Efficient detection and characterization of targets of natural selection using transfer learning

Area of Science:

• Genomics
• Evolutionary Biology
• Bioinformatics

Background:

• Natural selection shapes genomes, creating patterns of altered spatial diversity crucial for understanding evolution.
• Machine learning, particularly Convolutional Neural Networks (CNNs), shows promise in detecting adaptive signatures in population genomic data.
• Deep CNNs offer high performance but require substantial data and computational resources, while shallow CNNs have limited capacity.

Purpose of the Study:

• To develop and evaluate TrIdent, a novel tool leveraging transfer learning to enhance the detection of adaptive genomic regions.
• To improve upon existing methods for identifying genomic regions affected by natural selection using image representations of genetic variation.
• To adapt TrIdent for inferring selection parameters and exploring model interpretability.

Main Methods:

• Developed TrIdent, a tool employing transfer learning with deep CNNs pre-trained on large datasets.
• Represented multilocus variation as images for input into the CNN model.
• Evaluated TrIdent's performance across diverse genetic, demographic, and adaptive scenarios, including unphased data.
• Utilized class activation maps for model interpretability and adapted the model for parameter inference.

Main Results:

• TrIdent demonstrated superior detection of adaptive genomic regions compared to recent methods.
• The tool effectively identified known selection sweep candidates and novel genes associated with cancer and other diseases.
• Model interpretability was explored, and the tool was successfully adapted for selection parameter inference.

Conclusions:

• Transfer learning offers an effective approach to overcome data and computational limitations in detecting adaptive genomic regions.
• TrIdent provides a powerful and versatile tool for evolutionary genomics research, enhancing the identification of selection signatures.
• The application of TrIdent to whole-genome data revealed potential adaptive sweeps in human populations, including disease-associated genes.