iiinova — where your digital self keeps evolving

Diego del Alamo @delalamo.xyz · 09/06/2026

It’s representations of just the CDRs. The FWR is omitted from mean pooling. I’m basically just running a different version of the “junkyard” experiment from the paper www.nature.com/articles/s41... . They should be completely different

110 1

Diego del Alamo @delalamo.xyz · 09/06/2026

The training details for VHHBERT are light, but seems like there is no clustering at all. Anyway, the finding suggests that these models all represent CDRH3 residues as a linear combination of residue type features and position features, which is incredibly underpowered

010 1

Diego del Alamo @delalamo.xyz · 09/06/2026

Despite the huge amount of training data, ESM-C is still unable to distinguish mean-pooled CDRH3 representations of real antibodies from those of antibodies w/ scrambled CDRs. Only VHHBERT seems to do this, and only for natural sequences

270 7

Diego del Alamo @delalamo.xyz · 29/03/2026

Does anyone know what paper this figure is from? The citation I have for it turned out to be wrong

120 2

Diego del Alamo @delalamo.xyz · 26/03/2026

PDB 7E5Y is a funny case of two copies of the same protein in the same asymmetric unit, one of which has a register shift error (chain C, green) while the other does not (chain H, orange). Here's conserved J-gene Trp 118 (IMGT; 108 in PDB). Trastuzumab shown in teal for reference

020 2

Diego del Alamo @delalamo.xyz · 25/03/2026

The proteina-complexa paper is the 2nd this month suggesting that high K+E content, a staple of ProteinMPNN designs, is predictive of poor expression (the bits in bio benchmark paper showed a few weeks ago). In their case, they measure poor sequence recovery in phage display

2234 23

Diego del Alamo @delalamo.xyz · 18/03/2026

This is new? Instead of relying on intermediate estimates of the denoised state, Proteina-Complexa just does the full denoising, calculates whatever it needs to, and sends the info back to continue guiding the unfinished diffusion roll-out

130 3

Diego del Alamo @delalamo.xyz · 18/03/2026

Sometimes you hit a home run, and sometimes you get to first base

010 1

Diego del Alamo @delalamo.xyz · 15/03/2026

So I can't say I've ever seen residual cross-attention before (where the final representations attend to earlier representations of the input data); is there any literature on when and where to use this?

070 7

Diego del Alamo @delalamo.xyz · 10/03/2026

One of the most interesting parts of this workflow is the "sunk cost fallacy" estimator that predicts how promising a particular mutational line of inquiry is, and whether it is worth abandoning in favor of others

020 2

Diego del Alamo @delalamo.xyz · 05/03/2026

I regret to inform you that if you are job hunting on LI and this is your profile pic then you’re ngmi

030 3

Diego del Alamo @delalamo.xyz · 03/03/2026

RIP my LinkedIn after mentioning a job opening in my group ☠️

2100 10

Diego del Alamo @delalamo.xyz · 03/03/2026

The two co-first authors of this research paper, Yang & Yang, has decided to sort their names alphabetically

041 4

Diego del Alamo @delalamo.xyz · 28/02/2026

Protenix trained an identical model with way more training data (2025 cutoff instead of 2021), demonstrating that antibody-antigen modeling, but not protein-ligand modeling, is currently data-limited (DQ SR % means % DockQ≥0.23) with this architecture

3100 10

Diego del Alamo @delalamo.xyz · 26/02/2026

First fig below is the relevant code for zero-shot fitness prediction w/ ESM in the ProteinGym repo for example. It would explain why the results in the preprint, second fig, have many masked sequence-only models near the bottom of the rankings (3/3)

000 0

Diego del Alamo @delalamo.xyz · 26/02/2026

Hot take: this "chain X[auth Y]" notation on FASTA files pulled from the PDB is needlessly confusing, adds nothing, and needed to be changed yesterday

270 7

Diego del Alamo @delalamo.xyz · 26/02/2026

This plot is quite the indictment of fine-tuned PLMs, showing how performance is entirely data-dependent and, at the upper end of performance, equally achievable with randomized model weights

