In current research on protein-ligand binding affinity prediction, deep learning scoring functions based on Interaction Graph Neural Networks (IGNN) have shown excellent performance and development potential. This is largely due to the integration of physicochemical inductive biases into protein-ligand interaction representations in such models, enabling them to learn and capture key interaction features in complex structures for effective geometric affinity prediction. However, many IGNN models still use homogeneous graphs for representation, which partially ignores non-covalent interactions that dominate ligands. In message passing, this approach weakens the impact of covalent bonds within ligand molecules on node feature updates, leading to potential information loss, as shown in Figure 1(a).

Additionally, most 3D-GNN models adopt a "bottleneck" architecture: after multiple graph convolution layers, the input graph structure is typically compressed into a single vector representation via global pooling (as shown in Figure 1(b)). This processing may cause the loss of 3D structural information during compression. To address these issues, the paper proposes an improved affinity scoring method called EHIGN. EHIGN uses a heterogeneous graph modeling approach (Figure 1(c)) and treats binding affinity as the sum of contributions from non-covalent interactions between protein and ligand atoms, with potential biases corrected via a bias correction term (Figure 1(d)). This significantly enhances the model’s generalization ability. Related results were published in TPAMI (https://ieeexplore.ieee.org/abstract/document/10530021).

In compound property prediction, deep neural networks have made significant progress in accelerating and enhancing drug discovery. However, these technologies often require large amounts of labeled data to accurately predict molecular properties. In practice, during the early stages of drug development, the lack of data on physicochemical properties and biological activities of novel molecules or their analogs remains a major challenge. This predicamentmakes applying deep neural networks to few-shot drug discovery extremely difficult.

To address this, the paper proposes a meta-learning framework combining Graph Attention Networks (GAT) — Meta-GAT — for compound property prediction under low-data conditions, as shown in Figure 3. GAT captures local influences of atomic groups at the atomic level through a triple attention mechanism, effectively learning the contribution of atomic groups to overall compound properties. Additionally, the research team constructed a meta-learning benchmark dataset specifically for compound property prediction (Meta-molnet), as shown in Figure 4. Related results were published in TNNLS (https://ieeexplore.ieee.org/abstract/document/10436119;https://ieeexplore.ieee.org/abstract/document/10059171).

In the field of retrosynthetic optimal step prediction, current retrosynthetic models are limited by their prediction accuracy, diversity, and interpretability, restricting their practical application in synthetic route planning. Enhancing AI-based retrosynthetic prediction models from the perspective of how chemists reason about reactions remains an urgent and important research topic.

To this end, based on the simplified mechanism of reaction transformation, the paper proposes a graph-to-edit architecture called Graph2Edits based on graph neural networks for retrosynthetic prediction, as shown in Figure 5. Specifically, the model frames retrosynthetic reaction prediction as a process of deriving product-intermediate-reactant through a series of interconnected graph edits to learn reaction transformation rules, mimicking how chemists reason about reaction mechanisms. The end-to-end architecture generates graph edit sequences of arbitrary length in an autoregressive manner, strengthening connections between multiple generation steps and improving applicability in multi-center reactions and prediction diversity.

Directional Message Passing Neural Networks (D-MPNN) are used to encode local atom/bond and global graph features, fully leveraging compound structural information to predict atom/bond edits and generate termination symbols. Subgraphs as leaving groups are added to intermediates to complete reactant generation, approximating more realistic reaction transformation processes, significantly reducing generation steps, and improving prediction performance. Related results were published inNature Communications(https://www.nature.com/articles/s41467-023-38851-5).