Ens-Chemage:\nRobust Molecular Image-Based Ensemble\nTransfer Learning Framework for Small Contaminant Property Data Sets

Contaminant property data sets are typically small, posing\nchallenges\nfor developing accurate deep learning (DL) models. In this study,\nwe pretrained ResNet18 models on the PubChem data set (∼10\nmillion molecules) using molecular RGB images as inputs and their\nMACCS fingerprints as labels, generating five models (Chemage1 to\nChemage5) with various pretraining accuracy, and fine-tuned them on\n10 MoleculeNet and 12 contaminant property data sets. We found that\nappropriate model architectures and fine-tuning techniques significantly\nimprove the transfer learning efficacy. We then developed an ensemble\nmodel, Ens-Chemage, to combine the strengths of these individual models.\nEns-Chemage outperformed conventional machine learning (ML) models\nand ImageMol on almost all tested data sets. Through model interpretation,\nwe found that Ens-Chemage learned more accurate and distinct features\nthan the other models. Additionally, we defined its applicability\ndomain (AD) by using an uncertainty-based approach. Finally, Ens-Chemage\nhas been deployed for free public use at https://ens-chemage.streamlit.app/. This study presents significant advancements in the application\nof DL for small contaminant property data sets.