Have you ever found yourself frustrated by the poor accuracy of metabolic flux predictions when using unconstrained genome-scale metabolic models? Perhaps you've gathered extensive data and sought a solution through machine learning models for predicting metabolic fluxes. In this blog post, we'll delve into our journey of conceiving an innovative hybrid model that addresses these challenges and share the steps we took to bring this new model to life.

On the other side of the modeling realm, seemingly opposed to mechanistic models, machine learning models have proven to be very useful in guiding experimentation ². They also have proven to handle high-complexity problems, such as protein folding ³. We assessed these capabilities should be helpful, given the complex relationship (based on many regulatory mechanisms) between the controlled experimental setup of an experimentalist, and the metabolic fluxes found in an organism.

Observing the pros and cons of GEMs (mechanistic) and ML models, we investigated the field of hybrid modeling. Flux Balance Analysis (FBA) is the usual computation framework to infer predictions with GEMs. Consequently, the first step of the project was to build FBA-surrogating methods that can be integrated into neural network models. We successfully developed three kinds of FBA-surrogating methods. We found inspiring approaches, based on recurrent signaling networks ⁴, Hopfield’s networks ⁵, and Physics Informed Neural Networks ⁶. Using these approaches, and not without much effort, we decided to attempt an integration of GEMs with neural networks.

Then, we integrated these FBA-surrogating methods inside a neural network architecture, yielding our final hybrid model architecture. Naturally, we assessed their performance on many datasets, including FBA simulations and experimental data from bacterial species. Our results exhibited drastically improved regression performance on the growth rate, between FBA on GEMs and our hybrid models. We also observed that our hybrid models needed an amount of training data much lower than traditional machine learning methods. However, we still lacked a guarantee that hybrid model predictions were actually meaningful and reliable.

The final step of the project was to build a ‘reservoir computing’ approach. A hybrid model accurately mimicking FBA was obtained from training on simulation data. After freezing its parameters, we called this model a ‘reservoir’. Then, we learn from experimental data the best inputs for this reservoir to make accurate predictions. This approach enabled us to extract upper-bounds values of nutrients for all experimental conditions and use them as input of FBA. The FBA running with such inputs showed similar performance to the hybrid models alone, thus demonstrating that our hybrid models could indeed make meaningful predictions.

In summary, our study serves as a pioneering step towards advancing genome-scale metabolic network computation frameworks. As we embark on this journey of hybrid modeling for GEMs, we anticipate the community's collective efforts in developing more robust and fine-tuned hybrid GEM models. Ultimately, we envision a transformative shift in the computational framework paradigm for metabolic network modeling.

References:

1. Reed, J. L. & Palsson, B. Ø. Thirteen years of building constraint-based in silico models of Escherichia coli. J. Bacteriol. 185, 2692–2699 (2003).
2. Borkowski, O. et al. Large-scale active-learning-guided exploration for in vitro protein production optimization. Nat. Commun. 11, 1872 (2020).
3. Jumper, J. et al. Highly accurate protein structure prediction with AlphaFold. Nature 596, 583–589 (2021).
4. Nilsson, A., Peters, J. M., Meimetis, N., Bryson, B. & Lauffenburger, D. A. Artificial neural networks enable genome-scale simulations of intracellular signaling. Nat. Commun. 13, 3069 (2022).
5. Hopfield, J. J. & Tank, D. W. ‘Neural’ computation of decisions in optimization problems. Biol. Cybern. 52, 141–152 (1985).
6. Raissi, M., Perdikaris, P. & Karniadakis, G. E. Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations. J. Comput. Phys. 378, 686–707 (2019).