Figure 1.

The distribution of real and predictive affinity of GraphSAGE with targets, pockets, and their combination based on the test set. The performance of GraphSAGE, when harnessed in conjunction with both the sequence and structural knowledge of targets and pockets, surpassed that achieved by solely relying on the sequence and structural features of targets or pockets individually. Specifically, as compared to using only the features of targets, the MAE and RMSE exhibited reductions of 0.047 and 0.04, respectively. Moreover, improvements of 0.018 in PCC, 0.028 in Spearman, 0.013 in CI, and 0.024 in R² were realized. Similarly, in comparison to using only the features of pockets, the MAE and RMSE experienced decreases of 0.021 and 0.43, while enhancements of 0.017 in PCC, 0.018 in Spearman, 0.009 in CI, and 0.026 in R² were achieved.