Fig 1

Overview of the MolUNet++ architecture. (a) The architecture of MolUNet++ incorporates pooling and unpooling operations as down-sampling and up-sampling processes, respectively. Each layer following the unpooling operation retains feature representations from the previous layers of the same scale through skip connections. Ultimately, molecular embeddings at various granularities are combined for prediction tasks. (b) The edge feature update process within the edge message passing (E-MP) module. (c) The pipeline of the MESPool process: node features and edge features are integrated to represent undirected bonds, which are then scored by a query feature. The edges are classified as inner-substructure and backbone bonds, with nodes clustered accordingly and shrunk into super-nodes. (d) An illustration of the pooling and unpooling operations applied to the dopamine toy structure. The unpooling operation reconstructs the original graph structure, copies and aligns the super-node feature to its member nodes, and resets the features of the inner-substructural edges to zero vectors. (e) The substructure explainer: MolUNet++ encodes atom features and merges them into a molecular representation through global pooling, subsequently predicting the property $ y $. By masking the atom features of substructures identified by MESPool, a new prediction, $ y_{\text{sub}} $, is obtained for the molecule with those substructures removed. The contribution of each substructure to the property is determined by $ y\text{-}y_{\text{sub}} $, and this contribution is visualized to provide insights."

Fig 2

Utilization of structural identification query (Equation3) across different tasks. (a) Molecular property prediction: Three molecular structural fingerprints are concatenated and encoded using a multi-layer perceptron (MLP). The resulting feature serves as the query feature $ H_{\text{query}} $ in MESPool, and is subsequently decoded by another MLP for property prediction in conjunction with the graph embedding. (b) A pair of molecules is encoded using the same MolUNet++ network, with each molecule mutually utilizing the other's graph features as $ H_{\text{query}} $. The output embeddings are integrated through a co-attention layer to predict DDI interactions. (c) DTI prediction: The protein sequence is encoded using 1D CNNs, and the average of the representation rows is transformed through an MLP to serve as the $ H_{\text{query}} $ for structural identification in the drug encoder. The drug and protein representations are then fed into a bilinear attention network to learn their pairwise local interactions."

"

Fig 3

Benchmark performance. (a) Property prediction: the figures compare non-pretrained models (the first 4 bars) with pretrained models (the last 8 bars). The first 5 charts utilize AUC-ROC as the metric, where higher bars indicate better performance. The last 3 charts employ RMSE, where lower bars indicate better performance. (b) Interaction prediction: in each figure, the left two bar groups show DDI performance, while the right three show DTI performance. Each figure represents a different metric, where higher bars indicate better performance."

Fig 4

Visualization of pooling results and substructure attribution values in 2 property prediction tasks. Atoms of the same color represent an identified fragment, while independent atoms are also regarded as a single atomic substructure. (a) $ \text{log}D $ prediction: positive values indicate an increase in hydrophobicity (lipophilicity), whereas negative values indicate the contrary; (b) $ {\text{LD}}_{\text{50}} $ toxicity prediction: positive values suggest increased lethality."

"

Fig 5

Visualization of the case study in DDI prediction. (a) Attention scores visualization in the original SSI-DDI; (b) pooling results and identified substructure contribution visualization in MolUNet++."

Fig 6

Visualization of the case study in DTI prediction. (a) The ligand-protein interaction from the corresponding crystal structures (visualized through the Molecular Operating Environment (MOE) software); (b) visualization of the most important atoms in the bilinear attention layer of DrugBAN, that contribute to the binding; (c) pooling results and identified substructure contribution visualization in MolUNet++."