0132 13

Diego del Alamo @delalamo.xyz · 22/02/2026

What are you talking about? It’s literally in the app

120 2

Diego del Alamo @delalamo.xyz · 21/02/2026

Another LinkedIn influence confidently declaring that cryo-EM and X-ray crystallography are no longer necessary because we have AI now

5230 23

Diego del Alamo @delalamo.xyz · 20/02/2026

A detailed comparison of de novo designed proteins by NMR finally puts to rest the idea that structure prediction confidence metrics have any correlation with local flexibility

0103 10

Diego del Alamo @delalamo.xyz · 19/02/2026

My reply to @anshulkundaje.bsky.social on the other site has been quoted in this @nature.com report on Isomorphic's new DDE model, which relies on private data to a degree that is probably unprecedented in the history of this field 🧪🧬

0101 10

Diego del Alamo @delalamo.xyz · 16/02/2026

The fitness landscapes of binders targeting protein surfaces that are evolutionarily predisposed to engage in PPIs is much narrower and deeper than those that target surfaces with no known interacting proteins. This might be b/c random binding events is bad, and selected against

1245 24

Diego del Alamo @delalamo.xyz · 15/02/2026

Metadynamics meets diffusion-based structure prediction: the same trick that drives enhance sampling of MD trajectories can be adapted and applied as steering terms during generation to sample diverse conformations. Seems obvious in hindsight!

0176 17

Diego del Alamo @delalamo.xyz · 07/02/2026

Whoops, forgot to include the screenshot on the other site that prompted this post: why is ByteDance, the parent company of TikTok, training protein structure prediction neural networks?

030 3

Diego del Alamo @delalamo.xyz · 27/01/2026

Training antibody language models on evolutionary trajectories inferred from ancestral sequence reconstruction and NGS data, rather than a standard masked objective, leads to improved prediction of fitness and expression

090 9

Diego del Alamo @delalamo.xyz · 24/01/2026

There's an interesting parallel between the in silico affinity maturation presented in the PPIFlow paper, which turns the antibody design problem into a motif scaffolding problem, and what the adaptive immune system does with the so-called germline-encoded GRAB motifs www.biorxiv.org/content/10.6...

162 6

Diego del Alamo @delalamo.xyz · 22/01/2026

Something connecting these two is a recent observation that larger PLMs implicitly cluster sequences at finer resolutions (higher seq ID cutoff). That means big PLMs and local-only MSAs do the same thing: they filter out distant homologs that hinder VEP /🧵 bsky.app/profile/step...

030 3

Diego del Alamo @delalamo.xyz · 22/01/2026

The authors found that MSAs built from sequences with ≥90% identity (inadvisable for non-viral proteins) actually overperform, and that the number of such sequences made available during inference is predictive of accuracy when using MSA-based ML methods

120 2

Diego del Alamo @delalamo.xyz · 22/01/2026

The second is that strategies for assembling and interpreting MSAs that are built for prokaryotic and eukaryotic proteins (cluster at ~50% sequence identity) do not work as well for viral proteins, which are, e.g., not guaranteed to show coevolutionary signatures from MSAs

121 2

Diego del Alamo @delalamo.xyz · 22/01/2026

The first is that the scaling hypothesis for sequence-only PLMs is not as dead as it seems. Variant effect prediction on proteins with few homologs in training DBs, which include a large number of viral proteins, do appear to benefit from scale. Overrepresented proteins do not

130 3

Diego del Alamo @delalamo.xyz · 22/01/2026

The Twitter algorithm is completely broken but occasionally some gold slips through

030 3

Diego del Alamo @delalamo.xyz · 21/01/2026

Anyone have access to the PPIFlow paper? It’s only hosted on Dropbox for some reason

110 1

Diego del Alamo @delalamo.xyz · 08/01/2026

Some more evidence that BioEmu et al have largely rendered clever AF2 template and MSA hacks obsolete for quick conformational sampling of soluble monomeric proteins

