Content area
Machine learning and, in particular, deep neural network models have significantly transformed regression modeling and property prediction tasks in all scientific domains. To take advantage of this powerful tool, we integrated two state-of-the-art deep learning frameworks, PyTorch and TensorFlow, into the ChemML software package as part of an AutoML pipeline optimized for chemical property prediction. We benchmarked the classical machine learning models alongside our deep neural network implementations paired with a genetic algorithm for hyperparameter optimization with small and large datasets. Our findings reveal that PyTorchRegressorWrapper achieves the highest ranking according to standard regression metrics, significantly outperforming traditional regression models such as SVR, Ridge, and Lasso. The TensorFlowRegressorWrapper currently ranks fourth in the with ongoing efforts to further optimize its architecture. Both neural network models show strong generalizability when tested on larger chemical datasets. These findings illustrate the advantages of integrating a customizable deep neural network model into AutoML framework and how this approach, powered by a genetic algorithm, can be easily extensible to other datasets and regression tasks in computational chemistry.