Machine learning-assisted LOV protein engineering.
LOV photoreceptors control light-dependent processes such as phototropism in higher plants, but are also found in algae, fungi, and bacteria. They can perceive light below a wavelength of 470 nm and are therefore blue light sensors. LOV stands for the light-oxygen-voltage-sensing domain of this protein class. As a cofactor, the LOV domain contains a non-covalently bound flavin molecule as a chromophore. Engineered variants of LOV proteins are used as reporter proteins and optogenetic switches. An interdisciplinary team of scientists from all over Germany and the IPB have recently further optimized these photoreceptors with the help of machine learning. They recently presented their results in JACS Au.
The photocycle of LOV proteins begins when UV-A or blue light is absorbed by the oxidized flavin, which triggers chemical rearrangements and conformational changes in the protein. Depending on the type of LOV receptor, this leads to various physiological downstream reactions. In the dark, the LOV protein returns to its inactive state. This reverse reaction may comprise complex sub-steps and takes several seconds, or even days in the case of some LOVs.
It is precisely this transition to the dark state, known as dark recovery, with its large range of reaction times that is of particular interest to scientists, especially with a view to producing sensors with desired photocycle durations in the future. Such designer photoreceptors could enable optimized photosynthesis processes and thus increased plant biomass production.
For the dark recovery due to many contributing factors, no conclusive mechanistic model has yet been established. Previously, only individual contributing factors were investigated in detail and mutational studies were carried out on various critical amino acid residues around the chromophore. However, the simultaneous consideration of multiple mutations poses a combinatorial challenge. In addition, typical modeling at the molecular level is not feasible, as the dark state recovery is still a relatively slow process, even when lasting a few seconds, and would require enormous computing time. The research team therefore resorted to machine learning (ML) methods, which are ideally suited to predicting effects on protein function based on existing data from mutational studies and without the classical physics-based modeling.
The research team performed three rounds of prediction and validation on the phototropin 1 LOV2 domain of Avena sativa (AsLOV2), one of the best-studied LOV domains. They trained their ML model with suitable data sets from mutational studies on AsLOV2 to subsequently predict single, double, and triple mutants of AsLOV2 with specific dark recovery rates. The results from the experimental validation of predicted variants were always incorporated into the subsequent round of predictions. In the present study, the authors discuss in detail the predicted and tested substitutions, their structure-function implications, and their effects on the time of the recovery to the dark state. Their findings enabled the scientists to predict and generate several LOV domain variants whose dark recovery lasted from 0.4 to approx. 105 seconds. These variants represent the broadest spectrum to date in terms of photocycle duration, i.e. the fastest and slowest reverting AsLOV2 domains, which also proved to be stable in experimental studies.
At the same time, the researchers discuss the limitations of their ML approach. As the model was only based on a limited amount of mutation data, there is a certain bias in the predictions. A more comprehensive saturation mutagenesis scan could provide more balanced training data for an ML model in the future. Nevertheless, the potential of the ML-based approach for the evolution of protein properties that elude rational design is very promising, the authors conclude.
Original publication:
Hemmer S, Siedhoff NE, Werner S, Ölçücü G, Schwaneberg U, Jaeger K-E, Davari MD, Krauss U. 2023 Machine Learning-Assisted Engineering of Light, Oxygen, Voltage Photoreceptor Adduct Lifetime. JACS Au. 3, 3311–3323. 10.1021/jacsau.3c00440