safe. Absence of a report may reflect limited use, inadequate follow-up or poor recognition of an event. Ultimately, the study offers a sophisticated prioritization tool rather than a clinical verdict: the strongest signals will need confirmation through controlled toxicology experiments, mechanistic laboratory work and carefully designed human studies that account for dose, exposure timing, underlying illness and co-medications.