Benchmarking Molecular Fingerprints and Vector Embeddings for Predicting Indoleamine and Tryptophan-2,3-Dioxygenase Inhibition Selectivity Publisher



Naijian R ; Valipour M ; Amini M ; Irannejad H
Source: Letters in Drug Design and Discovery Published:2026

Abstract

Background Indoleamine and Tryptophan 2,3-dioxygenase (IDO/TDO) are crucial immunosuppressive enzymes in cancer immunotherapy. This study aimed to systematically benchmark the predictive utility of diverse molecular descriptors, encompassing classical fingerprints and modern deep learning embeddings, for selectivity prediction of IDO/TDO inhibitors. Methods A dataset comprising 760 compounds was utilized for multi-class classification. The investigation evaluated a comprehensive array of molecular representations, including Morgan fingerprints (both bit- and count-based), MACCS keys, Mol2vec, ChemDist, and ChemBERTa embeddings. Classification models included k-Nearest Neighbors (kNN) and Support Vector Machine (SVM), tested both with and without internal and external class balancing techniques such as SMOTE and ADASYN. Model performance was assessed using both random and scaffold-based splitting protocols. Results SVM demonstrated superior classification scores compared to kNN in random data splitting. Internal class weighting outperformed external resampling techniques (SMOTE/ADASYN) regarding minority class F1-scores. The best model, SVM trained on the 2048 count-based Morgan fingerprints (AUC-ROC = 0.88, MCC = 0.65), achieved superior performance on the internal test set. Interestingly, kNN models were shown to be more predictive than SVM applied on the Bemis-Murcko scaffold split data. Accordingly, kNN model applied on 300-dimensional Mol2vec embeddings showed competitive results (AUC-ROC = 0.77, MCC = 0.51) compared to the kNN applied on the 1024-bit Morgan fingerprints (AUC-ROC = 0.77, MCC = 0.53). Conclusion Established circular fingerprints remain an effective and reliable representation for selectivity prediction of IDO/TDO inhibitors compared to modern neural network embeddings. The obtained results are context-dependent and cannot be generalized to other classes of molecules. © 2026 The Authors.