A Multitask Approach to Learn Molecular Properties

Abstract

The endeavors to pursue a robust multitask model to resolve intertask correlations have lasted for many years. A multitask deep neural network, as the most widely used multitask framework, however, experiences several issues such as inconsistent performance improvement over the independent model benchmark. The research aims to introduce an alternative framework by using the problem transformation methods. We build our multitask models essentially based on the stacking of a base regressor and classifier, where the multitarget predictions are realized from an additional training stage on the expanded molecular feature space. The model architecture is implemented on the QM9, Alchemy, and Tox21 datasets, by using a variety of baseline machine learning techniques. The resultant multitask performance shows 1 to 10% enhancement of forecasting precision, with the task prediction accuracy being consistently improved over the independent single-target models. The proposed method demonstrates a notable superiority in tackling the intertarget dependence and, moreover, a great potential to simulate a wide range of molecular properties under the transformation framework.

Artificial Intelligence in Drug Toxicity Prediction: Recent Advances, Challenges, and Future Perspectives

“…Supporting Information Table S1 summarizes AI-based studies on predicting toxicity end points using the Tox21 data set, focusing on those published from 2019 to the present. While some approaches used ML methods, such as RF, GBDT, and SVM, , the majority used DL methods, such as NN, DNN, RNN, CNN, GNN, and GCNN. ,,,,− Most models achieved highly accurate predictions, with an ACC of up to 99.5% and ROC-AUC of up to 99.1%.…”

Machine learned calibrations to high-throughput molecular excited state calculations

“…There is thus a trade-off in using multitasking as it could reduce accuracy for T 1 predictions but improve accuracy for S 1 , while also reducing overall computation time. Because of the time savings of the multi-task model and previous works showing the benefits of multi-property prediction, [62][63][64] this was used for ML for the remainder of this study. Hyperparameter optimization was not performed for the following models due to the only marginal improvement seen.…”

A deep learning framework for predictions of excited state properties of light emissive molecules

“…Specifically for each model, DNN has the best prediction accuracies on the three singlet oscillator strengths. GCN has moderate performance almost on all predicting targets, indicating a possible insufficiency of the molecular 2-dimensional graph in predicting the excited state properties (note that the result is in significant contrast with the ground state property predictions where GCN usually gives the best precision 15 ). RNN performs good enough for most of the properties although the model only takes the SMILES string as input, which does not involve any structural characteristics.…”