Fig. 1. Overview of RetroExplainer.

a The pipeline of RetroExplainer. We formulated the whole process as four distinct phases: (1) molecular graph encoding, (2) multi-task learning, (3) decision-making, and (4) prediction or multi-step pathway planning. b The architecture of the multi-sense and multi-scale Graph Transformer (MSMS-GT) encoder and retrosynthetic scoring functions. We considered the integration of multi-sense bond embeddings with both local and global receptive fields, blending them as attention biases during the self-attention execution phase. Upon obtaining shared features, we employed three distinct modules to evaluate the probabilities of five retrosynthetic events. These comprise: the reaction center predictor (RCP), which includes both a bond change predictor (RCP-B) and a hydrogen change evaluator (RCP-H); the leaving group matcher (LGM), enhanced with an additional contrastive learning strategy; and the leaving group connector (LGC). It is noteworthy to mention that the acronym MLP stands for multi-layer perceptron. c The dynamic adaptive multi-task learning (DAMT) algorithm. This algorithm is intended to acquire a group of weights according to the descent rates of losses and their value ranges to optimize the five evaluators equally. $l_i^t$ denotes the $i$ th kind of loss score in the $t$ th iteration. The $l_i^{avg}$ means the average of $i$ th type of loss value over the loss queue from $l_i^t$ to $l_i^{t-n}$, where $n$ is the length of queue we take into consideration. $w_i^t$ is the obtained weight of the $i$ th kind of loss score at the $t$ th iteration.$\tau$ is a temperature coefficient. d. The chemical-mechanism-like decision process. We designed a transparent decision process with six stages, assessed by five evaluators to obtain the energy scores $E_1,\dots E_5$. The $\mathrm{\Delta }{E}_i$ is the gap between the $E_i$ and $E_{i+1}$.