Improving Compound Activity Classification via Deep Transfer and Representation Learning.

Area of Science:

• Molecular machine learning
• Cheminformatics
• Computational drug discovery

Background:

• Deep neural networks, particularly graph neural networks (GNNs), show promise for predicting structure-activity relationships (SAR).
• The performance of these models is often constrained by the need for extensive training datasets.
• Transfer learning offers a solution by leveraging data from related tasks to improve performance on data-scarce target tasks.

Purpose of the Study:

• To develop novel deep transfer learning methods, TAc and TAc-fc, for SAR modeling that overcome data limitations.
• To enhance the generalization of molecular features across different domains.
• To improve classification performance in target tasks with limited data.

Main Methods:

• Developed TAc, a deep transfer learning method focused on generating generalizable molecular features.
• Introduced TAc-fc, an extension of TAc incorporating selective feature-wise and compound-wise transferability.
• Utilized bioassay screening data from PubChem, curating 120 bioassay pairs with specific similarity criteria.

Main Results:

• TAc achieved the highest average ROC-AUC of 0.801, significantly improving 83% of target tasks with an average gain of 7.102% over the dmpna baseline.
• TAc demonstrated superior performance across a wide range of target tasks compared to all other baselines.
• TAc-fc, while having a slightly lower average ROC-AUC (0.798), excelled in PR-AUC and F1 scores on several tasks, indicating strong performance.

Conclusions:

• TAc and TAc-fc represent effective deep transfer learning strategies for SAR modeling, particularly in data-limited scenarios.
• These methods significantly enhance predictive performance in computer-aided drug discovery.
• TAc-fc offers competitive or superior performance to TAc on specific metrics and tasks, providing a valuable alternative.