Functional characterization of red-shifted rhodopsin channels from giant viruses explored by a machine-learning model for long-wavelength optogenetics

Shunki Takaramoto, Chunlan Zhao, Masayuki Karasuyama, Frederik Schulz, Masaya Watanabe, Yuma Kawasaki, Takashi Nagata, Masae Konno, Naoya Morimoto, Masahiro Fukuda, I. Yu, Hiromu Yawo, Oded Béjà, Hideaki Kato, Tanja Woyke, Ichiro Takeuchi, Keiichi InouebioRxiv 2025 · DOI

TL;DR

This paper tackles the challenge of finding better optogenetic tools by using machine learning to predict the light absorption properties of channelrhodopsins (ChRs). They developed an ML model that successfully identified ChR024, a viral rhodopsin with a red-shifted absorption spectrum, paving the way for more efficient and less toxic optogenetic manipulations.

What Problem Does This Paper Address?

Optogenetics, a powerful technique for controlling neural activity with light, relies on light-gated ion channels like channelrhodopsins (ChRs). However, the ChRs commonly used absorb light in the blue-green spectrum. This short-wavelength light is problematic because it causes cell toxicity and doesn't penetrate deeply into tissues, limiting the scope and effectiveness of optogenetic interventions. While red-shifted ChRs exist, there's a continuous need for even longer-wavelength variants for deeper tissue penetration and reduced phototoxicity. The traditional method of finding these improved ChRs involves extensive and costly experimental screening. This paper aims to overcome this bottleneck by leveraging machine learning to predict the absorption characteristics of uncharacterized ChRs, thereby accelerating the discovery of novel, red-shifted optogenetic tools.

Methods

The core of this study is the development and application of a machine-learning (ML) model. Specifically, the authors employed an elastic-net regression model. This type of model is chosen for its ability to perform feature selection and regularization, which is useful when dealing with potentially high-dimensional biological data. The model was trained to predict the maximum absorption wavelength (λmax) of ChRs based on the amino acid sequences surrounding their retinal chromophore. The researchers focused on 24 specific residues that were previously identified as critical for determining λmax. They extracted 18 physico-chemical features for each of these residues, resulting in a 432-dimensional feature vector for each ChR. After training, this model was used to predict the λmax of novel ChR homologs. Promising candidates were then synthesized and experimentally validated.

Datasets

The ML model was trained on a dataset of 1,163 experimental data points. These data were collected from existing literature and from datasets generated by some of the authors. The model was then used to predict the λmax of 259 new ChR homologs identified through homology searches of public gene databases. Out of these, 23 genes predicted to have the most red-shifted absorption were synthesized and experimentally characterized.

Key Insights and Ideas

  • ML as a powerful screening tool: The study demonstrates that machine learning can significantly reduce the cost and effort associated with discovering functional proteins by pre-screening candidates based on predicted properties.
  • Viral rhodopsins as a promising source: The identification of ChR024, a viral rhodopsin, as a highly red-shifted cation channel highlights the potential of exploring diverse biological sources for novel optogenetic tools.
  • Sequence-function relationship: The focus on specific amino acid residues surrounding the retinal chromophore underscores the detailed understanding of how protein sequence dictates light absorption properties.
  • ChR024's unique properties: ChR024 exhibits a red-shifted absorption spectrum and a distinct ion transport mechanism compared to canonical ChRs, suggesting it could offer unique advantages for specific optogenetic applications.

Key Results

The ML model, trained on 1,163 data points, successfully predicted the λmax of 259 uncharacterized ChR homologs. From these predictions, 23 candidates with red-shifted absorption were selected for experimental validation. Among these, ChR024 was identified as a cation-conducting channel with a λmax of approximately 578 nm, making it the second-longest absorbing ChR known, closely following Chrimson (~580 nm). The study also characterized ChR024's electrophysiological properties, confirming its cation transport capabilities and revealing a distinct ion transport mechanism compared to other ChRs. Spectroscopic studies further elucidated its pH-dependent absorption and photocycle.

Limitations and Future Work

The authors acknowledge that while their ML model effectively predicts λmax, it does not predict protein expression levels. This limitation means that some ChRs predicted to have desirable absorption properties might not be efficiently expressed in cells, hindering high-throughput screening. Future work could involve acquiring experimental datasets that include both λmax and expression levels to improve the predictive power of ML models. Additionally, the authors suggest combining experimental design based on active learning with robotic molecular characterization to further accelerate the discovery and development of functional proteins. The unique properties of ChR024 also open avenues for exploring its specific applications in optogenetics.

Summary updated

Abstract

Abstract Channelrhodopsins (ChRs) are light-gated ion channels. These proteins are widely used in optogenetics to optically manipulate neural activity. However, manipulation using short-wavelength light to activate ChRs causes cell toxicity and is hampered by low tissue penetration. To overcome these difficulties, although several red-shifted ChR variants have been identified, further red-shift is required for more efficient and noninvasive neural control. While molecular screening of ChRs requires high-cost experiments, recent machine-learning-based protein functionality prediction enables more efficient selection of target proteins for characterization. Here, we constructed an elastic-net machine-learning model trained on 1,163 experimental data to predict the maximum absorption wavelength ( λ max ) of uncharacterized ChRs. The model suggested several red-shifted candidates, and we identified a viral rhodopsin channel, ChR024, with the second-longest λ max as a cation-conducting ChR ( λ max = ∼578 nm) after Chrimson ( λ max = ∼580 nm). This result demonstrates the high impact of ML on reducing the screening costs of functional proteins.
BibTeX
@misc{takaramoto2025functional,
  author = {Shunki Takaramoto and Chunlan Zhao and Masayuki Karasuyama and Frederik Schulz and Masaya Watanabe and Yuma Kawasaki and Takashi Nagata and Masae Konno and Naoya Morimoto and Masahiro Fukuda and I. Yu and Hiromu Yawo and Oded Béjà and Hideaki Kato and Tanja Woyke and Ichiro Takeuchi and Keiichi Inoue},
  title = {{Functional characterization of red-shifted rhodopsin channels from giant viruses explored by a machine-learning model for long-wavelength optogenetics}},
  year = {2025},
  doi = {10.1101/2025.09.16.676488},
  howpublished = {bioRxiv},
}
Figure