Table 2 UQ evaluation metrics for the latent space uncertainties of a GCNN model trained on 9k+1k data points of Crippen’s logP. For NN uncertainties, the latent space used is the very last layer of the NN. For GCN uncertainties, the latent space is the vector right after the pooling layer (LS-GCN)
LS | \(R^2\) | a | b | \(\rho _{rank}\) | \(\rho _{rank}^{sim}\) | \(A_{mis}\) | NLL | NLL\(^{sim}\) |
|---|---|---|---|---|---|---|---|---|
NN\(_{10\text {k}}\) | 0.31 | 1.12 | – 0.05 | – 0.04 | 0.07 (0.01) | 0.07 | 0.14 | 0.17 (0.01) |
NN\(_{150\text {k}}\) | 0.91 | 1.20 | – 0.03 | 0.17 | 0.24 (0.01) | 0.03 | – 0.50 | – 0.51 (0.01) |
GCN\(_{10\text {k}}\) | 0.24 | 0.65 | 0.09 | – 0.02 | 0.11 (0.01) | 0.07 | 0.14 | 0.18 (0.01) |
GCN\(_{150\text {k}}\) | 0.85 | 1.85 | – 0.13 | 0.23 | 0.13 (0.01) | 0.05 | – 0.46 | – 0.45 (0.01) |