iiinova — where your digital self keeps evolving

I've had the pleasure of working on this in collaboration with my very sharp colleagues at Absci. We look forward to feedback on all aspects of our design pipeline and preprint! [9/n, n=9]

140

This solves the false-negative problem completely! And you can use fancy asymmetric versions of ipTM @rolanddunbrack.bsky.social. Interestingly we find that, of the 5 AFM-v2.3 models, model 2 is almost always the best on ab-ag complexes. We can cut runtime by 5x with little performance hit! [8/n]

131

It turns out that masking template AAs in "AF2Rank-Unmasked" may not be the optimal choice for ab-ag complexes. This is a much trickier modality for AFM—it is systematically less capable & confident there than on many PPIs seen in protein design papers. So, we restore the template AAs! [7/n]

120

Unmasking AlphaFold to integrate experiments and predictions in multimeric complexes - Nature Communications

The hack involves "unmasking" cross-chain template information. It's not how AFM was trained, but Mirabello et al showed that AFM can leverage the newly-unmasked information (www.nature.com/articles/s41...). However, we find that "AF2Rank-Unmasked" suffers from similar issues as AF-IG. Why? [6/n]

nature.com

Integrating AlphaFold (AF) predictions with experimental data is not straightforward. Here, authors introduce AF_unmasked, a tool to integrate AF with experimental information to predict large or chal...

AF-IG is serving the protein design community well, but it falls short on antibody-antigen complexes that are OOD to the training set—it throws away too many correct PDB structures. Sth else? With a fun AF-Multimer hack, one can run AF2Rank on complexes (this is available on ColabDesign!) [5/n]

120

Improving de novo protein binder design with deep learning - Nature Communications

A popular way to score/rank designed binders with AF is "AF Initial Guess" (AF-IG), introduced by the Baker lab. Here, the designed complex is fed as an initialization to AF's trunk (this is possible because of recycling). Both AF-IG and AF2Rank are ways to provide AF with a structural hint. [4/n]

nature.com

Recently, a pipeline for the design of protein-binding proteins using only the structure of the target protein was reported. Here, the authors report that the incorporation of deep learning methods in...

AF2Rank worked even better when template AA tokens were replaced with gap symbols. Otherwise, AF2 thinks you are giving it "the right answer" and, in the case of protein monomers, spits the decoy structure back to you over-confidently. Why is this potentially useful to binder discrimination? [3/n]

Umberto Lupo @umbislupo.bsky.social · 14/01/2026

This was fun work and a remarkable effort across the computational and wet-lab teams! Strategies for in-silico filtering and ranking of antibody designs have been under-discussed in the literature, e.g. in most technical reports on antibody design that I've seen. Let's talk about these here! [1/n]

1188

At Absci, we performed de novo antibody design campaigns against "zero-prior" epitopes—lacking structural data from antibody-antigen or protein-protein complexes. Model architectures, training data curation, and scoring protocols are fully described. Preprint: www.absci.com/wp-content/u...

Carl Quintanilla @carlquintanilla.bsky.social · 09/01/2026

Reposted by Umberto Lupo

GERMAN PRESIDENT STEINMEIER: “.. the United States has broken with the values that it helped to establish .. “.. we have now moved beyond the stage where we can lament the lack of respect for international law or the erosion of the international order; we are far beyond that, I believe.”

358120854461

Umberto Lupo @umbislupo.bsky.social · 07/01/2026

Recently found out about @jvkersch.bsky.social's `tmtools` github.com/jvkersch/tmt.... It works nicely! Being able to pass user-defined sequence alignments is a nice (if simple) feature that is missing from OpenStructure's own `tmtools` @torstenschwede.bsky.social.

github.com

GitHub - jvkersch/tmtools: Python bindings for the TM-align algorithm and code for protein structure comparison developed by Zhang et al.

Python bindings for the TM-align algorithm and code for protein structure comparison developed by Zhang et al. - jvkersch/tmtools

121

Reposted by Umberto Lupo

Edward Vogel @mathartforall.bsky.social · 06/01/2026

Black swan moment for energy production and storage?

