Lightnews — Scholar-powered news

Diego del Alamo

@delalamo.xyz

3.2K followers 2K following 1.8K posts

Computational protein engineering & synthetic biochemistry Opinions my own https://linktr.ee/ddelalamo

linktr.ee

Posts Media Videos Starter Packs

Pinned

Diego del Alamo @delalamo.xyz · May 15

"Adapting ProteinMPNN for antibody design without retraining"

Our several-years-old fix to ProteinMPNN's tendency to make weird antibody CDR sequences is finally preprinted

Adapting ProteinMPNN for antibody design without retraining

The neural network ProteinMPNN designs protein sequences capable of folding into predefined tertiary structures and quaternary assemblies. It has become widely used due to its high success rates when ...

www.biorxiv.org

1 14 47

Diego del Alamo @delalamo.xyz · 1d

Another one - PDB 8F2X chain H (3.5 Å resolution, shown in gold). Totally different authors. Seems like aromatics immediately before the conserved J-gene tryptophan throws off assignment in this area; 7czv has a phenylalanine before the tryptophan

Diego del Alamo @delalamo.xyz · 1d

If you've been including PDB 9DMQ chain F with ANARCI/AbNum-derived numbering in your training set for learning something like CDR design, then you'll be surprised to learn that the CDRH3 is only 7 AAs long and encompasses exclusively the region in green opig.stats.ox.ac.uk/webapps/sabd...

Diego del Alamo @delalamo.xyz · 2d

Does anyone know how IMGT numbers residues in the DE loop (residues 80-84)? I can't figure out the pattern, seems to be more or less entirely random and not following the logic of CDR1-3

Diego del Alamo @delalamo.xyz · 4d

The entire J-gene of PDB 7czv (3.3 Å, shown in salmon) seems to have fallen victim to a register/misassignment error, with the C-terminal valines (right) and tryptophan (center-left) on the wrong side of the beta sheet compared to where they would be expected to be

2 8

Diego del Alamo @delalamo.xyz · 5d

The worst offender I've found so far is PDB 5OMM chain C on Sabdab, which has many unassigned residues despite being 1.7 Å resolution. For example its C-terminal VTVSS starts at IMGT pos 115 instead of pos 124. Probably AbNum/ANARCI struggling w/ the gaps

opig.stats.ox.ac.uk/webapps/sabd...

Diego del Alamo @delalamo.xyz · 5d

I've put together a tool to renumber antibody structures using only main chain atomic coords. No sequence info is used at all. In the process I found a few numbering mistakes in Sabdab/PDB, which is numbered by sequence & thus sensitive to gaps/register errors (deleted prev post due to bugs)

GitHub - delalamo/SAbR: Structure-based antibody renumbering

Structure-based antibody renumbering. Contribute to delalamo/SAbR development by creating an account on GitHub.

github.com

1 2 15

Diego del Alamo @delalamo.xyz · 10d

I flirted with doing something like this a year or two ago - prepending a mAb-specific structure prediction NN to the AF2 template input for mAb design by hallucination - but was stopped by the fact that every Bio/ML NN is in PyTorch except AF2 which is written in freaking JAX

AI x Bio Discovery @aixbiobot.bsky.social · 11d

mBER: Controllable de novo antibody design with million-scale experimental screening [new]
Ab design via seq/struct conditioning designs VHHs. Million-scale screens show success.

1 1 8

Diego del Alamo @delalamo.xyz · 11d

One mo'

mBER: Controllable de novo antibody design with million-scale experimental screening

Recent machine learning approaches have achieved high success rates in designing protein binders that demonstrate in vitro binding to their targets. While open models for unconstrained "minibinder" de...

www.biorxiv.org

Reposted by Diego del Alamo

Nick Boyd @nboyd.bsky.social · 14d

It's a very cool result but IMO there are caveats. Inference is (mostly) slower. There is existing work on faster models (e.g. MiniFold or protenix mini), and also existing work on ensemble prediction. I doubt this works without training on AFDB, which bakes in inductive bias from triangle layers

1 1 10

Diego del Alamo @delalamo.xyz · 15d

Oh interesting, what’s the TLDR for where these mutations are located? Will read later

1 1

Diego del Alamo @delalamo.xyz · 15d

Still boggles my mind that we can leave frameworks as-is and only edit CDRs and still reliably get nM binders. Complete opposite of what I would have expected given how hard it is to graft CDRs from one FW to another (such as humanizing mouse mAbs). First RFantibody and now this

Arc Institute @arcinstitute.org · 16d

In another preprint from the @brianhie.bsky.social Lab and @synbiogaolab.bsky.social, they introduce Germinal, a generative AI system for de novo antibody design.

