Synthetic organic chemistry involves in drug discovery and chemical biology. Retrosynthesis is one of the most complex issues in the field of organic chemistry, which is the design of efficient synthetic routes for a given target. It is an efficient and environmental-friendly synthesis of valuable molecules with well-designed and feasible routes. Evidence show that similar products tend to be produced by similar reactions (Reaxys or SciFinder). For general retrosynthesis planning, a proposed step can come from any reaction class. Key considerations include the need to identify a cascade of disconnections schemes, suitable building blocks and functional group protection strategies. Many additional considerations in synthetic route planning are not limited to cost, process complexity, reaction yield, workup difficulty, safety, and toxicity of intermediates.
Artificial intelligence (AI), driven by improved computing power, data availability and algorithms, underpins chemical drug development. AI-assisted retrosynthetic analysis is starting from the target compound and working backward, which has been well-reviewed over the years. AI approaches have also been reported for prediction of reaction outcomes and optimization of reaction conditions. Potential application of retrosynthetic program may play an important role in de novo molecular design and automated synthesis of molecules.
Computer-aided Retrosynthetic Route PlanningThe earlier retrosynthesis programs are mainly computer-aided retrosynthetic analysis tools. Template-based methods for retrosynthetic analysis rely on human knowledge of organic synthesis, as well as the encoding of organic and mechanistic rules.
Confirm the extent of generalization and abstraction. Choose reaction databases. Encode reaction templates or synthon generation rules. Generalize subgraph matching rules (subgraph isomorphism problem). Extract the meaningful context around the reaction center. Baseline model (Liu et al). Synthia (formerly Chematica) developed by Grzybowski and coworkers. ReactionPredictor from Baldi's group based on mechanistic views. Retrieve reaction precedents from the knowledge base. Template extraction approaches. Only the atoms which are immediately
involved in the reaction are contained in these templates. It is specified by atomic identity, aromaticity,
number of hydrogen atoms, and chirality if applicable. Score and rank candidate precursors relying on quantitative
similarity scores (Morgan2noFeat / Morgan3noFeat / Morgan2Feat /Morgan3Feat fingerprints and the Tanimoto metric,
graph neural networks). Trade-off between generalization and specificity. Template extraction algorithms only consider reaction centers
and their neighboring atoms rather than take chemical environment of molecules into considerations. Unsettled problems of mapping the atoms between products and reactants,
and the abundance of distinct leaving groups for equivalent reaction sites. Mechanisms outside the knowledge database cannot be predicted.AI-assisted Retrosynthesis Planning Encode chemical reactions as sentences using reaction SMILES in the natural language (NL) framework. Treat forward- or retro- reaction prediction as a translation problem, using different types of neural machine translation architectures.
We present a template-free approach that is independent of reaction templates, rules, or atom mapping, to implement automatic retrosynthetic route planning.
Forward reaction prediction (Molecular Transformer). Neural model based on the sequence-to-sequence (seq2seq) architecture (Liu et al.). Template-free self-corrected retrosynthesis predictor (Zheng et al.). Monte-Carlo tree search algorithms (Segler et al.) One-step retrosynthetic model. Multistep synthesis plans. Hyper-graph exploration strategy for automatic retrosynthesis route planning.
Figure 1 Schematic of the Multi-step retrosynthetic workflow. (Philippe Schwaller, et al. 2019)
Figure 2 Workflow of Transformer-based Retrosynthesis
In CD ComputaBio, the developed approach mimics the retrosynthetic strategy, which is defined implicitly by a corpus of known reactions without the need to encode any chemical knowledge. We have also tried token- and character-based methods to tokenize the SMILES strings as model input, and used the open-source chemoinformatics software RDKit to validate the model. CD ComputaBio has assessed the entire framework by reviewing several retrosynthetic problems to highlight strengths and weaknesses.
A