Abstract
A carcinogenicity assessment of possibly carcinogenic chemicals (International Agency for Research on Cancer: IARC class 2B) was conducted using a consensus framework constructed from three complementary machine learning models: BiLSTM with MACCS fingerprints, LightGBM with RDKit descriptors, and Random Forest (RF) with E-state features. These models were developed and rigorously evaluated on benchmark carcinogenicity data sets, with LightGBM emerging as the top performer (accuracy = 0.800, MCC = 0.615, AUROC = 0.882, sensitivity = 0.739, specificity = 0.857). Consistent and competitive performance was also observed for RF and BiLSTM, affirming the reliability of individual predictions. Notably, LightGBM maintained strong generalization ability on independent human carcinogen test sets from IARC and IRIS (accuracy = 0.753, MCC = 0.535, AUROC = 0.842). For the ISSCAN internal test set, the top three models achieved MCC values ranging from 0.564 to 0.615, with AUROC scores between 0.858 and 0.882. For the human carcinogen test set, the top three models attained MCC values from 0.335 to 0.535 and AUROC scores ranging from 0.785 to 0.842. The consensus model was subsequently applied to 47 within-domain compounds from the 2B category, classifying them into 16 potential carcinogens, 8 presumed noncarcinogens, and 23 cases with inconclusive results. To uncover structural correlates, a SHAP-based interpretation of the BiLSTM model was performed, revealing discriminative molecular features including MACCS fingerprint keys and core Bemis-Murcko scaffolds associated with predicted carcinogenicity. To support practical applications, a freely accessible web server for carcinogenicity assessment has been developed and is available at https://carcinogenicity-predictor.streamlit.app.