Follow

Representing the machine learning community by giving a talk at in jeans and sneakers.

Every other speaker wore a whole suit and tie!

(also every speaker was male but that's a different topic)

@KevinKaichuang I enjoyed your talk! Have you considered using some of the more recently developed equivariant architectures to incorporate more refined geometric information for the structural component of your architecture? 🙂

@mspells I've played around a bit but they don't make a big difference for these sorts of things

@KevinKaichuang
That's interesting, thanks! I would have thought that being able to incorporate structural info about neighbors that are nearby in space but not necessarily in sequence order could be useful, but maybe it depends on how much training data are available or something like that.

@mspells That's what the MIF and MIF-ST papers do right? I just didn't think the architectures were that recent.

@KevinKaichuang
Maybe I'm misunderstanding, but I thought that the MIF work just used the backbone torsion angles of neighbors in sequence order; I was thinking that structural info derived from Cartesian coordinates of some number of nearest neighbors (in 3D space for the folded structure, not in sequence order space, I mean) might also be informative.

@KevinKaichuang Sorry for being unclear; I'm suggesting that, rather than using the rotation-invariant attributes of torsion angles for neighbors along the sequence, you could use a set of learned, rotation invariant attributes that represent the geometry of the nearby point cloud (i.e. not restricted to neighbors as you travel along the sequence, but rather k-nearest neighbors in 3D) for each amino acid. Sorry again if my point is still unclear; happy to clarify further if it would help!

Sign in to participate in the conversation
mastodon.cloud

Everyone is welcome as long as you follow our code of conduct! Thank you. Mastodon.cloud is maintained by Sujitech, LLC.