Fig 1.
The overarching architecture of the TGF-M, which is divided into two main components: feature engineering and predictor.
Fig 2.
TGF-M adopts a prediction architecture that comprises three key components: K-hop convolution, virtual nodes, and linear attention.
Table 1.
The benchmark model of the PCQM4MV2.
Fig 3.
Bar chart of model’s parameters and MAE.
Light blue bars indicate the parameter count, while dark blue bars indicate the MAE metric.
Fig 4.
Comparison of scatter plots between predicted and true values across models.
The horizontal axis denotes the true value, while the vertical axis denotes the predicted value. The closer the sample points are to the diagonal red line, the better the predictive performance.
Fig 5.
Violin plot comparison of the predictions from models.
Thicker sections represent areas of higher data density, while thinner sections indicate fewer observations. The central marker typically shows the median.
Table 2.
Experimental results of adding different topological information.
Fig 6.
Line chart and Bar chart of models’ parameters and MAE.
The line chart represents the parameter values, while the bar chart illustrates the MAE.
Table 3.
Ablation study results on different molecular encoding features.
Table 4.
Ablation study results on predictor modules.
Fig 7.
T-SNE visualization at different stages of the training process.
The color intensity denotes the magnitude of the molecular energy gap, and the training process organizes the molecular representations within the latent space in a structured manner.
Fig 8.
T-SNE visualization of Conjugated vs. Non-Conjugated molecules.
Red indicates conjugated molecules, while blue indicates non-conjugated ones. The results illustrate the projection of various molecules in two-dimensional and three-dimensional space after model training.
Fig 9.
T-SNE visualization of Aromatic vs. Non-Aromatic molecules.
Purple indicates Aromatic molecules, while yellow indicates Non-Aromatic molecules. The results illustrate the projection of various molecules in two-dimensional and three-dimensional space after model training.