henhaobio
Sign in

bioRxivSynthetic BiologyPreprint

The Human Bindome: A Proteome-scale Atlas of Designed Binder Candidates

Julius WencksterniD, Anna M. Díaz-RoviraiD, Julia Kuhn, Arvid Ban, Rahma HamdaniiD, Roser Pruano-Milla, Evgenia ElizarovaiD, Sandrine GeorgeoniD, Kaiden ThompsoniD, Matthias HinterndorferiD, Devanarayanan Siva SankariD, Maximilian DünnebackeiD, David Desscan, Sreenath Sasidharan NairiD, Marcelo Querino Lima AfonsoiD, Jennifer FlemingiD, Sameer S VelankariD, Andrea AblasseriD, Paola PicottiiD, Georg E. WinteriD, Mikko TaipaleiD, Bruno E. CorreiaiD

PDFDOICodeData
In-depth full-text digest available · Jump to digest ↓

Abstract

Affinity reagents such as antibodies are indispensable for interrogating proteins biological function. Yet they are costly and frequently unreliable, with unknown sequences, posing challenges to reproducible experimental research. Deep learning-based protein design can now in silico generate affinity reagents achieving reliable experimental success rates, but has remained largely confined to specialist laboratories. Here we present the Human Bindome, a proteome-scale atlas of high-confidence in silico protein binder candidates. By embedding the experimentally benchmarked BindCraft method in an accelerated, parallelized framework with automated domain-level target selection, we generated 306,146 binder candidates covering 8,296 human proteins (40.9% of the full proteome). Every candidate carries a defined sequence, a predicted binder-target structure model, and in silico confidence metrics. We characterize proteome-wide coverage and show that binder epitopes frequently overlap functional sites. This positions the Bindome as a resource of genetically encodable perturbagens for site-specific, modular control of protein function. The Bindome is freely available through a web interface (https://bindome.epfl.ch), with agentic, natural-language querying and as data splits for machine-learning model development. We anticipate that the Bindome will be valuable for the scientific community by providing affinity and perturbation reagents with broad applications in dissecting biological mechanisms as well as in drug and target discovery.

In-depth summary

Claude Sonnet 4.6 · xhigh effort · 8 Nov 2026 · from the full text of version 1 (30 Jul 2026) · CC BY 4.0

A parallelized BindCraft pipeline generated 306,146 in silico binder candidates against 8,296 human proteins, each with sequence, predicted complex and confidence metrics, freely accessible online.

Existing affinity-reagent resources either lack sequences and in-depth characterization or cover few proteins (SAbDab 493 targets, ABCD 1,409), and design tools require expertise and compute. The Bindome precomputes sequence-defined, structure-modelled binder candidates at proteome scale, but none are yet experimentally validated.

Key results

  1. ShownThe Bindome covers 77.8% of the targetable human proteome. It is the second-largest resource by target count among those compared, and the largest source of binder candidates with open sequences and models.

    306,146 candidates; 8,296 targets (77.8% of targetable, 40.9% of all human proteins) of 20,279 human target proteins. Antibodypedia 18,640 (91.9%), ABCD 1,409 (6.9%), Biocompare 579 (2.9%), SAbDab 493 (2.4%). Median pLDDT 0.92, median iPAE 0.23. · In silico confidence of the binder candidates; Proteome-scale target coverage

  2. ShownStructural-superfamily coverage is high, but per-protein epitope coverage is sparse: binder contacts cover only 13.9% of total surface area, although 64.5% of the modelled surface is targetable.

    78.2% of 1,055 CATH superfamilies contain at least one domain with a binder; 5.4% are targetable but uncovered. 64.5% of the modelled surface is targetable, yet binder contacts cover 13.9% of total surface area. · Proteome-scale target coverage

  3. ShownBinder epitopes are more evolutionarily conserved than residue-count-matched surface controls, and conservation rises as more binders converge on the same residues.

    At each of three consensus levels (any binder, ≥30%, ≥90% of a target's binders), a paired two-sided Wilcoxon signed-rank comparison by domain gave P < 1×10^-300 against the matched control, using Rate4Site evolutionary rates. · Epitope distribution and conservation; Epitope definitions and evolutionary-rate…; Figure legends

  4. SuggestedMany binder epitopes overlap annotated functional sites, which supports site-specific perturbation hypotheses. Overlap alone does not show that function is altered.

