mutationTCN

mutationTCN predicts the effects of genetic variations by modeling inter-residue dependencies in biological sequences using a deep autoregressive generative model with dilated causal convolutions and an attention mechanism.

Key Features:

• Deep Autoregressive Generative Model: mutationTCN combines dilated causal convolutions with an attention mechanism to capture inter-residue correlations within biological sequences.
• Competitive Performance: On 42 high-throughput mutation scan experiments, mutationTCN achieved a mean improvement in Spearman rank correlation of approximately 0.023 compared to a variational autoencoder (VAE).
• Efficiency with Limited Data: The model leverages information from multiple sequence alignments and performs effectively with lower effective numbers of sequences, applicable to viral sequence families.
• Semi-Supervised Learning Framework: The architecture supports semi-supervised learning to incorporate both labeled and unlabeled data during training.
• Direct Optimization and Stable Training: mutationTCN enables direct optimization of data likelihood, facilitating a simple and stable training process.

Scientific Applications:

• Genetic Research: Predicting mutation impacts to aid studies of genetic disease mechanisms and variant interpretation.
• Viral Genomics: Analyzing viral sequence families and studying viral evolution when sequence data are limited.
• Protein Engineering: Predicting how mutations affect protein function to inform design of proteins with desired properties.

Methodology:

Computational methods include a deep autoregressive generative model using dilated causal convolutions and an attention mechanism, utilization of multiple sequence alignments, a semi-supervised learning extension, direct optimization of data likelihood, and evaluation versus a variational autoencoder using Spearman rank correlation across 42 high-throughput mutation scan experiments.