Deep Learning vs Classical Methods in Potency and ADME Prediction: Insights from a Computational Blind Challenge.

Area of Science:

• Computational chemistry
• Drug discovery
• Artificial intelligence in medicine

Background:

• Predicting compound potency and ADME profiles is essential for successful drug discovery.
• The efficacy of AI and deep learning against classical methods for these predictions is not well-established.
• The ASAP-Polaris-OpenADMET Antiviral Challenge offered a benchmark for evaluating predictive models.

Purpose of the Study:

• To benchmark AI and deep learning against classical methods for predicting compound potency and ADME.
• To analyze the performance of top-performing models in a large-scale computational challenge.
• To identify key factors for building robust predictive models in drug discovery.

Main Methods:

• Retrospective analysis of modeling strategies from the ASAP-Polaris-OpenADMET Antiviral Challenge.
• Rigorous statistical benchmarking of classical, traditional machine learning, and deep learning algorithms.
• Leveraging public datasets and feature augmentation for enhanced model performance.

Main Results:

• Deep learning models significantly outperformed traditional machine learning for ADME prediction.
• Classical methods remained highly competitive for predicting compound potency.
• Top performance achieved in pIC50 prediction for SARS-CoV-2 Mpro and aggregated ADME.

Conclusions:

• AI and deep learning offer significant advantages for ADME prediction, complementing classical methods for potency.
• Data curation and feature augmentation are crucial for developing effective predictive models.
• Future opportunities include integrating structure-guided modeling for enhanced computational drug discovery.