Kinetic parameter prediction using neural networks identifies limitations to C 4 photosynthesis

Abstract

Kinetic models of photosynthesis enable time-resolved predictions of traits related to this key process and provide the means to identify factors limiting photosynthesis. However, the use of large-scale models is currently limited by the lack of efficient approaches to estimate the hundreds of genotype-specific kinetic parameters. Here, we present C4TUNE, an artificial neural network that can efficiently predict parameters of a large-scale photosynthesis model from photosynthesis response curves. C4TUNE was trained on a biologically relevant synthetic dataset comprising matched samples of parameters and response curves obtained using a C 4 photosynthesis kinetic model. To speed up the training of C4TUNE, we devised a surrogate neural network to predict photosynthesis response curves directly from the model parameters and environmental inputs. Given response curves as input, we showed that over 99% of the parameter vectors predicted by C4TUNE could be used directly in simulation of the kinetic model and resulted in excellent fits. Finally, we applied C4TUNE to predict parameters for a population of 68 maize genotypes across two seasons. The predicted genotype-specific parameters allowed pinpointing factors that limit photosynthetic efficiency, validated using simulations. Therefore, the use of C4TUNE presents a fast and precise approach for parameter prediction based on minimal datasets.

Code and data availability

The paper's Data Availability Statement provides a public GitHub repository with the authors' custom code for artificial dataset generation, neural network definition/training, and predicted maize genotype parameters. Zenodo datasets (gas exchange measurements and synthetic training data) are mentioned via DOIs but no