(449j) Transition State Geometry Prediction Using Neural Embeddings of Transition State Graphs

Transition State Theory (TST) with modern quantum chemistry methods allow reaction rates coefficients to be calculated with high accuracy, but finding the transition state is computationally expensive and difficult to automate. The key step in TST is to determine the transition state (TS) geometry, which is mostly determined by reaction family, reaction center, and reactant geometries. Inspired by neural word embeddings techniques, we introduce rxn2vec, which is an unsupervised machine learning approach to create vector representations of transition states. We build upon a neural embedding framework named graph2vec [1] to learn data-driven distributed representations of arbitrarily sized graphs. In this case, fully optimized TS geometries generated by our automated TST calculator (AutoTST) [2] are converted to undirected graphs using connectivity matrices. These training data then provide the framework to predict key TS reaction center geometries. We demonstrate our tool on a set of reactions, predict key distances from these reaction, and implement AutoTST to perform partial geometry optimizations to converge on a finalized geometry. So far this tool can be used to predict the reaction centers of three reaction families, namely hydrogen abstraction, intramolecular hydrogen migration, and radical addition to multiple bond.

[1] graph2vec: Learning Distributed Representations of Graphs
A. Narayanan, M. Chandramohan, R. Venkatesan, L. Chen, Y. Liu, S. Jaiswal (2017)
arXiv:1707.05005 [cs.AI]

[2] Automated Transition State Theory Calculations for High-Throughput Kinetics
Pierre L. Bhoorasingh, Belinda L. Slakman, Fariba Seyedzadeh Khanshan, Jason Y. Cain, and Richard H. West. J. Phys. Chem. A.121 (37), 6896-6904, (2017) https://doi.org/10.1021/acs.jpca.7b07361