henhaobio
Sign in

bioRxivBioengineeringPreprint

Latent generative search unlocks de novo design of untapped biomolecular interactions at scale

Kieran DidiiD, Danny ReidenbachiD, Matthew PenneriD, Supriya RavichandraniD, Marshall Case, Mike Nichols, Erik Swanson, Alex Reis, Maggie Prescott, Yue Hong QianiD, Dongming QianiD, J.S. YangiD, Weiji LiiD, Le LiiD, Daichi ShonaiiD, Sean M. GayiD, Bhoomika Basu MallikiD, Ho Yeung ChimiD, Liurong CheniD, Miguel Atienza JuanateyiD, Hubert Klein, D. L. RIEGERiD, Phillip SchlegeliD, Anna U. Macintyre, Maxim SecoriD, , Karsten KreisiD

PDFDOICode

In-depth summary

Claude Opus 5.5 · xhigh effort · 8 Nov 2026 · from the full text of version 1 (18 Sept 2026) · CC BY 4.0

Reward-guided inference-time search over the Proteína-Complexa sequence-structure codesign model yields validated binders across 127 targets, kinases and a free carbohydrate, with codesigned sequences outperforming ProteinMPNN redesign.

Binder design against polar, solvent-exposed epitopes and small flexible ligands such as carbohydrates is hard for current methods, which also depend on a separate inverse-folding step. The paper pairs a learned generative prior with inference-time reward search and tests it experimentally at scale, including a screen of more than one million designs.

Key results

  1. ShownIn a compute-matched phage-display comparison on 83 targets, codesigned Proteína-Complexa sequences gave the highest specific on-target hit rate and beat post hoc redesign of the same backbones.

    Specific hit rate 1.87% vs 0.60% (BoltzGen native) and 1.44% (BoltzGen with ProteinMPNN). On Proteína-Complexa backbones: 702 on-target hits (607 specific) codesigned vs 323 (275) with ProteinMPNN and 333 (289) with fixed-interface redesign. · 5 Benchmarking design methods at scale; Massive-scale benchmark; S4.1 Large-scale design campaign; S4.4 Model temperature ablation

  2. ShownAcross a 127-target panel, the framework gave on-target hits for most targets, but off-target binding was also widespread.

    492,636 Proteína-Complexa sequences screened. 115 of 127 targets (91%) received one or more on-target hits; 92 targets (72%) had target-specific binders. 1,982 of 23,966 active binders (8.3%) engaged the intended target, a 10.5-fold enrichment over the 0.79% expected. Each panel target was also bound specifically by some binder designed for a different target. · 5 Benchmarking design methods at scale; Massive-scale benchmark

  3. ShownDirect binder campaigns on PDGFR, PD-L1 and the Nipah virus attachment glycoprotein yielded binders with a wide range of hit rates, including picomolar PDGFR binders and nanomolar binders from scaffold re-engineering.

    192 designs per target tested by surface plasmon resonance (SPR). Hit rates 63.5% (PDGFR, KD 93.6 pM to 1.34 µM) and 10.4% (PD-L1, KD 108 nM to 3.71 µM). Nipah: 20 designs tested; the de novo set gave one binder (KD = 56 nM), and 5 of the 6 re-engineered designs bound with nanomolar affinity. · 4 Protein binders; S3.2 Hotspot conditioning analysis

  4. ShownDesigns against the ActRIIA receptor bound with nanomolar affinity and blocked signalling in a cell reporter assay without affinity maturation.

    Of 200 designs, 192 expressed; SPR found 16 binders, 8 sub-micromolar, tightest KD = 36 nM. Myostatin-induced Smad2/3 luciferase signalling in HEK293 cells was inhibited with IC50 from 169 to 1,532 nM. · 6 Binders to a muscle-wasting receptor; ActRIIA; Figure legends

  5. SuggestedDe novo proteins bound a free blood-group A pentasaccharide with weak affinity, and one design (design 3) agglutinated type A2 but not type B red cells.

    23 designs tested by fluorescence polarization; 15 exceeded the +10 mP threshold. Five well-constrained KD values of about 12 to 340 µM; fitted KD for five representative designs 10.1 to 814.3 µM (design 3 extrapolated). Design 3 had no sequence homologue and no Foldseek hit at E < 10^-2. · 8 De novo carbohydrate-antigen binders; Carbohydrate binders; S7.3 Red blood cell agglutination assay

Methods, data and reproducibility

Proteína-Complexa generates binder backbone, side chains and sequence jointly through a variational-autoencoder latent space and a flow-matching model conditioned on the target. Inference-time search (beam search, Feynman-Kac steering, Monte Carlo tree search) with multi-term rewards steers sampling, and designs are tested by phage display, SPR, cell assays and biophysics.

