Understanding structure-guided variant effect predictions using 3D convolutional neural networks.

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Predicting missense variant pathogenicity is crucial for molecular diagnostics but remains challenging.
• Existing tools often struggle to fully integrate diverse data types like evolutionary information and structural features.

Purpose of the Study:

• To introduce DeepRank-Mut, a framework for predicting missense variant pathogenicity.
• To leverage physicochemical features of amino acids in 3D structural environments using deep learning.

Main Methods:

• Extraction of atomic and residue-level features from the variant's 3D structural environment.
• Representation of features in multi-channel 3D voxel grids.
• Training a 3D convolutional neural network (3D-CNN) for pathogenicity prediction.

Main Results:

• DeepRank-Mut achieves performance comparable to existing resources combining sequence and structural data.
• An average accuracy of 0.77 was obtained on independent test datasets using 10-fold cross-validation.
• Evolutionary information and solvent accessibility of neighboring residues significantly influence predictions.

Conclusions:

• DeepRank-Mut offers a robust deep learning approach for protein structure-guided pathogenicity prediction.
• Understanding variant neighborhood contributions and disease mechanisms is key to improving predictive models.
• The study provides insights for adopting deep learning in variant pathogenicity assessment.