Thanks for your work on MaSIF and for making the source code available! I'm running MaSIF in the supplied docker container, with some tweaks (for example, I changed the scripts a little to output the high-scoring docked structures for manual review).
I'm running on a very small list of proteins of interest to me. I first ran the proteins through the default pipeline where it downloads the complexes directly from the PDB to recompute the benchmark, followed by the second stage alignment. This was fairly successful.
However, in the next stage, I decided to try something more like a "real world" experiment where I would not necessarily have a native structure; instead, I would have two chains that I suspect interact but don't know how. So I used the same pairs from above and did some standard light protein prep (no backbone changes) and rotation to emulate the perhaps-unknown position of the two proteins to one another. There was a marked decrease in performance, and upon further trials it became clear that the protein preparation was not an issue but that rotation of the chains contributed to a huge decrease in performance, and even an inability to find anything close to a native structure despite previously finding near-native structures with ease.
Based on what I understand from the paper and the code, the interface descriptors and alignment phase should be agnostic to initial rotation of the chain. Is there something I'm missing here?
Hello,
Thanks for your work on MaSIF and for making the source code available! I'm running MaSIF in the supplied docker container, with some tweaks (for example, I changed the scripts a little to output the high-scoring docked structures for manual review).
I'm running on a very small list of proteins of interest to me. I first ran the proteins through the default pipeline where it downloads the complexes directly from the PDB to recompute the benchmark, followed by the second stage alignment. This was fairly successful.
However, in the next stage, I decided to try something more like a "real world" experiment where I would not necessarily have a native structure; instead, I would have two chains that I suspect interact but don't know how. So I used the same pairs from above and did some standard light protein prep (no backbone changes) and rotation to emulate the perhaps-unknown position of the two proteins to one another. There was a marked decrease in performance, and upon further trials it became clear that the protein preparation was not an issue but that rotation of the chains contributed to a huge decrease in performance, and even an inability to find anything close to a native structure despite previously finding near-native structures with ease.
Based on what I understand from the paper and the code, the interface descriptors and alignment phase should be agnostic to initial rotation of the chain. Is there something I'm missing here?
Thanks!