1. Introduction

Chemical reaction selection and design play a key role in drug and material synthesis.1 The synthesis conditions (temperature, pressure, solvents, etc.), time and yield of the product can be greatly optimized through selecting an appropriate chemical reaction pathway. Therefore, the design of evaluation indicators for Computer Aided Synthesis Planning (CASP) has evolved in recent years.2-7 CASP evaluation indicators are mainly divided into two catalogs: expert knowledge-based evaluation indicators8,9 and synthesis complexity/accessibility-based evaluation indicators.10-17 For expert knowledge-based evaluation indicators, the rank of synthesis results are determined by experts.8,9 Although this kind of methods have a high confidence level, it still suffers from ambiguity and lack of objectivity. So, it is difficult to be applied in retrosynthesis tasks and provide an objectivity guidance on synthesis route.6 For synthesis complexity/accessibility-based evaluation indicators, the feasibility of synthesis is qualified by molecular structures and the reaction relationship between reactants and products.10-17 Although synthesis complexity/accessibility evaluation indicator eliminates the ambiguity and objectivity problems, the influence of reaction agents and conditions are still unable to be considered in these indicators.
Here, a brief overview of existing synthesis complexity/accessibility-based evaluation indicators is given. SAscore13 uses Extended Connectivity Fingerprints (ECFPs) 18 fragment analysis obtained from the compounds of PubChem database19. According to the frequency of each fragment occurrence, each fragment is assigned a numerical score. After combining the fragment score with the penalty for complexity and the bonus for symmetry, SAscore is able to measure compound synthesis accessibility on a high-throughput scale. SAscore is widely used in guiding synthesis directions in retrosynthesis.20,21 Based on the assumption that the complexity of the reactants is lower than products, a data-driven metric SCscore14 was designed to describe real syntheses. Trained by 22 million reactant-product pairs from the Reaxys22 database, SCscore is able to describe the complexity of the synthetic route.4 Although this evaluation metric differs from the metric of synthetic accessibility, it can also be used as a guide for retrosynthesis through the Morgan Fingerprints input. SYBA15 is a fragment-based method for the rapid classification of the synthesis difficulty of organic compounds. It uses Bernoulli Naïve Bayes classifier to assign SYBA score contributions to individual fragments based on their frequencies in the database of easy- (ES) or hard-to-synthesize (HS) molecules. Although it can be used to quickly rank large molecular datasets for high-throughput screening or molecular design, it still cannot compete with more sophisticated synthetic path reconstruction methods that enable the incorporation of other factors23. RAscore and GASA are the evaluation metrics using a similar method in retrosynthesis accessibility.16,17 Machine Learning (ML) is used in these methods to generate the probability of retrosynthesis accessibility. The data-driven models of RAscore and GASA were trained by using ES or HS labels generated by multistep retrosynthetic planning algorithm such as Retro*24 and AiZynthFinder.25 Although these developed evaluation metrics are able to clearly determine the difficulty of molecular synthesis, the impact of reaction agents is still unable to be considered.
With the development of ML, Graph Neural Networks (GNN) are gradually used in chemistry. In addition to predicting molecular thermodynamic properties in the dataset such as QM9,26-29 it has also been used in molecular generation,30,31reinforcement learning for molecular design,32molecular representation learning33-37 and reaction yield prediction38 in recent years. For the molecular representation learning method, SMILES Contrastive LeaRning (SMICLR) framework was proposed which embraces multimodal molecular data. It jointly trains a graph encoder and SMILES encoder to perform the contrastive learning. Through data augmentation on graphs and SMILES sequences, SMICLR model successfully reduced the prediction error for the energetic and electronic properties of the QM9 dataset.33 MolCLR is a self-supervised learning framework which performs graph data augmentation and contrastive learning method on a large unlabeled molecular database to achieve representation learning of molecules. Benefiting from pre-training on a large unlabeled database, MolCLR even achieves state-of-the-art results on several challenging benchmarks after fine-tuning.34GeomGCL designs a novel geometric graph contrastive scheme to enable collaborative supervision between 2D and 3D molecular graph geometric views, aiming to improve model generalization ability on molecular graph classification and regression.35 MoCL is a contrastive learning framework which utilizes domain knowledge at both local and global levels to learn molecular representations. By replacing valid substructures with bioisosteres that share similar properties, MoCL achieves accurate prediction of molecular properties, providing a suitable and powerful augmentation method for molecular graph.36 KCL builds a knowledge graph data augmentation module by using fundamental chemical attributes to connect atoms that are not directly connected by bonds.37 By using a double MPNN model, extensive experiments demonstrated that KCL obtained superior performance against state-of-the-art baselines on eight molecular datasets, demonstrating the feasibility of the framework for molecular representation learning. In summary, contrastive learning method shows a better performance on molecular properties prediction. It illustrates that contrastive learning method is able to help the model extract more features and improve prediction effect of molecular properties.
In this work, we migrate the generation method of molecular synthesis accessibility to reaction superiority and design a reaction total atom-atom mapping algorithm to complement the atomic mapping relationship in the chemical reaction database. By using the reaction descriptors constructed from the reaction mapping relationships and reaction reagents, a chemical reaction representation learning model is constructed through a contrastive learning method. After fine-tuning the model on a binary classification task for determining reaction superiority, reaction superiority score (RSscore) is generated to evaluate the superiority of chemical reactions and further applied on reaction evaluation and synthesis route analysis.