032

Umberto Lupo @umbislupo.bsky.social · 17/12/2025

Thank you for your reply! I see, so the complete statement is something like "this is the first time that the reversible folding of a protein is computed with an all-atom foundation machine learning model trained on DFT-generated data", correct?

Umberto Lupo @umbislupo.bsky.social · 17/12/2025

Exciting! Concerning your chignolin example, you write: "Previous studies tackling this system with a ML model were limited to non-transferable protein models [...] this is the first time that the reversible folding of a protein is computed with a foundation machine learning model"

Umberto Lupo @umbislupo.bsky.social · 29/10/2025

🏛️ 🔍 👀 "There are no optical corrections in the Parthenon"

Kresten Lindorff-Larsen @lindorfflarsen.bsky.social · 02/10/2025

Reposted by Umberto Lupo

Visiting Ephesus at #embointegmod25 to look for the missing AlphaFold 3 code

0182

Reposted by Umberto Lupo

Lenka Zdeborová @zdeborova.bsky.social · 30/09/2025

Would you like to come to my lab for a postdoctoral position? Apply for the EPFL AI Centre fellowship: epfl.ch/research/fun.... FYI, I have funding (ERC & Simon's collaboration) also for 1-2 top candidates who will not secure the fellowship. Same for @krzakalaf.bsky.social lab.

epfl.ch

EPFL AI Center and Swiss AI Initiative Postdoctoral Fellowships

The 2nd call is now open with a deadline for submissions of 3 November (17.00 CET)!Applications are encouraged from researchers at the postdoctoral level with a keen interest in collaborative, interdi...

031

Umberto Lupo @umbislupo.bsky.social · 16/09/2025

Very neat video, and equally neat housekeeping :)

Patrick Kidger @patrickkidger.bsky.social · 12/09/2025

Reposted by Umberto Lupo

💥 We are *also* organizing Machine Learning for Structural Biology @ EurIPS, Copenhagen!! Topics include anything in the ML+bio intersection. Submit your ML+bio short papers! Authors can even present in both locations if they have people in both locations 🚀

0172

Umberto Lupo @umbislupo.bsky.social · 04/09/2025

Cool-looking work by @piompons.bsky.social @albecazzaniga.bsky.social @fra-cutu.bsky.social, going beyond existing works on sampling different conformational ensembles with AF2.

020

Umberto Lupo @umbislupo.bsky.social · 02/09/2025

Is there any way, any way at all, that this can be followed remotely? (jk, but if you have a link...)

100

Reposted by Umberto Lupo

Nature @nature.com · 29/08/2025

Posts about research on Bluesky receive substantially more attention than similar posts on X, formerly called Twitter go.nature.com/45Ftiw4

go.nature.com

Research posts on Bluesky are more original — and get better engagement

Bluesky posts about science garner more likes and reposts than similar ones on X.

2294104

Umberto Lupo @umbislupo.bsky.social · 29/08/2025

Two exciting openings with us! 🤖🧬🆎🧫💉 - AI Scientist 👉 lnkd.in/eDXHH4E8 - AI Scientist, Drug Creation 👉 lnkd.in/eEvGyaTR You'll work on antibody sequence/structure design, antibody-antigen co-folding, antibody-antigen binding prediction, physics-based methodologies, and more! DMs welcome!

032

Reposted by Umberto Lupo

Damiano Sgarbossa @damianosg.bsky.social · 21/08/2025

🎉 Excited to share that the last paper of my PhD is now published in PRX Life! We introduce RAG-ESM, a retrieval-augmented framework that makes pretrained protein language models (like ESM2) homology-aware with minimal training cost. 📄 Paper: journals.aps.org/prxlife/abst...

072

Umberto Lupo @umbislupo.bsky.social · 21/08/2025

@cyrilmalbranke.bsky.social is 🦋! "ProteomeLM significantly outperforms DCA in recovering experimentally validated interactions. In H. sapiens, ProteomeLM achieves an AUC of 0.83, compared to 0.73 for DCA. Among the top 10 million scored pairs, it recovers 50% of known PPI, versus only 20% for DCA"

Umberto Lupo @umbislupo.bsky.social · 21/08/2025

Great to see this long-form version of your ICLR 2024 work finally out!