281 8

Diego del Alamo @delalamo.xyz · 01/01/2026

Please give me your best .pymolrc.py functions. Here's one that implements "fetch_imgt <pdb_file>" and "fetch_imgt <pdb_file>_<chain_id>" for fetching renumbered antibody chains from sabdab

090 9

Diego del Alamo @delalamo.xyz · 26/12/2025

11/ Last thing; regardless of what happens, I don't expect an improvement in Ab-Ag complex prediction or antibody loop modeling. Advances in structure-based design and structure prediction used to go hand-in-hand. Yet progress here continues to be slow www.tandfonline.com/doi/full/10.... /end

051 5

Diego del Alamo @delalamo.xyz · 26/12/2025

9/ Broadly neutralizing mAbs tend to have longer CDRH3s, sometimes longer DE loops, and often more conformational dynamism than their traditional counterparts. It's unclear how effective the current tech stack is at producing such molecules (fig from elifesciences.org/articles/33038)

141 4

Diego del Alamo @delalamo.xyz · 26/12/2025

8/ The big one we've been missing is broad binding/neutralizing capabilities. Chai showed their tech can narrow selectivity to only target antigens of interest while avoiding paralogs. But I'm not aware of de novo design intentionally moving in the opposite direction.

140 4

Diego del Alamo @delalamo.xyz · 26/12/2025

5/ We might also begin to see examples of multi-epitope targeting by a single Fv, which is already seen in naturally occurring antibodies, such as Fu2. This molecule binds two copies of SARS-CoV-2 RBD from different oligomers such that neither molecule can bind to ACE2.

151 5

Diego del Alamo @delalamo.xyz · 26/12/2025

4/ I expect antibody de novo design methods to explore binding/releasing targets in the presence of effectors. Shugai has a molecule that only binds its target when environment ATP is very high (for example, in the TME), though it was obtained via display pubmed.ncbi.nlm.nih.gov/32847940/

140 4

Diego del Alamo @delalamo.xyz · 26/12/2025

2/ The field lags standard binder design by 12-18 months. This past year, we saw nM binding achieved using diffusion methods like BoltzGen as well as those based on NN backpropogation (Germinal). I suspect further affinity gains in single-shot design are likely data-constrained

140 4

Diego del Alamo @delalamo.xyz · 19/12/2025

MFW you have an awesome idea and need to pull up a random biorxiv paper for critical details

020 2

Diego del Alamo @delalamo.xyz · 19/12/2025

Pacesa et al and PXDesign both define success as number of sequences that recover the design pose when refolded w/ structure prediction methods, normalized over wall time. So RFDiffusion's backbone design efficiency doesn't seem to be useful

100 1

Diego del Alamo @delalamo.xyz · 19/12/2025

Not sure how to reconcile A) bindcraft consistently outperforming RFDiffusion, but B) AF3-gen diffusion-based design outperforming AF3-gen hallucination-based design. My best guess is that there's something about AF2 but not AF3-gen methods that makes them amenable to backprop?

492 9

Diego del Alamo @delalamo.xyz · 16/12/2025

New latent labs report on de novo antibody design closely follows what we’ve seen from Chai and Nabla. Promising results, and I respect that they admit that half of the targets were unsuccessful with their approach www.latentlabs.com/wp-content/t...

140 4

Diego del Alamo @delalamo.xyz · 11/12/2025

Also this is a pretty awesome way to do conformational sampling

040 4

Diego del Alamo @delalamo.xyz · 11/12/2025

Folding funnels are back - now we just need to convince scientists to resume using P_near, a funnel-ness metric that deserved far more use than it received

181 8

Diego del Alamo @delalamo.xyz · 17/11/2025

Anyway, if anyone in academia has some spare time & compute, I'd love to see if the MD-trained NNs are any better than the vanilla models at multi-state prediction, as it wasn't tested in the paper (model weights have a non-commercial license so they're off-limits to me)

010 1

Diego del Alamo @delalamo.xyz · 17/11/2025

A randomly valuable thing that the simplefold authors did was fine-tune all six structure prediction models on MD-ATLAS data, and honestly I can't tell if there's any benefit to scaling these things if the goal is ensemble prediction

150 5