Germinal produces functional nanobodies in just dozens of tests, making custom antibody design more accessible than ever before.

1 3 11

Reposted by Diego del Alamo

Frank Noe @franknoe.bsky.social · 16d

Postdoc position available in the BioEmu team at @msftresearch.bsky.social AI for Science - Berlin DE or Cambridge UK. Looking for candidates with backgrounds in #MachineLearning #AI Biophysics or Bioinformatics

jobs.careers.microsoft.com/global/en/jo...

9 13

Diego del Alamo @delalamo.xyz · 16d

Apple now has a protein folding NN?...

arxiv.org/pdf/2509.18480

7 19

Diego del Alamo @delalamo.xyz · 16d

How to request your swiss criminal record:
1. Fill out a form online
2. Pay a processing fee, also online
3. Print form and put it in an envelope and MAIL it
4. Receive criminal record by email

1 2

Diego del Alamo @delalamo.xyz · 20d

It’s a bit goofy. I think RosettaFold2 is also still in peer review and RosettaFold3 is out?

Diego del Alamo @delalamo.xyz · 20d

The only thing more compelling than this preprint's science are the names of some of its authors

Alexis Verger 🧬🧫🧪 @alexis-verger.cpesr.fr · 21d

De novo design of All-atom biomolecular interactions with RFdiffusion3 www.biorxiv.org/content/10.1...

De novo Design of All-atom Biomolecular Interactions with RFdiffusion3

Deep learning has accelerated protein design, but most existing methods are restricted to generating protein backbone coordinates and often neglect interactions with other biomolecules. We present RFd...

www.biorxiv.org

Diego del Alamo @delalamo.xyz · 20d

This upcoming week is my last week at GSK. I’m going to move on to greener pastures but first a few weeks off to … close on a house! Go for some bike rides! Paint 40k models! Etc.!

Reposted by Diego del Alamo

Diego del Alamo @delalamo.xyz · 21d

Yep. Reminds me of the bar plot error bars that were just capital Ts

1 1 6

Diego del Alamo @delalamo.xyz · 21d

Yep. Reminds me of the bar plot error bars that were just capital Ts

1 1 6

Diego del Alamo @delalamo.xyz · 21d

Protein language models can predict epistasis (non-additivity of multiple point mutations) if predictions are nonlinearly transformed first www.biorxiv.org/content/10.1...

2 15

Diego del Alamo @delalamo.xyz · 22d

I posted this on the other site and will repost here: something is fishy with Fig 2F results on inverse folding performance (left). Variance is way too low. Compare that to my results preprinted earlier this year (right). Definitely causing me to scrutinize other results in this paper more closely

1 8

Diego del Alamo @delalamo.xyz · 22d

Does anyone know of an NN-based drop-in replacement for protein side chain repacking (e.g., Rosetta repack/SCWRL)? Ideally something lightweight

2 1

Reposted by Diego del Alamo

Kresten Lindorff-Larsen @lindorfflarsen.bsky.social · 23d

New review on computational design of intrinsically disordered proteins 🖥️🍝 by @giuliotesei.bsky.social @fpesce.bsky.social & 👴

doi.org/10.48550/arX...

Figure 3 from the paper with the caption: "Role of machine learning in de novo design of IDRs. (A) Machine-learning models can be trained on diverse data sources, from molecular dynamics simulations to annotations of cellular localization and protein structures from the Protein Data Bank. (B) Often implemented as neural networks using sequence-encoded features as input, these models can initially be trained on a limited region of sequence space as surrogate models. Through active learning, additional simulations are performed during the design campaign to generate new data, and the surrogate model is retrained on the expanded dataset to progressively improve its accuracy. (C) Machine-learning models have been developed to predict biophysical observables, biological annotations, and protein structures. When combined, machine-learning models can be used to identify a set of sequences that strike a trade-off between multiple design objectives, defining a Pareto front."

16 48

Diego del Alamo @delalamo.xyz · 27d

This was a great read! If it were up to me, this specific part of rule #4 would be expanded into a full Opinion piece, because proteins for structural characterization are definitely not sampled IID from uniprot/wherever

2 1 6

Reposted by Diego del Alamo

Stephanie Wankowicz @stephanieaw.bsky.social · 29d

Structural bioinformatics is incredibly powerful on its own or when paired with theory or experiment. One of the PDB's superpowers isn’t from one structure, but comparing many to uncover folds, binding sites, and subtle conformational shifts. chemrxiv.org/engage/chemr...

10 Rules for a Structural Bioinformatic Analysis

The Protein Data Bank (PDB) is one of the richest open‑source repositories in biology, housing over 277,000 macromolecular structural models alongside much of the experimental data that underpins thes...

chemrxiv.org

1 15 54