Expression → Binder

An open-source pipeline that joins cohort RNA-seq target discovery to de novo protein-binder design in one reproducible workflow — with machine-readable provenance from every ranked binder back to the patient samples it came from.

Genomics stops at “here are the interesting genes”. Protein design starts at “given a target structure”. bindsight is the reproducible, citable bridge between them.

Open research workspace See real results What is bindsight? Run it yourself GitHub

In plain terms

bindsight compares RNA-seq counts between conditions to prioritise candidate cell-surface targets. It can design protein sequences and assess their predicted complexes while recording inputs, settings and outputs. RNA abundance is not surface protein abundance, and model confidence does not establish binding, specificity or clinical safety. New to the terms? See the Glossary.

1 of 291

CA9 surfaced

clear-cell kidney, patient-matched TCGA-KIRC contrast without biomarker stratification

0.88

best ipTM

20 de novo ERBB2 binders on a free Kaggle T4

30% / 50%

designs / shuffled controls

6/20 designs and 10/20 shuffles exceed ipTM 0.65 under the seeded five-draw calibration. The threshold-crossing rate is withdrawn as a measure of design quality.

1 of 17

recall @ rank 20

approved-agent antigens, 95% CI 0.01–0.27

Those numbers come from recorded computations. Committed summaries, predicted structures and provenance metadata can be inspected here; the raw calibration inputs, cohort caches and complete provenance crate are not all included. See the evidence and its reproducibility limits.

These figures come from the corrected protocol

An earlier run invoked ProteinMPNN without --pdb_path_chains, so it redesigned the HER2 target chain as well as the binder, and its numbers (best ipTM 0.84, 50% success) are withdrawn. The run above holds the target fixed, and that is checked rather than asserted: all 20 designs carry a target chain byte-identical to the native 142-residue domain IV. The corrected mean ipTM is lower — 0.51 against 0.59 — which is what you would expect once designs stop being scored against a surface they helped invent.

What 1 of 17 does and does not show

That is the recall of clinically approved antigens across a pre-registered panel of 22 antigen-cohort pairs across fifteen TCGA projects, using patient-matched tumour/normal samples without biomarker stratification. The interval matters at this panel size, and repeated antigens make the pairs correlated. Eleven of the seventeen fail the composite significance rule, one is downregulated, and three fall beyond the enrichment cutoff. These are different outcomes, not a blanket claim of absent overexpression. Approved-agent tier does not mean approval in every listed cancer. An earlier six-cohort version reported ERBB2 at rank 4 and recall@5 of 33%; it is withdrawn, because its breast cohort was stratified by a PAM50 subtype call and ERBB2 is one of the fifty genes that classifier is built from.

How it works¶

Patient RNA-seqcounts + design

Differential expressionpydeseq2

Cell-surface filterSURFY surfaceome

Safety + tractabilityGTEx · Open Targets

Targetable siteSURFACE-Bind · AlphaFold

Binder designRFdiffusion + MPNN

Structure + affinityBoltz-2

Ranked candidatesmulti-objective

ProvenancePROV-O · RO-Crate

Amber stages need a GPU. A free Kaggle T4 is the verified route, and the one the committed benchmark used. Local protein design requires a separately configured NVIDIA/CUDA environment; the default Docker image is CPU-only. CPU discovery is subject to dataset memory requirements. See the local workspace guide for the downloadable setup and platform limitations.

Try it, three ways¶

In your browser

bindsight ui serves the interface locally in seconds: the committed evidence, the twenty designed binders in 3-D, and a one-click run of the discovery half on a real TCGA cohort. Its evidence surface is published here too, with nothing to install: the twenty complexes draw in this page's browser, and the input checker is the same file bindsight ui loads.

One command

bindsight demo runs the whole discovery half locally on a real NIH/GDC cohort and writes a self-contained HTML report.

Your own cohort

Point it at your counts matrix and sample design, choose a GPU backend when you are ready to design, and export an RO-Crate for deposit.

pip install "bindsight[discover,report]"
bindsight demo      # real TCGA-BRCA cohort, CPU only, full provenance
bindsight ui        # the web interface, locally

Start here¶

Real results — what it has actually demonstrated.
What is bindsight? — the 5-minute pitch.
How to use it — install, the demo, and the full discover → design → validate → rank → report → export flow.
Use cases — concrete scenarios.
Designing on Colab — the GPU half on free Colab.

How it fits together¶

The CLI (bindsight …) and an optional Snakemake front-end both drive the same Python pipeline. The discovery half is CPU-only; the design half (RFdiffusion → ProteinMPNN → Boltz-2, plus BindCraft / BoltzGen / Chai-1r / AF2-IG) runs on a GPU backend you choose (Modal / local Docker / Kaggle / Colab).

See ARCHITECTURE.md, LICENSING.md, and CONTRIBUTING.md in the repository for design rationale, the per-component license inventory, and how to add a designer / validator / runner plugin.

Cite it¶

bindsight is AGPL-3.0-or-later. This release is not archived under a DOI — if it helps your work, cite the repository, the tag you ran and the checksums attached to the release. Ready-made entries are in CITATION.cff.