Molecular active learning approaches for predicting skin cytotoxicity

Skin cytotoxicity assessment of small molecules is vital for assessing their risk of cytotoxicity upon topical contact. However, identifying cytotoxic compounds within vast chemical spaces is costly and time-consuming. In this study, we developed a data-efficient training approach to reduce the cost and time of cytotoxicity screening using an active deep learning framework. To mimic a real-world drug screening scenario, an experimental budget was imposed, allowing only a limited number of molecules to be selected from a large, unlabeled pool for cytotoxicity evaluation. A curated cytotoxicity dataset against keratinocytes was constructed. The active learning (AL) experiments were conducted using three deep learning (DL) models with seven different acquisition function strategies. From these experimental setups, we found that the self-supervised autoencoder acquired cytotoxic enrichments better than random chance and the other two baseline models. Secondly, we found that the top-k highest-probability acquisition function identified more new cytotoxic molecules than the other acquisition strategies. However, when measuring the diversity of acquired cytotoxic and non-cytotoxic molecules, we found that the novelty and diversity demonstrated the highest structural differences compared to other acquisition functions. Thirdly, we found that the overall cytotoxic enrichments from the best AL model (77.0 ± 4.0) are greater than those from baseline one-shot learning (71.3 ± 4.7), indicating that this AL method is useful for identifying more cytotoxic molecules. Next, we found that the AL model attains performance on par with a one-shot training approach that uses 100% of the pool data, while the AL model trained on only 16.6% of the pool data demonstrates a low data requirement for model training. Furthermore, we validated the predictive performance of the new AL model on cytotoxicity testing, exhibiting that the method is effective for prioritizing cytotoxic molecules in low-resource settings for drug design and development.