End-to-end sequence-structure-function meta-learning predicts genome-wide chemical-protein interactions for dark...

Area of Science:

• Computational biology and cheminformatics
• Genomics and proteomics
• Drug discovery and medicinal chemistry

Background:

• Discovering protein-ligand interactions is crucial for chemical genomics, protein function prediction, and drug discovery.
• Over 90% of gene families lack identified small-molecule ligands, termed "dark" gene families, due to experimental and historical biases.
• Current computational methods struggle with predicting ligands for novel or dissimilar proteins in dark gene families.

Purpose of the Study:

• To develop a deep learning framework, PortalCG, to systematically discover protein-ligand interactions in understudied "dark" gene families.
• To address the limitations of existing computational approaches in predicting ligands for proteins with unknown interactions.
• To enhance drug discovery pipelines by identifying novel targets and screening compounds for under-explored proteomes.

Main Methods:

• Developed PortalCG, a deep learning framework with four novel components: 3D binding site sequence pre-training, end-to-end pretraining-fine-tuning, out-of-cluster meta-learning, and stress model selection.
• Employed a sequence-structure-function paradigm, integrating protein structural information as an intermediate layer within a differentiable deep learning framework.
• Utilized extensive benchmark experiments and external validation, including out-of-distribution (OOD) scenarios and multi-target compound screening.

Main Results:

• PortalCG significantly outperformed state-of-the-art machine learning and protein-ligand docking techniques on dark gene families.
• Demonstrated strong generalization power for target identification and compound screening in OOD scenarios.
• Outperformed rational design by medicinal chemists in external validation for multi-target compound screening, highlighting its potential in drug discovery.

Conclusions:

• PortalCG offers a viable solution to the out-of-distribution problem in exploring understudied protein functional space.
• The differentiable sequence-structure-function deep learning framework shows superiority over conventional methods using predicted structures.
• PortalCG has significant potential in drug discovery, exemplified by applications in designing dopamine receptor antagonists for opioid use disorder and illuminating the human genome for novel therapeutics.