Graph neural processes for molecules: an evaluation on docking scores and strategies to improve generalization

Table 1 Coefficient of determination (\(R^2\)) of molecular NP models and baselines in FSL prediction of PARP1, ESR2 and PGR

Each cell shows the mean and standard error (in parentheses) of 10 random repetitions. The MG-CNP ranked consistently as the best model in terms of \(R^2\). Fine-tuning led to substantial improvements, especially when the meta-test task was poorly correlated with meta-training tasks (as in ESR2 and PGR) and there were many observations available

ISSN: 1758-2946