iiinova — where your digital self keeps evolving

ok55991.bsky.social @ok55991.bsky.social · 22/06/2026

- The non-folding frontier: Predicting intrinsically disordered regions (IDRs) and transient states. - Inter-molecular logic: Protein-RNA interactions, ligand binding, and complex assemblies. - Protein generation and design: Models for de novo protein generation and targeted design.

000

ok55991.bsky.social @ok55991.bsky.social · 22/06/2026

We invite contributions, including, but not limited to, the following main focus areas: - Conformational landscapes: Modeling protein dynamics and Boltzmann distributions. - Genotype-to-phenotype relationships: Predicting the effect of mutations on protein structure, dynamics, and function

110

ok55991.bsky.social @ok55991.bsky.social · 22/06/2026

This workshop aims to connect ML researchers and structural biologists to discuss the current state and perspectives of structural biology. We invite submission of abstracts for an oral presentation or a poster.

100

ok55991.bsky.social @ok55991.bsky.social · 22/06/2026

We are excited to announce the Call for Abstracts for the first "Structure of Biological Molecules" Workshop taking place at the 40th German Conference for Bioinformatics (GCB) on September 22, 2026 in Saarbruecken. We depend on your contributions! old-shatterhand.github.io/workshop_sbm...

old-shatterhand.github.io

Structures of Biological Molecules @ German Conference on Bioinformatics

121

ok55991.bsky.social @ok55991.bsky.social · 27/02/2026

I totally see your point, but since we focus on VEPs I think it is fair to refer to the whole pipeline (including summing up logits) as “the method” :) It’s not meant to be criticizing LMs as the whole, but in this particular aspect, which I believe has been largely ignored

130

ok55991.bsky.social @ok55991.bsky.social · 27/02/2026

On the other hand, if PLMs had seen enough epistatic sequences, they should have learned the epistatic effects in the embeddings of individual amino acid positions. Which raises a major question: do we have enough data to train masked PLMs to learn complex biological effects?

100

ok55991.bsky.social @ok55991.bsky.social · 27/02/2026

As @delalamo.xyz rightfully notes bsky.app/profile/dela..., the poor performance of classical PLMs can be explained by how PLMs calculate variant effects: by adding individual logits for all mutations.

110

ok55991.bsky.social @ok55991.bsky.social · 27/02/2026

What works better, but still not satisfactorily, are multi-modal DL models that combine protein sequence, structure and evolution and non-DL models with specifically designed features.

100

ok55991.bsky.social @ok55991.bsky.social · 27/02/2026

We show that none of them can predict effect of multiple mutations with strong epistatic interactions, and classical PLMs fail particularly badly.

100

ok55991.bsky.social @ok55991.bsky.social · 27/02/2026

With Anastasia Kolchina, as well as Igors Dubanevics and @fkondras.bsky.social of @oistedu.bsky.social, we identified protein variants with strong epistatic effects and checked performance of 41 VEP tools including many protein language models (PLMs) based on the data from the ProteinGym benchmark.

100

ok55991.bsky.social @ok55991.bsky.social · 27/02/2026

We are happy to present a piece of analysis that I consider to have a major impact on our understanding of how good variant effect prediction (VEP) tools really are: www.biorxiv.org/content/10.6....

biorxiv.org

121

ok55991.bsky.social @ok55991.bsky.social · 27/02/2026

Unfortunately this is the state of the art and we with our paper want to raise awareness to the necessity of developing more appropriate approaches

120

ok55991.bsky.social @ok55991.bsky.social · 27/02/2026

Hi Diego! Thanks for sharing your thoughts! We also think that this is not an appropriate way to predict variant effects for epistatic sequences, and agree that this is probably the reason for the poor performance.

110

Reposted by @ok55991.bsky.social

Roman Joeres @romanjoeres.bsky.social · 24/07/2025

This week I presented DataSAIL at the #ISMBECCB2025 conference in #Liverpool It has been an amazing chance and experience to meet many people working on information leakage. And getting great ideas to extend it Nicely supervised by & colaborated with @dbblumenthal.bsky.social @ok55991.bsky.social

054

Reposted by @ok55991.bsky.social

Alexander Gress @alexandergress.bsky.social · 18/06/2025

I am very excited to share the publication of our tool StructMAn 2.0 in the NAR webserver issue (10.1093/nar/gkaf381). If you want to annotate protein structures to protein sequences or if you are looking for structural evidence for PPIs, consider using: tools.helmholtz-hips.de/structman/

031

Reposted by @ok55991.bsky.social

Roman Joeres @romanjoeres.bsky.social · 21/05/2025

DataSAIL's "press tour" stops at ISMB/ECCB in Liverpool for a talk and a poster. If you're in ML, come around and check the data splitting out. I'm looking forward to this and the keynotes by the Nobel Laureates John Jumper and David Baker! @ok55991.bsky.social @dbblumenthal.bsky.social

011

Reposted by @ok55991.bsky.social

Roman Joeres @romanjoeres.bsky.social · 09/04/2025

DataSAIL is out in @naturecomms.bsky.social Since the preprint, we have improved the work a lot, thanks to countless reviewers and feedback. You can find it here: nature.com/articles/s41... Thanks, @dbblumenthal.bsky.social and @ok55991.bsky.social, for helping and supervising me on this journey.

052