Prediction of mutation effects using a deep temporal convolutional network.

Area of Science:

• Computational biology
• Genomics
• Machine learning

Background:

• Predicting genetic variation effects is crucial for biological research.
• Machine learning models, particularly variational autoencoders (VAEs), leverage evolutionary sequence data.
• Current models face challenges in efficiently utilizing diverse sequence data.

Purpose of the Study:

• To develop a novel deep autoregressive generative model for predicting genetic variation effects.
• To improve the modeling of inter-residue correlations in biological sequences.
• To enhance the efficiency and accuracy of predictive models, especially with limited sequence data.

Main Methods:

• Proposed mutationTCN, a deep autoregressive model utilizing dilated causal convolutions and an attention mechanism.
• Modeled inter-residue correlations within biological sequences.
• Extended the architecture to a semi-supervised learning framework.

Main Results:

• mutationTCN demonstrated competitive performance against VAE models in 42 high-throughput mutation scan experiments (mean Spearman rank correlation improvement of ~0.023).
• The model efficiently captured information from multiple sequence alignments with fewer sequences, outperforming latent variable models on viral families.
• The semi-supervised framework achieved high prediction accuracy with a simple and stable training process.

Conclusions:

• mutationTCN offers a powerful and efficient alternative for predicting genetic variation effects.
• The model's ability to handle limited sequence data is particularly valuable for applications like viral family analysis.
• The proposed architecture facilitates direct data likelihood optimization and stable training.