Abstract
【Objective】
QSAR prediction helps avoid costs and delays associated in Ames tests for new compounds. We developed DeepSnap-DL, a QSAR technique using 3D compound images for deep learning. This study sought to enhance prediction accuracy by integrating DeepSnap-DL with traditional descriptor-based models using ensemble and consensus methods for Ames models.
【Methods】
This study used Ames test results sourced from the mutagen by MHLW and the Hansen data set. For the descriptor method, we developed a lightGBM-based model using molecular descriptors from MOE. The ensemble method built models similar to the approach of the descriptor method, using the prediction probability of DeepSnap-DL and molecular descriptors. The consensus method involved building a model with an applicability domain for compounds in which DeepSnap-DL and descriptor predictions matched. We assessed generalization performance with a 20% external test set, utilizing metrics like balanced accuracy and MCC.
【Results and Discussion】
An Ames test data set comprising 6950 compounds (positive/negative ratio: 1.31/1) was assembled. Compared to the descriptor method, no improvement was observed in prediction performance between the DeepSnap-DL and ensemble methods. Yet, the consensus method showcased superior generalization performance among all prediction models. These results suggest that the prediction model’s applicability domain, established by the consensus method (i.e., the group of compounds for which the descriptor and DeepSnap-DL methods predictions coincide), is useful for QSAR prediction of the Ames test.
