FunctionaL Assigning Sequence Homing (FLASH) maps phenotype to sequence with deep and machine learning

Area of Science:

• Genomics
• Machine Learning
• Microbiology

Background:

• Genome-wide association studies (GWAS) have limitations in predicting phenotypes from novel genetic variations and integrating structural variants.
• Existing deep and machine learning models struggle with consistent prediction of microbial resistance phenotypes.

Purpose of the Study:

• Introduce FLASH, a statistically-based deep learning framework for direct analysis of raw sequencing reads.
• To achieve accurate and consistent prediction of microbial phenotypes, including resistance and virulence.
• To enable prediction of phenotypes currently impossible with GWAS, such as phage host range.

Main Methods:

• Developed FLASH, a deep learning framework operating directly on raw sequencing reads.
• Applied FLASH to over 35,000 isolates of bacteria, fungi, and viruses.
• Validated FLASH's performance on independent test data, including previously unseen variations.

Main Results:

• FLASH demonstrated uniformly high accuracy across diverse microbial species.
• The framework successfully identified canonical drug targets and novel pan-species virulence predictors.
• FLASH predicted phenotypes beyond GWAS capabilities, including bacterial phage host range.

Conclusions:

• FLASH offers a new, interpretable, and statistically-based deep learning approach for predicting gene function and phenotype.
• The framework is efficient, valuable for complex microbial genomes, and overcomes limitations of experimental validation.
• FLASH advances microbial genomics by enabling accurate prediction from raw sequencing data for various phenotypes.