Thanks for the answer! I hope then to see you in CASP (and CAMEO too, it is a great tool to test/refine your method). I was discussing a paper with a co-worker of mine (we also work on psp, we work on RBO Aleph). We had a hard time pinpointing the thing that made your method finally work. You have mentioned in your blog post that you have been working on it for years now, and I guess a lot of other people had the idea of using deep learning for psp. But what was the insight that made it all work, using LSTM? or was it many small refinements and hacks?

I would say the biggest thing is obviously the architecture, coupling LSTMs with the geometric units that spit out the actual 3D structure that can then be directly optimized via the dRMSD loss function. That's the biggest point of distinction from everything else out there (no contact map prediction, etc.) So it really is about end-to-end differentiability IMO, which hasn't been done before.

As for why it took so long, it is and it is not fine-tuning. Getting RGNs to train _at all_ was a rather difficult process, and required a lot of finicking around. But since I got them working, I haven't actually spent all that much time fine-tuning them, and so I expect there to be a lot of low-hanging fruit in terms of optimizing performance (starting from the baseline I found.)