iiinova — where your digital self keeps evolving

Kieran Didi @kdidi.bsky.social · 18/09/2026

With colleagues at Cambridge, we tested our carbohydrate binders in cell-agglutination assays, closer to their intended use case. Several binders showed specificity for A antigen, B antigen, or both - all of them useful behavior for the application. 3/n

110

Kieran Didi @kdidi.bsky.social · 18/09/2026

Sampling temperature gives us a useful dial: low temperature produces stable codesigns, higher temperature yields native-like structures. With collaborators in Munich/Leipzig, we pushed this to an 800-aa codesign—and obtained a crystal structure that closely matches the design! 2/n

130

Kieran Didi @kdidi.bsky.social · 17/03/2026

Case study 5/5: first fully de novo Carbohydrate binders We target blood-group B antigen and got a a 21% hit rate. Massive shoutout to my long-term enzyme collaborator @mattpenner.bsky.social at Cambridge for powering this through, excited about what codesign will unlock for enzymes! 17/n

110

Kieran Didi @kdidi.bsky.social · 17/03/2026

Case study 4/5: Nipah virus In the Adaptyv binder competition setting, Proteina-Complexa produced a nanomolar binder (56 nM) to NiV-G, targeting a recessed receptor-binding site. We also show how to go beyond partial diffusion and perform joint sequence-structure redesign 16/n

100

Kieran Didi @kdidi.bsky.social · 17/03/2026

Case study 3/5: Targeting Kinases (PAK1 + CK1δ) with peptide binders Across mini-protein and short-peptide regimes (<31 aa peptides and 49–74 aa mini-binders), we observed strong hit rates on difficult target sites. 15/n

100

Kieran Didi @kdidi.bsky.social · 17/03/2026

Case study 2/5: ActRIIA, a target for GLP-1 associated muscle wasting We designed de novo binders that block myostatin signaling in cells. Tightest binder reached KD = 36 nM, with functional downstream inhibition. 14/n

110

Kieran Didi @kdidi.bsky.social · 17/03/2026

Case study 1/5: PDGFR We achieved a 63.5% hit rate, with top binders reaching double-digit picomolar affinity (best reported at 93.6 pM). Very cool collab with @novonordisk.bsky.social 13/n

100

Kieran Didi @kdidi.bsky.social · 17/03/2026

As part of this campaign, we also ran a the biggest head-to-head wetlab comparison of design methods ever. Proteina-Complexa outperformed baselines in this setting. My personal highlight: we are the only method where codesign shines, outperforming MPNN for the first time! 12/n

100

Kieran Didi @kdidi.bsky.social · 17/03/2026

Massive campaign: ~1M designed candidates screened across 127 diverse/challenging targets with all-to-all binding readouts. We solve more than 2/3 of the targets given a very limited compute budget and often quite challenging hotspots and crops 11/n

100

Kieran Didi @kdidi.bsky.social · 17/03/2026

That’s the method side (come to ICLR in Rio to hear more about at our Oral!). But in silico only goes so far, so now the wet-lab side! This was a major collaboration with @manifoldbio.bsky.social , @vivabiotech, @novonordisk.bsky.social , @CambridgeUniversity, @DukeUniversity, @lmu.de . 10/n

100

Kieran Didi @kdidi.bsky.social · 17/03/2026

Quantitatively, these inference-time scaling strategies outperform prior hallucination-based methods under normalized compute budgets as well as other generative models, setting a new SOTA in in-silico binder design. 9/n

100

Kieran Didi @kdidi.bsky.social · 17/03/2026

One of my highlights: because search is reward-guided, we can optimize for biophysical objectives during generation - including interface hydrogen-bond terms that promote denser interaction networks. With MLFF improvements being rapid, this becomes more and more powerful! 8/n

100

Kieran Didi @kdidi.bsky.social · 17/03/2026

Data was also key: with @martinsteinegger.bsky.social and @sooyoung-cha.bsky.social lab we introduce Teddymer, a large synthetic binder-target pretraining resource built from domain-domain interactions derived from AFDB monomers (details + links on the project page). 7/n

120

Kieran Didi @kdidi.bsky.social · 17/03/2026

In practice, this means scaling inference-time compute using strategies like beam search, MCTS, and Feynman–Kac steering to improve candidate quality and physical plausibility. Force field metrics, interaction energies, folding scores, you choose you reward! 6/n

120

Kieran Didi @kdidi.bsky.social · 17/03/2026

We then combine the strengths of generative models with optimisation methods like hallucination. We call the inference recipe latent generative search: combine a learned generative prior with search at test time to steer toward better binders. 5/n

100

Kieran Didi @kdidi.bsky.social · 17/03/2026

Core design choice: - joint sequence-structure generation in a partially latent atomistic framework (building on La-Proteina). - No discrete sequence tokenization loop. - No mandatory post hoc inverse-folding redesign step, the first codesign model that outperforms MPNN! 4/n

100

Kieran Didi @kdidi.bsky.social · 17/03/2026

Proteina-Complexa is a generative binder design framework that supports diverse targets: single-chain proteins, multi-chain complexes, and small-molecule binding contexts. 3/n

100

Kieran Didi @kdidi.bsky.social · 17/03/2026

Two papers, one story: 1) ICLR 2026 Oral: method + core modeling advances 2) Experimental validation: large-scale wet-lab evidence across many targets and campaigns Let’s start with the method paper. 2/n

101

Kieran Didi @kdidi.bsky.social · 17/03/2026

📢 We’re launching Proteina-Complexa — and after the Jensen keynote mention, we definitely had to post this thread now ;) Atomistic binder design with generative pretraining + test-time compute, plus large-scale wet-lab validation. Project page: research.nvidia.com/labs/genair/... 🧵 1/n

13716

Kieran Didi @kdidi.bsky.social · 15/08/2025

The preprint shows how atomworks leads to better reference conformers (and better predictions!), enables advanced features in RF3 like chirality-aware training or ligand templating and narrows the performance gap to closed-source models. 5/6

100

Kieran Didi @kdidi.bsky.social · 15/08/2025

`atomworks.ml` on the other hand offers advanced dataset featurization and sampling for deep learning workflows, all operating on the canonical AtomArray object from @biotite_python so that all transforms are traceable and generalizable between models. 4/6

121

Kieran Didi @kdidi.bsky.social · 15/08/2025

AtomWorks has two main components: atomworks.io takes a file (cif, sdf, ...) and does parsing, cleaning and more. You can also look at your structures in a notebook or via PyMol thanks to pymol-remote, so you can directly inspect if your code does what you want! 3/6

111

Kieran Didi @kdidi.bsky.social · 15/08/2025

In the past, every BioML model had its own data pipeline, creating loads of overhead. With AtomWorks, >80% of code is shared between models like ProteinMPNN, RF3 or design models. 2/6

100

Kieran Didi @kdidi.bsky.social · 15/08/2025

AtomWorks is out! Building upon @biotite_python, we built a toolkit for all things biomolecules and trained RF3 with it. All open-source, test it via `pip install atomworks`! AtomWorks: github.com/RosettaCommo... RF3: github.com/RosettaCommo... Paper: tinyurl.com/y2w4z65b 1/6

1215

Kieran Didi @kdidi.bsky.social · 23/04/2025

Excited to present my first paper officially as a PhD student now as an ICLR Oral this week! Super fun work with the GenAIR team at NVIDIA. Talk: Fr 10:54 - 11:06 (Oral Session 3B, Garnet 213-215) Poster: Fr 15:00-17:30 (Hall 3 + Hall 2B #5) Come by the poster/reach out to chat

0122