Studying genetic variations and disease outcomes

• Deep learning is well-suited for analyzing the intricate relationships between genetic variations and disease outcomes,

  • ultimately leading to accurate predictions of variant pathogenicity

  • E.g., AlphaMissense from Google DeepMind

AlphaMissense

• AlphaMissense is developed to predict the pathogenicity of missense variants

  • i.e., Whether variants that change the protein sequences are likely to cause diseases (pathogenic) or not (bening).

• AlphaMissense is based on AlphaFold model

  • AlphaFold - Accuratley predicts 3D structure of proteins from their amino acid sequences using advanced deep learning techniques and biological knowledge-bases

AlphaMissense deep learning workflow

• AlphaMissense deep learning workflow can be described in 4 main steps

graph TD; 1(Data preparation) --> 2(Feature engineering) 2 --> 3(Model training) 3 --> 4(Model evaluation) classDef highlight fill:#99ccff; classDef white fill:#ffffff; class 1,2,3,4 highlight;

Note

Genomics context of each step :mega:

ML Concept	Genomics conext
Data preparation	Collecting and processing a large dataset of missense variants along with annotations indicating their pathogenicity (disease-causing or benign)
Feature engineering	Convert amino acid sequences into representations suitable for deep learning models
Model training	Fine-tune AlphaFold model that predicts protein structure to predict variant pathogenicity
Model evaluation	Assess the accuracy and generalizability of variant pathogenicity prediction using independent datasets