We thank the reviewers for their detailed and thoughtful feedback. We respond to each reviewer individually below. So it is a mean (across training seeds) of the median (across complexes) of AUROC. We will clarify this in the Table 2 and Figure 1 captions. We will clarify this point.

The authors propose the first end-to-end learning model for protein interface prediction, the Siamese Atomic Surfacelet Network (SASNet). The novelty of the method is that it only uses spatial coordinates and identities of atoms as inputs, instead of relying on hand-crafted features. The authors also introduce the Dataset of Interacting Protein Structures (DIPS) which increases the amount of binary protein interactions by two orders of magnitude over previously used datasets (DB5). The results outperform state-of-the-art methods when trained on the much larger DIPS dataset and are still comparable when trained on the DB5 dataset, showing robustness when trained on bound or unbound proteins. The paper is very well written and easy to follow.