D-LIM: A neural network for interpretable gene-gene interactions

Area of Science:

• Genomics
• Computational Biology
• Systems Biology

Background:

• Predicting the functional impact of genetic mutations, especially across multiple genes (epistasis), remains a significant challenge in biology.
• Existing methods like biochemical models require extensive parameter knowledge, while standard machine learning approaches often lack biological interpretability.
• Understanding genotype-fitness relationships is crucial for fields ranging from disease research to synthetic biology.

Purpose of the Study:

• To develop a novel machine learning framework, D-LIM (Dimensionality-Limiting Inference of Molecular Phenotypes), for inferring interpretable low-dimensional fitness landscapes from mutation-fitness data.
• To enable accurate prediction of mutation effects by modeling genes as independent molecular phenotypes with nonlinear interactions.
• To provide a tool that not only predicts fitness but also reveals the underlying biological mechanisms and limitations of low-dimensional models.

Main Methods:

• Introduction of D-LIM, a neural network architecture designed to infer fitness landscapes directly from mutation-fitness data.
• D-LIM assumes that genes contribute through independent, gene-specific molecular phenotypes whose interactions determine overall fitness.
• Application of D-LIM to diverse biological datasets, including deep mutational scanning data from metabolic pathways, protein-protein interactions, and yeast adaptation studies.

Main Results:

• D-LIM achieved state-of-the-art predictive accuracy across various biological systems, outperforming existing methods.
• The model successfully inferred interpretable effective phenotypes, revealing the nature of gene interactions and identifying potential trade-offs.
• In cases where the low-dimensional assumption failed, D-LIM effectively indicated the insufficiency of such models, guiding further investigation.
• The model demonstrated the ability to estimate mutational effects on inferred phenotypes, allowing for limited extrapolation beyond the training data.

Conclusions:

• D-LIM offers a powerful and interpretable approach to modeling genotype-fitness relationships, particularly in complex biological systems.
• The framework provides insights into the structure of epistasis and the validity of low-dimensional approximations in biological fitness landscapes.
• By integrating structure constraints within a neural network, D-LIM facilitates biological inference and hypothesis generation, advancing our understanding of gene function and evolution.