Toward FOXM1 inhibitors based on supervised machine learning classification models and similarity-based screening

Ovarian cancer remains among the deadliest gynaecological malignancies, with aberrant activation of the Forkhead Box M1 (FOXM1) transcription factor strongly implicated in tumour progression, therapy resistance, and poor prognosis. Given the limited availability of validated FOXM1 inhibitors and the lack of prior dedicated machine learning-based QSAR studies identified in our literature survey, this study employed a supervised machine learning framework to predict the bioactivity of structurally diverse small-molecule inhibitors. A curated dataset of 101 compounds with reported IC₅₀ values was obtained from the ChEMBL database, and 1444 molecular descriptors were computed and filtered to retain 43 non-redundant features for modelling. Seven regression models, namely multiple linear regression, Ridge, partial least squares, Random Forest, Elastic Net, k-nearest neighbour, and support vector regression, were developed and assessed using training-set fitting, nested cross-validation, and external test-set validation. Among the evaluated models, Random Forest showed the strongest overall external predictive performance, achieving R²_external = 0.797, RMSE = 0.333, MAE = 0.250, Q²F₁ = 0.801, Q²F₂ = 0.797, Q²F₃ = 0.785, CCC = 0.875, and MRE = 6.00%, while satisfying the Golbraikh-Tropsha Criterion A. The k-nearest neighbour model also showed strong predictive ability, with R²_external = 0.781 and the highest nested cross-validated performance (Q²_CV_Nested = 0.728). SHAP analysis indicated that hydrophobic aliphatic features, electronic distribution, molecular rigidity, and hydrogen-bonding related descriptors contribute meaningfully to FOXM1 inhibition. Structural similarity analysis of the five most active compounds using ECFP4 fingerprints and the Tanimoto coefficient revealed substantial scaffold conservation, with similarity values ranging from approximately 64.71% to 99.42%. Bemis-Murcko scaffold analysis identified 59 unique scaffolds from 101 compounds, indicating moderate scaffold diversity, while PCA and t-SNE showed that active compounds occupy defined regions of chemical space. These findings provide an interpretable ML-QSAR framework for FOXM1 inhibitors and a useful computational basis for analysing and prioritizing compounds within the studied chemical space.