Hybrid mechanistic–machine learning modeling of photosynthetic efficiency across plant species and biohybrid systems

Abstract Enhancing photosynthetic efficiency is central to improving crop productivity and climate resilience. We developed a unified computational framework that integrates a mechanistic redox–photosynthesis model with Long Short-Term Memory (LSTM) neural networks to simulate gene–environment interactions across plant systems. Public RNA-seq datasets from Arabidopsis thaliana, Oryza sativa, and Zea mays were processed to derive gene-set activity scores for antioxidant capacity, PSII repair, and Calvin–Benson cycle flux, which parameterized a coupled ODE model of light absorption, ROS generation, AsA–GSH cycling, PSII damage-repair, and carbon fixation. Synthetic time-series generated under heat, drought, and high-light stress reproduced expected ROS bursts, photoinhibition, and recovery kinetics, with Zea mays (C₄) showing faster redox recovery and higher carbon flux compared to C₃ species. An LSTM trained on these mechanistic simulations achieved high predictive accuracy on held-out synthetic data (R² ≈ 0.95) and demonstrated moderate performance when evaluated against external transcriptomic datasets from GEO/Zenodo, supporting the model’s ability to generalize beyond internal training conditions. As a proof-of-concept extension, we simulated biohybrid enhancement by introducing a quantum-dot/organic nanoparticle excitation term into the photon-input function, yielding computationally predicted improvements (30–90%) in water-splitting and carbon-fixation efficiency. These biohybrid outcomes are predictive and require experimental verification. Overall, the hybrid mechanistic–machine-learning framework provides a scalable and interpretable platform for generating testable hypotheses toward future gene- or material-assisted optimization of photosynthesis under climate stress. External validation using publicly available chlorophyll fluorescence datasets showed moderate predictive performance (R² ≈ 0.35–0.53), confirming generalization beyond synthetic domain. Graphical abstract