Predicting reactants from a specified core product remains a critical challenge in retrosynthesis prediction. While semi-template-based and graph-edit-based methods have shown promising results in accuracy and interpretability, they struggle to handle complex reactions. In this paper, complex reactions refer to chemical reactions involving the participation of multiple bonds, such as those involving the multiple reaction center or the same leaving group being attached to multiple atoms. To address these limitations, we propose RetroSiG (Retrosynthesis via Search in (Hyper) Graph), a semi-template-based framework that reformulates retrosynthesis as a two-phase search problem: (i) reaction center identification as a search task in the product molecular graph, and (ii) leaving group completion as a search task in the leaving group hypergraph. RetroSiG’s novel search mechanism systematically explores subgraphs by leveraging reinforcement learning to guide node selection at each step. This approach ensures connectivity and efficient decision-making, thereby enabling the effective handling of complex reaction predictions. In addition, RetroSiG incorporates the one-hop constraint, a domain-specific prior inspired by the observation that reaction center molecular subgraphs and leaving group subgraphs are almost always connected. This constraint focuses exploration on first-order neighbors, significantly reducing the search space and improving computational efficiency without compromising accuracy. Furthermore, RetroSiG leverages a hypergraph structure to model implicit dependencies among leaving groups, which enhances robustness for reactions with multiple leaving group configurations. Comprehensive experiments demonstrate RetroSiG’s competitive performance across standard benchmarks, while ablation studies confirm the contributions of its key design components, including the hypergraph and the one-hop constraint. Our results highlight RetroSiG’s scalability and effectiveness in handling diverse and complex retrosynthesis tasks.

Graph similarity measurement is a fundamental task in various graph-related applications. However, recent learning-based approaches lack interpretability as they directly transform interaction information between two graphs into a hidden vector, making it difficult to understand how the similarity score is derived. To address this issue, we propose an end-to-end paradigm for graph similarity learning called Similarity Computation via Maximum Common Subgraph Inference (INFMCS), which is more interpretable. Our key insight is that the similarity score has a strong correlation with the Maximum Common Subgraph (MCS). We implicitly infer the MCS to obtain the normalized MCS size, with only the similarity score being used as supervision information during training. To capture more global information, we stack vanilla transformer encoder layers with graph convolution layers and propose a novel permutation-invariant node Positional Encoding. Our entire model is simple yet effective. Comprehensive experiments demonstrate that INFMCS consistently outperforms state-of-the-art baselines for graph-graph classification and graph-graph regression tasks. Ablation experiments verify the effectiveness of our proposed computation paradigm and other components. Additionally, visualization and statistical analysis of results demonstrate the interpretability of INFMCS.