From Theory to Experiment: Transformer-based Generation Enables Rapid Discovery of Novel Reactions

Deep learning methods have been proven their potential roles in the chemical field, such as reaction prediction and retrosynthesis analysis. However, the de novo generation of unreported reactions using artificial intelligence technology remains not be completely explored. Inspired by molecular generation, we proposed the task of novel reaction generation. In this work, we applied the Heck reactions to train the transformer model, state-of-art natural language process model and obtained 4717 generated reactions after sampling and processing. We then confirmed that 2253 novel Heck reactions by organizing chemists to judge the generated reactions, and adopted organic synthesis experiment to verify the feasibility of unreported reactions. In this process, it only took 15 days from Heck reaction generation to experimental verification, proving that our model learns reaction rules in-depth and can make great contributions in the novel reaction discovery.

Conformational Sampling for Transition State Searches on a Computational Budget

Transition state searches are the basis for characterizing reaction mechanisms and activation energies, and are thus central to myriad chemical applications. Nevertheless, common search algorithms are sensitive to molecular conformation and the conformational space of even medium-sized reacting systems are too complex to explore with brute force. Here we show that it is possible to train a classifier to learn the features of conformers that conduce successful transition state searches, such that optimal conformers can be down-selected before incurring the cost of a high-level transition state search. To this end, we have benchmarked the use of a modern conformational generation algorithm with our reaction prediction methodology, Yet Another Reaction Program (YARP), for reaction prediction tasks. We demonstrate that neglecting conformer contributions leads to qualitatively incorrect activation energy estimations for a broad range of reactions, whereas a simple random forest classifier can be used to reliably down-select low-barrier conformers. We also compare the relative advantage of performing conformational sampling on reactant, product, and putative transition state geometries. The robust performance of this relatively simple machine learning classifier mitigates cost as a factor when implementing conformational sampling into contemporary reaction prediction workflows.

On the influence of template size, canonicalization and exclusivity for retrosynthesis and reaction prediction applications

Heuristic and machine learning models for rank-ordering reaction templates comprise an important basis for computer-aided organic synthesis regarding both product prediction and retrosynthetic pathway planning. Their viability relies heavily on the quality and characteristics of the underlying template database. With the advent of automated reaction and template extraction software and consequently the creation of template databases too large to be curated manually, a data-driven approach to assess and improve the quality of template sets is needed. We therefore systematically studied the influence of template generality, canonicalization and exclusivity on the performance of different template ranking models. We find that duplicate and non-exclusive templates, \textit{i.e.} templates which describe the same chemical transformation on identical or overlapping sets of molecules, decrease both the accuracy of the ranking algorithm and the applicability of the respective top-ranked templates significantly. To remedy the negative effects of non-exclusivity, we developed a general and computationally efficient framework to deduplicate and hierarchically correct templates. As a result, performance improved for both heuristic and machine learning template ranking algorithms across different template sizes. The canonicalization and correction code was made freely available.

Constructing xenobiotic maps of metabolism to predict enzymes catalyzing metabolites capable of binding to DNA

Background The liver plays a major role in the metabolic activation of xenobiotics (drugs, chemicals such as pollutants, pesticides, food additives...). Among environmental contaminants of concern, heterocyclic aromatic amines (HAA) are xenobiotics classified by IARC as possible or probable carcinogens (2A or 2B). There exist little information about the effect of these HAA in humans. While HAA is a family of more than thirty identified chemicals, the metabolic activation and possible DNA adduct formation have been fully characterized in human liver for only a few of them (MeIQx, PhIP, A$$\alpha$$ α C). Results We have developed a modeling approach in order to predict all the possible metabolites of a xenobiotic and enzymatic profiles that are linked to the production of metabolites able to bind DNA. Our prediction of metabolites approach relies on the construction of an enriched and annotated map of metabolites from an input metabolite.The pipeline assembles reaction prediction tools (SyGMa), sites of metabolism prediction tools (Way2Drug, SOMP and Fame 3), a tool to estimate the ability of a xenobotics to form DNA adducts (XenoSite Reactivity V1), and a filtering procedure based on Bayesian framework. This prediction pipeline was evaluated using caffeine and then applied to HAA. The method was applied to determine enzymes profiles associated with the maximization of metabolites derived from each HAA which are able to bind to DNA. The classification of HAA according to enzymatic profiles was consistent with their chemical structures. Conclusions Overall, a predictive toxicological model based on an in silico systems biology approach opens perspectives to estimate the genotoxicity of various chemical classes of environmental contaminants. Moreover, our approach based on enzymes profile determination opens the possibility of predicting various xenobiotics metabolites susceptible to bind to DNA in both normal and physiopathological situations.