Another protein structure diffusion paper just got published and once again, the presented method runs into trouble with beta-sheets (sequence design via ProteinMPNN), frequently failing to reach a generous scTM-score cutoff of 0.5 doi.org/10.1038/s41467-024-45051-2 - ThreadSky

ddelalamo.bsky.social • 387 days ago

Another protein structure diffusion paper just got published and once again, the presented method runs into trouble with beta-sheets (sequence design via ProteinMPNN), frequently failing to reach a generous scTM-score cutoff of 0.5
https://doi.org/10.1038/s41467-024-45051-2

Comments

ddelalamo.bsky.social•387 days ago

Late last year it was noted that backbones w sheets refold are more difficult to design and thus get generally lower scTM than purely helical proteins, independent of sequence design method

https://arxiv.org/abs/2312.00080

lucasnivon.bsky.social•387 days ago

Since early after top7 it has been difficult to design mixed alpha beta and especially all beta proteins. NFT2 fold and jelly roll (mostly Enrique and Tamuka) were big steps forward. https://pubmed.ncbi.nlm.nih.gov/30374087/

ddelalamo.bsky.social•387 days ago

So my comparison point is the failed AF hallucinations from the ProteinMPNN paper. They crystallized a mixed a/b protein allegedly selected from 73 successfully expressed proteins (out of 96 total attempted genes). Anecdotally seems like hallucination, but not diffusion, gets those geometries right

ddelalamo.bsky.social•387 days ago

Here are similar results reported by authors of Chroma. Left: experimental sheet content is consistently lower than the design model. Right: confirmation of lower scTM.

ddelalamo.bsky.social•387 days ago

The diffusion model GENIE also shows a propensity against wanting to design beta-sheets, never exceeding <30% of total sequence length (compared to >90% helical for some proteins)

https://arxiv.org/abs/2301.12485

ddelalamo.bsky.social•387 days ago

For reference, ProteinMPNN seems to have a respectable sequence recovery of ~35% of beta barrels, and that its designs of experimentally determined beta-barrels can be refolded with RMSDs of ~1.8 Å

https://www.biorxiv.org/content/10.1101/2024.01.16.575764

ddelalamo.bsky.social•172 days ago

An addendum to this, Possu Huang mentioned at EMBL-EBI this spring that filtering w protein folding NNs might bias toward de novo helical backbones vs sheets; Cheng & Perez last month presented evidence that these are indeed better at folding helical backbones https://www.biorxiv.org/content/10.1101/2024.08.25.609581

Comments

Posting Rules

Reply