A Machine Learning Approach to Solving Biochemical Systems

Various application areas are making valuable use of machine learning techniques in the form of regression, classification, clustering, neural networks, and reinforcement learning. The development and deployment of such machine learning techniques are facilitated by the availability of large datasets together with the significant advances we are witnessing in computational storage and power, enabling more accurate predictions and faster research progress. In this work, we apply a machine learning framework called Nessie to some notable biochemical systems to generate the marginal probabilities that describe the evolution of those systems over time. We compare with the Stochastic Simulation Algorithm (SSA) and an adaptive Finite State Projection based on Quantized Tensor Train (FSP-QTT). The machine learning approach after training shows an accuracy at the level of the traditional SSA and FSP-QTT with the advantages of re-usability of trained models. Nessie can only train and predict for one species at a time unlike SSA although the whole process can be repeated for all species in a system. Also, Nessie only learned one-dimensional marginal distributions predicted by the CME, hence cannot capture joint distributions.