Multi-task convolutional neural networks for predicting in vitro clearance endpoints from molecular images

Area of Science:

• Computational chemistry
• Drug discovery
• Machine learning

Background:

• Compound metabolic stability is a key challenge in pharmaceutical research.
• Predictive in silico models can accelerate drug development by reducing iterative design-make-test-analyze cycles.
• Accurate prediction of in vitro clearance endpoints is essential for optimizing drug candidates.

Purpose of the Study:

• To investigate the efficacy of image-based molecular representations for predicting multiple in vitro clearance endpoints.
• To develop and evaluate multi-task convolutional neural network (CNN) models for clearance prediction.
• To benchmark CNN performance against other state-of-the-art machine learning methods.

Main Methods:

• Curated compound measurements for four in vitro clearance endpoints from internal sources.
• Built multi-task convolutional neural network (CNN) models utilizing image-based molecular representations.
• Employed rigorous data splitting strategies to validate model performance.
• Benchmarked CNNs against deep neural networks (DNNs) and graph convolutional neural networks (GCNs).

Main Results:

• CNN models successfully captured implicit chemical relationships within the data.
• Multi-task learning with CNNs demonstrated on par or improved accuracy compared to state-of-the-art methods.
• Image-based molecular representations proved effective for predicting multiple clearance endpoints.
• CNNs showed a clear benefit from multi-task learning across all investigated clearance endpoints.

Conclusions:

• Image-based molecular representations are a viable approach for predicting multiple compound clearance endpoints.
• Multi-task CNNs offer a powerful tool for accelerating drug discovery by improving predictive accuracy.
• Further research into model interpretability using molecular images is warranted.