  • Designed monomers (100-800 aa), codesigned vs ProteinMPNN sequences; crystal structure of one 800-residue design · n = 48 designs · PDB 32GJ
  • Multiplexed phage-display benchmark of designed binders against a commercial target panel · n = 1,015,870 sequences, 127 targets
  • PDGFR and PD-L1 binder designs tested by SPR · n = 192 designs per target
  • Nipah virus attachment glycoprotein binders (de novo and re-engineered) · virus · n = 20 designs
  • ActRIIA mini-protein binders tested by SPR and a Smad2/3 reporter assay in HEK293 cells · n = 200 designs (192 expressed)
  • PAK1 mini-proteins (split-NeoR screen) and CK1δ peptides (pull-down) · n = 50 PAK1 designs; 18 CK1δ peptides
  • Type II A pentasaccharide binders tested with red blood cell agglutination · human (commercial type A2 and B red blood cells) · n = 23 designs
Methods & tools
Proteína-Complexa, La-Proteína, AlphaFold2, Boltz-2, RosettaFold3, ProteinMPNN, BoltzGen, BindCraft, RFdiffusion3, Tmol, Foldseek, multiplexed phage display
Software & environment
Default search settings: beam width 4, branching factor 4, step length 100 over 400 denoising steps; Feynman-Kac inverse temperature 10. AlphaFold2 Multimer v2.3.2 used for PAK1 complex predictions.
Assay conditions & compute
More than 140,000 GPU hours for generation and filtering across the campaign. Baselines were given about 32 GPU hours per target; most Proteína-Complexa tasks finished in 20 to 32 GPU hours.
Reproducibility gaps
Plasmids and designed proteins only on reasonable request. The text gives no link to the phage-display screen data or the designed sequences, beyond the PDB deposit for the crystal structure.
Limitations (6)
  • The advantage of codesign over sequence redesign holds on average but differs from target to target, and the authors say it is unclear which interface properties gain most from joint optimisation. · stated by the authors
  • Off-target binding was common: for every target in the panel, some sequence designed against a different target showed specific binding to it. · stated by the authors
  • Phage input pool recovered only 73.2% of about 1.4 million designs; longer sequences and lysine/glutamate-rich sequences dropped out, so the screened set is compositionally biased. · stated by the authors
  • The BLI antigen-specificity screen of carbohydrate binders was run in crude lysate and is described as supporting, not definitive, evidence. Design 15 also bound fucose, shared by the A, B and O antigens. · stated by the authors
  • The benchmark was designed and run by the method's developers with their own baseline implementations and selection rules, and hits were called from phage display, not from affinity measurements. · noted in this summary
  • Carbohydrate affinities are weak (tens to hundreds of µM, one value extrapolated), and blood-group discrimination was shown for a single design in one agglutination format. · noted in this summary

Read this if you design protein or peptide binders and want large-scale experimental evidence on whether sequence-structure codesign with inference-time search beats inverse-folding pipelines.

Citations
4

Abstract

De novo protein design has advanced rapidly, yet designing binders to polar, solvent-exposed epitopes and small, flexible ligands remains challenging. Such hydrated surfaces and flexible molecules, including carbohydrates, provide few of the hydrophobic contacts favoured by current methods and have largely resisted de novo binders. To address this challenge, here we introduce latent generative search for binder design, a novel framework that uses reward-guided search at inference time to steer the Proteina-Complexa generative model. The model codesigns sequence and structure - generating them together in a continuous latent space - and thereby removes the inverse-folding step on which current methods rely. In a screen of more than one million designs by multiplexed phage display, latent generative search produced more validated binders than every other method tested, its codesigned sequences surpassing post hoc redesign. It delivered high-affinity binders across therapeutic receptors, a viral attachment protein and intracellular signalling targets. Our approach also accessed previously untapped biology, generating the first de novo proteins that bind a free carbohydrate, including one that discriminates between blood-group antigens - a polar, flexible target class beyond the reach of current design methods.

Reproducibility ledger

Nobody has reported trying to reproduce this paper yet. Did you replicate an experiment or run the code? Your report saves the next lab a week.

Sign in to file a reproducibility report.

Related papers

Similarity Map
SelectedSame fieldCross-field
THISCao '21Lourenço '24Team '25Stark '25Sehgal '26Eguchi '22
72% semantic matchSynthetic Biology
Longxing Cao, Brian Coventry, Inna Goreshnik+28 more·bioRxiv·Sep 4, 2021· 12 citations