SMILES-Transformer-Assisted SAVI Screening for Novel Chronobiotics

The development of new chronobiotics, substances capable of selectively modulating the parameters of circadian rhythms, is hampered by the fragmented nature and limited volume of available experimental data.In the present study, a comprehensive evaluation of the applicability of the SMILES-Transformer architecture to the classification of circadian rhythm modulators was performed using the specialised ChronobioticsDB resource, and the first systematic virtual screening of the SAVI (Synthetically Accessible Virtual Inventory) library of synthetically accessible compounds for chronobiotic activity was carried out. Rigorous protocols were applied for model training and validation: Data-Efficient Modeling (DEM) assessment with 20 repeats, repeated scaffold validation (5 × 5), and a comparative analysis of training strategies (feature-based vs. end-to-end fine-tuning). The influence of three variants of circadian-effect labelling (raw, aggregated, and expert-curated) and three loss functions (BCE, Focal Loss, and Asymmetric Loss) on the quality of multi-label classification was investigated. The results demonstrate that systematic hyperparameter optimisation in end-to-end mode provides the best predictive performance (ROC-AUC 0.666 for the effect_coarse task), whereas standard fine-tuning without optimisation leads to overfitting (ROC-AUC 0.470). Scaffold validation confirmed the ability of the model to generalise to structurally novel compounds (ROC-AUC 0.587). Expert aggregation of labels improved the recognition of rare classes (F1-macro 0.254 versus 0.148 for the raw labelling). Based on the trained models, a consensus virtual screening of the SAVI library was performed using four independent classifiers (classf, effect_coarse, target, mechanism). From more than five million compounds, 10,000 of the most promising candidates were selected, among which 34 super-candidates (consensus score > 0.9) and 435 strong candidates (> 0.8) were identified. Analysis of the predicted targets revealed dominance of the CLOCK-BMAL1 complex (60.49%), while among effects the circadian phase shift prevailed (37%). All identified candidates are synthetically accessible and are recommended for prioritised experimental verification.