Exploration of natural red-shifted rhodopsins using a machine learning-based Bayesian experimental design

Keiichi Inoue, Masayuki Karasuyama, Ryoko Nakamura, Masae Konno, Daichi Yamada, Kentaro Mannen, Takashi Nagata, I. Yu, Kei Yura, Oded Béjà, Hideki Kandori, Ichiro TakeuchibioRxiv 2020 · DOI

TL;DR

Researchers have developed a machine learning model that can predict the absorption wavelengths of microbial rhodopsins, enabling the efficient discovery of new red-shifted variants. This approach significantly speeds up the search for optogenetic tools that can operate with longer, less scattering wavelengths, potentially leading to more precise neural control.

What Problem Does This Paper Address?

Optogenetics, a powerful technique for controlling neural activity with light, relies on light-sensitive proteins called rhodopsins. However, current optogenetic tools often require strong light, which can scatter in biological tissues, limiting their effectiveness and depth of penetration. To overcome this, there's a growing need for rhodopsins that absorb longer, red-shifted wavelengths of light. Discovering these red-shifted rhodopsins through traditional methods like gene screening and mutation analysis is time-consuming and expensive. This paper addresses the challenge of efficiently identifying novel red-shifted rhodopsins from vast genomic databases.

Methods

The core of this work is a machine learning (ML) model, specifically a Bayesian sparse modeling approach called BLASSO. This model was trained to predict the absorption wavelength ($\lambda_{max}$) of microbial rhodopsins based on their amino acid sequences. Key aspects of their approach include:

  • Feature Engineering: Instead of using raw amino acid sequences, they transformed them into a numerical feature vector. This involved considering 24 specific amino acid residues known to be important for color tuning around the retinal chromophore, and for each residue, using 18 different physicochemical features (e.g., weight, hydropathy, isoelectric point). This resulted in a 432-dimensional feature space per protein.
  • Bayesian Modeling: Unlike standard regression, their Bayesian approach provides not just a point prediction but also a measure of uncertainty (variance) for each prediction. This is crucial for balancing exploration (searching for novel candidates with high uncertainty) and exploitation (selecting candidates with high predicted red-shift gains).
  • Expected Red-Shift Gain: They defined a screening criterion called "expected red-shift gain" which considers both the predicted wavelength and the uncertainty, guiding the selection of promising candidates.

Datasets

  • Training Data: The ML model was trained on a dataset of 893 microbial rhodopsins with known absorption wavelengths, augmented with 97 newly reported rhodopsins.
  • Target Data: For screening, they used a dataset of 3,064 new candidate rhodopsin genes from public databases (NCBI, Tara Oceans).

Key Insights and Ideas

  • ML as a powerful screening tool: The study demonstrates that ML can dramatically accelerate the discovery of functional proteins by filtering vast genomic datasets down to a manageable number of promising candidates.
  • Exploration-exploitation trade-off is critical: Simply predicting the highest wavelength isn't enough. The model's ability to quantify prediction uncertainty allows for a more strategic search, balancing the risk of selecting a poorly predicted candidate with the potential for discovering something truly novel.
  • Specific amino acid positions are key: The research reinforces the importance of specific amino acid residues around the retinal chromophore in determining rhodopsin color. Their model leverages this by focusing on these critical positions.
  • Red-shifted rhodopsins are abundant: The success in finding 32 out of 40 experimentally tested candidates with red-shifted gains suggests that nature has evolved a diverse range of these useful proteins.

Key Results

  • The ML model predicted the absorption wavelengths of microbial rhodopsins with an average error of ±7.8 nm.
  • From 5,558 candidate rhodopsins, their ML model selected 40 promising candidates.
  • Experimental validation showed that 32 out of these 40 candidates (80%) exhibited red-shifted absorption wavelengths.
  • Four of these showed significant red-shift gains (> 20 nm).
  • Two of these also possessed desirable ion-transporting properties, making them potential optogenetic tools.
  • The ML-based screening was significantly more efficient than random sampling (p < 0.0002).

Limitations and Future Work

  • Modeling Errors: The authors acknowledge that observed wavelengths and gains were generally smaller than predicted. They attribute this to modeling errors, possibly due to insufficient information (like 3D structures) or limitations in the model's flexibility (e.g., not fully capturing non-linear relationships).
  • Selection Bias: The ML model might inadvertently select candidates with high prediction values due to modeling errors, leading to a form of selection bias. Developing statistical methods to mitigate this bias is suggested.
  • Broader Application: While this study focused on ion-pumping rhodopsins, the authors suggest their ML model could be extended to ion channel and enzymatic rhodopsins, even those from eukaryotic origins, for future optogenetics applications.
  • Mechanism Investigation: The identified red-shifted rhodopsins provide a basis for further investigation into the specific amino acid residue changes responsible for color tuning.

Summary updated

Abstract

Abstract Microbial rhodopsins are photoreceptive membrane proteins utilized as molecular tools in optogenetics. In this paper, a machine learning (ML)-based model was constructed to approximate the relationship between amino acid sequences and absorption wavelengths using ~800 rhodopsins with known absorption wavelengths. This ML-based model was specifically designed for screening rhodopsins that are red-shifted from representative rhodopsins in the same subfamily. Among 5,558 candidate rhodopsins suggested by a protein BLAST search of several protein databases, 40 were selected by the ML-based model. The wavelengths of these 40 selected candidates were experimentally investigated, and 32 (80%) showed red-shift gains. In addition, four showed red-shift gains &gt; 20 nm, and two were found to have desirable ion-transporting properties, indicating that they were potentially useful in optogenetics. These findings suggest that an ML-based model can reduce the cost for exploring new functional proteins.
BibTeX
@misc{inoue2020exploration,
  author = {Keiichi Inoue and Masayuki Karasuyama and Ryoko Nakamura and Masae Konno and Daichi Yamada and Kentaro Mannen and Takashi Nagata and I. Yu and Kei Yura and Oded Béjà and Hideki Kandori and Ichiro Takeuchi},
  title = {{Exploration of natural red-shifted rhodopsins using a machine learning-based Bayesian experimental design}},
  year = {2020},
  doi = {10.1101/2020.04.21.052548},
  howpublished = {bioRxiv},
}
Figure