Detecting Interspecific Positive Selection Using Convolutional Neural Networks

Area of Science:

• Computational Biology
• Evolutionary Genetics
• Bioinformatics

Background:

• Traditional statistical methods (maximum likelihood, Bayesian inference) detect positive selection using phylogenies and codon alignments.
• These methods suffer from false positives due to alignment errors, especially with high indel rates and divergence.
• Existing frameworks struggle to handle noisy sequence data and alignment errors effectively.

Purpose of the Study:

• To develop and evaluate convolutional neural network (CNN) models for detecting positive selection in DNA sequences.
• To improve accuracy and robustness compared to traditional statistical methods, particularly with noisy data.
• To explore the generalizability and scalability of CNNs for large-scale evolutionary analyses.

Main Methods:

• Trained and tested CNN models on simulated codon sequence alignments.
• Compared CNN performance against traditional statistical methods under various phylogenetic scenarios.
• Utilized saliency maps to interpret CNN decision-making and explore site-wise inference.

Main Results:

• CNN models achieved higher accuracy in detecting positive selection, especially on simulated noisy data with misalignments.
• The CNN approach demonstrated robustness against alignment errors where traditional methods faltered.
• Trained CNN models are computationally faster at test time, enabling scalable, large-scale analyses.

Conclusions:

• CNNs provide a powerful and accurate alternative for detecting positive selection in molecular evolution.
• This AI-driven method offers improved handling of data imperfections like misalignments.
• CNNs present a scalable solution for evolutionary genomic studies, with potential for site-specific selection inference.