Understanding Colour Tuning Rules and Predicting Absorption Wavelengths of Microbial Rhodopsins by Data-Driven Machine-Learning Approach

Masayuki Karasuyama, Keiichi Inoue, Ryoko Nakamura, Hideki Kandori, Ichiro TakeuchiScientific 2018 · DOI

TL;DR

This paper introduces a machine learning approach to predict the absorption wavelengths of microbial rhodopsins based on their amino acid sequences. By building a comprehensive database and applying a group-wise sparse learning model, the researchers not only accurately predict spectral properties but also uncover novel insights into the "color tuning rules" governing these proteins, identifying previously overlooked residues crucial for spectral shifts.

What Problem Does This Paper Address?

Microbial rhodopsins are versatile proteins used in optogenetics for controlling neural activity. A key challenge in their application is the limited range of available absorption wavelengths, particularly for longer wavelengths (red-shifted) which are desirable due to lower phototoxicity and better tissue penetration. While experimental methods exist to discover and engineer new rhodopsins with desired spectral properties, these are often time-consuming, costly, and rely heavily on intuition or extensive trial-and-error. Existing computational methods for predicting spectral shifts are computationally expensive. This paper aims to bridge this gap by developing a data-driven machine learning approach that can efficiently predict absorption wavelengths and reveal the underlying rules of color tuning, thereby accelerating the design of novel rhodopsins for optogenetics.

Methods

The core of this work is a machine learning (ML) approach based on group-wise sparse learning. The process involves:

  1. Data Representation: Amino acid sequences are converted into a binary representation. Each of the 20 amino acid types at each of the 210 transmembrane residues is represented by a binary variable. This results in a high-dimensional feature space (4,200 variables).
  2. Linear Model: A linear model is employed to predict absorption wavelength (λmax) from the binary sequence representation. This model has an intercept and coefficients for each amino acid-residue combination.
  3. Group-wise Sparsity: To handle the high dimensionality and ensure interpretability, a group-wise sparsity constraint is applied. This regularization technique forces coefficients for all amino acid types at a specific residue to be zero if that residue is deemed unimportant for color tuning. This effectively identifies "active residues."
  4. Interpretability: The fitted coefficients (βi,j) provide insights into color tuning. Positive coefficients suggest a red-shifting effect, while negative coefficients indicate a blue-shifting effect for a particular amino acid at a specific residue.
  5. Prediction and Rule Discovery: The trained ML model is used to predict λmax for new, uninvestigated rhodopsins and to understand the rules governing spectral shifts by analyzing the active residues and their associated coefficients.

Datasets

The study utilizes a comprehensive database constructed from:

  • Previously reported data: 519 microbial rhodopsin sequences and their absorption wavelengths from existing literature.
  • Newly determined data: 277 rhodopsin sequences and their absorption wavelengths experimentally determined by the authors.

This combined dataset comprises 796 proteins in total. For demonstration purposes, the dataset was split into a training set (677 proteins, excluding KR2 and its variants) and a target set (119 KR2 wild-type and its variants). This split allows for evaluating the model's predictive power on unseen data.

Key Insights and Ideas

  • ML as a powerful tool for protein engineering: The study demonstrates that ML can effectively decipher complex relationships between protein sequence and function (spectral properties), moving beyond traditional intuition-based design.
  • Uncovering "hidden" color tuning rules: The group-wise sparsity approach successfully identified residues that significantly influence spectral shifts, including two residues (BR Glu161 and Ala126) that were not previously considered important in prior studies.
  • Quantitative prediction of spectral shifts: The ML model provides a quantitative way to estimate the impact of specific amino acid mutations on absorption wavelengths.
  • Bridging experimental and computational approaches: The work shows how to leverage existing experimental data to build predictive models that can guide future experimental design, reducing trial-and-error.

Key Results

  • Accurate Prediction: The ML model achieved an average prediction error of ±7.8 nm for the absorption wavelengths of KR2 wild-type and its variants.
  • Identification of Active Residues: The model identified 20 "active residues" that significantly influence the color tuning of microbial rhodopsins.
  • Novel Residue Identification: Two residues, BR Glu161 and Ala126, were newly identified as important for color shift, with no prior studies focusing on their specific roles.
  • Correlation: A high correlation coefficient of p=0.7736 was observed between predicted and experimentally observed wavelengths for the target set.
  • Experimental Validation: The ML model's predictions were further validated by experimentally measuring the absorption wavelengths of newly constructed KR2 mutants, showing an average error of 8.04 nm, closely matching the model's prediction.

Limitations and Future Work

  • Additive Effect Assumption: The current ML model assumes an additive effect of amino acid mutations on wavelength. However, real-world interactions can be more complex and combinatorial.
  • Database Size: While comprehensive, the database could be further expanded with more experimental data to improve model complexity and accuracy.
  • Protein Stability and Functionality: The current model focuses solely on spectral properties. Future work could incorporate other factors like protein stability and functional activity into the ML framework for more holistic protein design.
  • Combinatorial Effects: The authors suggest that future models could incorporate combinatorial effects of multiple amino acids to capture more nuanced interactions.
  • Broader Applicability: The ML approach is presented as a general framework that can be applied to predict other molecular properties of functional proteins beyond rhodopsins.

Summary updated

Abstract

The light-dependent ion-transport function of microbial rhodopsin has been widely used in optogenetics for optical control of neural activity. In order to increase the variety of rhodopsin proteins having a wide range of absorption wavelengths, the light absorption properties of various wild-type rhodopsins and their artificially mutated variants were investigated in the literature. Here, we demonstrate that a machine-learning-based (ML-based) data-driven approach is useful for understanding and predicting the light-absorption properties of microbial rhodopsin proteins. We constructed a database of 796 proteins consisting of microbial rhodopsin wildtypes and their variants. We then proposed an ML method that produces a statistical model describing the relationship between amino-acid sequences and absorption wavelengths and demonstrated that the fitted statistical model is useful for understanding colour tuning rules and predicting absorption wavelengths. By applying the ML method to the database, two residues that were not considered in previous studies are newly identified to be important to colour shift.
BibTeX
@inproceedings{karasuyama2018understanding,
  author = {Masayuki Karasuyama and Keiichi Inoue and Ryoko Nakamura and Hideki Kandori and Ichiro Takeuchi},
  title = {{Understanding Colour Tuning Rules and Predicting Absorption Wavelengths of Microbial Rhodopsins by Data-Driven Machine-Learning Approach}},
  booktitle = {Scientific},
  year = {2018},
  doi = {10.1038/s41598-018-33984-w},
}
Figure