Skip to content

Proteina-Complexa — Didi, Zhang, Kreis et al. (NVIDIA + Mila + Oxford + SNU)

One-line description — A fully atomistic protein-binder design method unifying flow-based generative pretraining (built on La-Proteina) with inference-time test-time-compute optimisation, achieving up to 63.5 % in-silico hit rate against PDGFR plus nanomolar binders to Nipah-virus glycoprotein and a myostatin-pathway receptor.

  • Authors: Kieran Didi, Zuobai Zhang, Guoqing Zhou, Danny Reidenbach, Zhonglin Cao, Sooyoung Cha, Tomas Geffner, Christian Dallago, Jian Tang, Michael M. Bronstein, Martin Steinegger, Emine Kucukbenli, Arash Vahdat, Karsten Kreis
  • Affiliation(s): NVIDIA; Mila / Université de Montréal / HEC Montréal; University of Oxford; Seoul National University
  • Track / workshop: ICLR 2026 main track — Oral / Poster (Sat Apr 25, Pavilion 3 P3-#1718; Oral session)
  • OpenReview: forum?id=qmCpJtFZra
  • Project page: research.nvidia.com/labs/genair/proteina-complexa
  • Code: github.com/NVIDIA-Digital-Bio/proteina-complexa
  • arXiv: 2603.27950
  • Status at ICLR 2026: Main-track Oral

What it does

Proteina-Complexa generates fully atomistic protein binders against protein and small-molecule targets. It unifies two design paradigms that were previously separate: (a) generative flow-matching models for protein structure, and (b) hallucination / inference-time optimisation against a structure-prediction oracle. The result is a single pipeline that emits binders with both high generative diversity and high computed hit rate.

How it works

Core idea. Combine a flow-matching atomistic generative model (La-Proteina) with binder-conditioning and test-time-compute optimisation, so the same pipeline emits both diverse generative samples and high-hit-rate binders refined at inference.

Inputs / outputs. Inputs are a target structure (protein receptor, small molecule, or carbohydrate) plus optional conditioning (fold class, hot-spot residues, hydrogen-bond constraints); outputs are fully atomistic binder designs ready for downstream wet-lab validation.

Key innovation. Prior generative binder methods split between (a) RFdiffusion-class structure-conditioned generators (trained on experimental complexes — data-bottlenecked) and (b) hallucination loops against AlphaFold/Boltz (compute-heavy, single-target). Proteina-Complexa unifies the two: pretraining on Teddymer, a synthetic dataset built from domain-domain interactions of computationally predicted monomers (sidesteps the experimental-complex bottleneck), plus a novel latent generative search at inference that beats classical hallucination loops under matched compute.

Parameters / training details. Backbone: La-Proteina flow-matching architecture extended with a binder-conditioning track. Pretraining corpus: Teddymer synthetic domain-domain dataset. Test-time-compute: latent generative search normalised against fixed compute budgets. Public github release + pretrained checkpoints confirmed in NVIDIA dev-blog companion.

Canonical experiment. On the standard PDGFR generative-binder benchmark, Proteina-Complexa reports a 63.5 % in-silico hit rate — state of the art. Wet-lab follow-up produced nanomolar binders against Nipah-virus glycoprotein and a myostatin-pathway receptor, plus the first de novo carbohydrate binders, generalising the method beyond protein-protein interfaces.

Headline benchmarks

  • 63.5 % hit rate against PDGFR (in-silico) — state of the art on the standard generative-binder benchmark suite.
  • Nanomolar binders against Nipah virus glycoprotein.
  • Nanomolar binders against a muscle-wasting receptor blocking myostatin signalling in cells.
  • First de novo carbohydrate binders demonstrated.
  • Markedly higher in-silico success rate than prior generative binder methods; novel test-time strategies beat prior hallucination methods under matched compute.

Where it fits in the corpus

  • AACR 2026: drug-discovery axis — pair with any RFdiffusion / Boltz-class binder dossiers in the AACR vault. PDGFR is the explicit cancer-target hook (PDGFR signalling is a known oncology pathway).
  • GEM workshop sibling: the GEM workshop's RBX-1 binder competition (RBX-1 = cancer-relevant E3 ligase) is the natural follow-on benchmark.
  • NeurIPS 2026 MLSB / ICML 2026 GenBio: expect an extended version with wet-lab follow-up.
  • MICCAI 2026 / AACR 2026 wet-lab confirmation: Bronstein / Steinegger / NVIDIA likely to ship a follow-on at SITC or AACR with experimental validation of the oncology targets.

Notes

This is the headline biomolecular-design paper of ICLR 2026 — the NVIDIA + Mila + Oxford + SNU consortium is the same team that shipped La-Proteina in 2025; Complexa is the binder-conditioned successor. The "test-time-compute" framing — scaling sampling at inference rather than at training — is shared with the broader 2025–2026 reasoning-LLM trend and is one of the methodological bridges between ICLR's LLM and bio tracks. NVIDIA's developer-blog companion piece confirms the public github release and pretrained checkpoints.