Accessible and robust machine learning approaches to improve the opsin genotype-phenotype map.

Area of Science:

• Genomics
• Molecular Biology
• Bioinformatics

Background:

• Predicting biological phenotypes from genetic variations is a significant challenge.
• Machine learning (ML) shows promise but faces limitations in accessibility, interpretability, and data availability (the 'data-cliff').
• Opsin genes, crucial for animal spectral sensitivity, serve as an excellent model system for genotype-phenotype prediction.

Purpose of the Study:

• To advance ML-driven genotype-phenotype prediction for opsin genes.
• To develop accessible and interpretable tools for predicting spectral sensitivity (λmax) from opsin sequences.
• To enhance the understanding of sequence-function relationships in opsins.

Main Methods:

• Introduction of the Opsin Phenotype Tool for Inference of Color Sensitivity (OPTICS) with SHapley Additive exPlanations (SHAP) and 3D structural mapping.
• Utilizing amino-acid physicochemical properties for sequence encoding, outperforming standard methods and protein language models in certain aspects.
• Development of the Mine-N-Match (MNM) pipeline to link opsin sequences with in-vivo λmax data.

Main Results:

• OPTICS provides user-friendly prediction of λmax with mechanistic insights.
• Physicochemical property encoding enhances predictive performance and biological explainability.
• MNM pipeline expands genotype-phenotype coverage, especially for underrepresented species.

Conclusions:

• The integrated framework improves confidence, accuracy, and interpretability of genotype-phenotype predictions for animal opsins.
• This work facilitates simulating molecular evolution, reconstructing visual history, designing proteins, and generating testable hypotheses.