Ens-Chemage: Robust Molecular Image-Based Ensemble Transfer Learning Framework for Small Contaminant Property Data Sets

Contaminant property data sets are typically small, posing challenges for developing accurate deep learning (DL) models. In this study, we pretrained ResNet18 models on the PubChem data set (∼10 million molecules) using molecular RGB images as inputs and their MACCS fingerprints as labels, generating five models (Chemage1 to Chemage5) with various pretraining accuracy, and fine-tuned them on 10 MoleculeNet and 12 contaminant property data sets. We found that appropriate model architectures and fine-tuning techniques significantly improve the transfer learning efficacy. We then developed an ensemble model, Ens-Chemage, to combine the strengths of these individual models. Ens-Chemage outperformed conventional machine learning (ML) models and ImageMol on almost all tested data sets. Through model interpretation, we found that Ens-Chemage learned more accurate and distinct features than the other models. Additionally, we defined its applicability domain (AD) by using an uncertainty-based approach. Finally, Ens-Chemage has been deployed for free public use at https://ens-chemage.streamlit.app/. This study presents significant advancements in the application of DL for small contaminant property data sets.