Tessera
Folgen

David B. Blumenthal

@dbblumenthal.bsky.social

Professor for Biomedical Network Science at FAU Erlangen-Nürnberg (https://bionets.tf.fau.de). Opinions are my own.

2 Posts52 Follower56 Folgt
David B. Blumenthal@dbblumenthal.bsky.social · last yr.

Very happy to see DataSAIL published in @naturecomms.bsky.social. Give it a try if you want to test if your ML models generalize to OOD scenarios. Great collaboration between @uni-saarland.de and @fau.de :-)

Roman Joeres@romanjoeres.bsky.social · last yr.

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.

FAU Erlangen-Nbg@fau.de · 2y ago

Are you ready to challenge yourself and compete with the brightest minds in AI and computer science? 🧠 Join the FAU AI Innovation Challenge 2025! Compete for 10.000€ in prizes in various categories ranging from game AI, cybersecurity & more. 👉 Learn more: go.fau.de/1beg-

A graphic depicting people sitting at desks and working on computers in the foreground. In the background, there are more people standing around in an abstract space. The space itself looks futuristic.
David B. Blumenthal@dbblumenthal.bsky.social · 2y ago

Deep learning models for sequence-based PPI prediction still fail to yield reliable predictions in challenging scenarios. Great work led by the amazing Timo Reim and @judith-bernett.bsky.social

Judith Bernett@judith-bernett.bsky.social · 2y ago

🧬🖥️ Proud to share our latest update on PPI predictions – "Deep learning models for unbiased sequence-based PPI prediction plateau at an accuracy of 0.65" doi.org/10.1101/2025... by T. Reim, published with @itisalist.bsky.social @dbblumenthal.bsky.social, A. Hartebrodt, and me. What did we do? 1/15 🧵

Graphical summary of the analyses done in the publication displayed on six panels a-f. (a) We computed ESM-2 embeddings of different sizes for the proteins of our data-leakage-free PPI dataset. The per-token embeddings have variable sizes depending on the protein length, while the per-protein embeddings have a fixed size by applying dimension-wise averaging. (b) We tested two models operating on the per-protein embeddings—a baseline random forest classifier and adaptions of the previously published Richoux model. Five models operated on the per-token embeddings: a 2d-baseline, the 2d-Selfattention and 2d-Crossattention models (which expanded the 2d-baseline through a Transformer encoder), and adaptations of the published models D-SCRIPT and TUnA. (c) Hyperparameter tuning gave us insight into the influence of each tunable parameter on the classification performance. (d) No model surpassed an accuracy of 0.65. The more advanced models had similar accuracies, leading us to believe that the information content of the ESM-2 embedding has more influence than the model architecture. Per-token models did not consistently outperform per-protein models. (e) We applied various modifications to test their influence: different embedding sizes, inserting a Transformer encoder into different positions, adding spectral normalization after the linear layers, self- vs. cross-attention, and removing the padding. (f) Finally, we compared the implicitly predicted distance maps of the 2d-baseline, 2d-Selfattention, 2dCrossattention, and D-SCRIPT-ESM-2 to real distance maps computed from PDB structures.