    26% of ASD-annotated targets have a binder overlapping an allosteric site, though the methods annotated allosteric sites only per target. About half of annotated PINDER interfaces overlap an epitope. Roughly one quarter of ligand-site targets overlap. About half of essential genes in A375, K562 and HCT116 are targeted. 148 targets have epitopes overlapping DNA-binding regions. · Coverage of allosteric sites; Functional landscape of candidate binders for…; Allosteric-site coverage

  5. ShownDesigned interfaces resemble native ones in size and polarity but are biased towards helical binder segments. Binder folds mostly match known folds, while the interfaces are largely novel.

    17.3% of binders have a max query TM-score < 0.5 against the AFDB; median interface TM-score about 0.3 against PPIRef; interface Cα RMSD above 2 Å between bound and unbound states in 4% of binder candidates and 8% of targets. · Designed versus natural protein-protein…

Methods, data and reproducibility

BindCraft with default filters (fixed 90-residue binders, no hotspots) was run as parallel SLURM array jobs on AlphaFold DB domains selected by PAE clustering and confidence/compactness filters, with full-length clash checks. Coverage, conservation, interface and functional-site analyses followed.

  • Reviewed human UniProtKB/Swiss-Prot targets with AlphaFold DB (v6) structures and PAE matrices used as design targets · Homo sapiens · n = 20,279 human target proteins; 10,664 targetable
  • Bindome designed binder-target complexes (90-residue binders) · Homo sapiens (targets) · n = 306,146 candidates; 8,296 target proteins
  • Machine-learning splits of binder-target complexes (leakage-controlled, interface clustering) · Homo sapiens (targets) · n = 205,553 train, 21,205 validation, 21,955 test
  • PINDER native protein-protein complexes (release 2024-02) used for interface comparison · Homo sapiens
  • Allosteric Database (ASD, release 2023-09) targets and modulators · Homo sapiens
  • DepMap Public 26Q1 CRISPR Chronos gene effect for A375, K562, HCT116 · Homo sapiens
  • Affinity-reagent databases compared: Antibodypedia, ABCD, Biocompare, SAbDab · Homo sapiens
Methods & tools
BindCraft, AlphaFold Database (v6), ProteinMPNN, pae_to_domain, Foldseek, Rate4Site (pyrate4site v0.1.0), freesasa v.2.0.3, Biopython v1.86 (DSSP, Shrake-Rupley), PINDER (pinder v0.5.0), PPIRef/iDist and US-align, 3D-Beacons API, MCP server
Software & environment
Biopython v1.86, freesasa v.2.0.3, HH-suite v3.3.0, pyrate4site v0.1.0, pinder v0.5.0, US-align version 20260527, Foldseek commit 718d421, AlphaFold DB v6.
Reproducibility gaps
Hardware and runtime of the design campaign are not reported; no experimental binding data accompany the candidates.
Limitations (6)
  • Binder candidates still await experimental validation; realized success rates across the atlas are uncertain given the wide target diversity, and selectivity has not been characterized in depth. · stated by the authors
  • Paralog cross-reactivity is a risk, heightened by the preference of binders for conserved epitopes and by epitope convergence within CATH superfamilies. · stated by the authors
  • Per-target epitope coverage is sparse (13.9% of total surface area covered), and further advances in design methods are needed; it is unclear whether every interface is tractable. · stated by the authors
  • Overlap with functional annotations does not ensure that a binder perturbs native protein function; the overlaps only frame testable hypotheses. · stated by the authors
  • The allosteric result is internally ambiguous: results and methods describe residue-level overlap, but one methods section says allosteric sites were reported only as whole-target annotations. · noted in this summary
  • The conservation control is a size-matched surface sample that may include epitope residues, and it does not account for other surface properties (e.g., hydrophobicity), so the conservation signal may be partly confounded. · noted in this summary

Read this if you want a proteome-scale source of sequence-defined designed binders, or leakage-controlled binder-target complexes for ML, along with coverage and functional-site statistics.

Citations
0

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
THISBalbi '26Pačesa '24Swanson '25Team '25Cao '21Lourenço '24
69% semantic matchSynthetic Biology
Petra E. M. Balbi, Ahmed Sadek, Anthony Marchand+8 more·Proceedings of the National Academy of Sciences·Jan 28, 2026· 6 citations