Graph neural networks are promising for phenotypic virtual screening on cancer cell lines
These videos have been matched automatically. Contact us if they are not relevant.
These videos have been matched automatically. Contact us if they are not relevant.
View abstract on PubMed
Summary
This summary is machine-generated.Deep learning models, particularly D-MPNN, show promise for phenotypic virtual screening (PVS) in cancer drug design. Rigorous testing on diverse chemical libraries confirms D-MPNN
Area Of Science
- Computational chemistry
- Drug discovery
- Artificial intelligence in medicine
Background
- Artificial intelligence (AI) is revolutionizing early drug design, with phenotypic virtual screening (PVS) offering new ways to predict cancer cell responses.
- Previous studies questioning deep learning's efficacy in PVS were limited by small datasets, inadequate metrics, and unrealistic chemical diversity.
- Real-world virtual screening libraries present significant chemical diversity, posing a challenge for predictive models.
Purpose Of The Study
- To rigorously evaluate machine learning algorithms for phenotypic virtual screening (PVS) using large, diverse datasets.
- To compare the performance of different algorithms under various validation conditions, including dissimilar-molecule splits.
- To identify the most effective AI approach for predicting compound activity in cancer cell lines.
Main Methods
- Prepared 60 datasets with 30,000-50,000 molecules each, testing growth inhibitory activity on NCI-60 cancer cell lines.
- Evaluated five machine learning algorithms for PVS, conducting approximately 14,440 training runs per algorithm.
- Employed both random and dissimilar-molecule split validation types, primarily using hit rate as the performance metric.
Main Results
- All evaluated models faced greater challenges with test molecules dissimilar to the training data.
- The D-MPNN (Deep-learning Molecular Message Passing Neural Network) algorithm consistently outperformed other methods across both validation types.
- Performance was significantly impacted by the chemical dissimilarity between training and testing molecule sets.
Conclusions
- D-MPNN, a graph-based deep neural network, is the most suitable algorithm for developing predictive models in phenotypic virtual screening.
- Robust evaluation with diverse chemical libraries and appropriate metrics is crucial for assessing AI models in drug discovery.
- Future PVS efforts should consider chemical diversity and employ validation strategies that mimic real-world screening challenges.
These videos have been matched automatically. Contact us if they are not relevant.
These videos have been matched automatically. Contact us if they are not relevant.
