Where bindsight fits — positioning & landscape¶
A short, honest map of where
bindsightsits in the protein-design / target-discovery ecosystem: who it's for, the gap it fills, and what's on the roadmap. If you're evaluating bindsight for your group, start here.
The gap bindsight fills¶
Two mature ecosystems sit side by side and barely talk to each other:
- Genomics / transcriptomics (DESeq2, edgeR, Seurat, scanpy, TCGA, recount3) stops at "here are the interesting genes."
- De novo protein / binder design (RFdiffusion, ProteinMPNN, BindCraft, BoltzGen, AlphaFold, Boltz-2) starts from "given a target structure…"
Every open binder-design tool we surveyed — BindCraft, BinderFlow, PXDesign, Latent-X — begins with a known target. The path from expression data to a designed, ranked, provenance-tracked binder candidate is built ad-hoc, per project, rarely reproducibly.
bindsight ships that bridge as one open, reproducible tool — and records the receipts all the way back to the patient cohort.
Who it's for¶
| Audience | What bindsight gives you |
|---|---|
| Translational researchers | A free, reproducible "data → designed binder" path on a laptop, with free-GPU offload. |
| Clinical / cancer biologists | An audit trail from any binder candidate back to the cohort it came from. |
| Method developers | A held-out evaluation harness (rediscovery of known antigens) to benchmark new designers/validators behind a stable plugin interface. |
| Early-discovery teams | An open, extensible comparator you can plug proprietary designers into — no fork required. |
How it relates to neighboring tools¶
bindsight is an orchestration + provenance layer, not a new model. It stands on, and credits, the best open tools in each step (see Acknowledgments). Its contribution is the connective tissue: the surfaceome/targetable-site filter (SURFACE-Bind), the multi-objective ranking, the cost-aware GPU offload, the failure taxonomy, and the PROV-O + RO-Crate provenance that makes a run citable and reproducible.
| The binder-design tools surveyed above | bindsight | |
|---|---|---|
| Input | Target structure | RNA-seq counts |
| Provenance | PDB + maybe a log | PROV-O JSON-LD + RO-Crate, audit trail to cohort |
| Hardware | HPC assumed | CPU laptop + a free Kaggle T4 for the GPU half (the one verified backend; Modal is the paid escape hatch and Colab needs a human) |
| Cost-awareness | None | --dry-run estimates GPU $ before running |
| Negative results | Discarded | Catalogued (failure_taxonomy.parquet) |
| Citability | Code dump | Tagged releases with checksums, schema-validated outputs (Pydantic v2) |
Roadmap¶
- v0.3.6 (now) — discovery half end-to-end on CPU; design + validation
demonstrated on a free Kaggle T4 under the corrected ProteinMPNN protocol:
20 de novo ERBB2 binders, best ipTM 0.88, 40% success@0.65 (8/20, 95% CI 15–70% over backbones — withdrawn as a
measure of design quality: shuffles of the designs' own sequences clear 0.65 more
often (50% against 30%), see
benchmarks/calibration/README.md), every one carrying a target chain byte-identical to the native domain IV. Multi-page web UI shipped (bindsight ui, local). - v0.4.0 — live (async, non-blocking) Modal/Colab job submission; BindCraft + BoltzGen plugins fully wired; scRNA-seq input. None of this is in v0.3.6: no shipped backend builds an environment BindCraft or BoltzGen can run in, which is why the three-way designer comparison is deferred rather than delivered.
- v1.0.0 — the JOSS resubmission, on or after 2026-12-07 and with the
rediscovery study posted as a preprint: it is the instance of the software
being used for research that JOSS's pre-review asks for, which is why it goes
first (
paper/README.mdsays why;paper/SUBMISSIONS.mdrecords every venue tried and what came back). The study itself is already run and published inbenchmarks/study/: fifteen unstratified TCGA projects against a pre-registered panel, recall at rank 20 of 1 in 17 with a 95% interval of 0.01 to 0.27. An earlier HER2-centred version is withdrawn as circular, so v1.0 is about strengthening the signal — single-cell evidence, co-expression, immunopeptidomics — not about repeating the same measurement.
Get involved¶
bindsight is AGPL-3.0-licensed and built in the open. If you run target discovery or binder design and want to compare notes — or you'd like to try it on your own cohort — open an issue or reach the author (@mikhaeelatefrizk, ORCID 0009-0006-1069-9558). Feedback from real workflows directly shapes the roadmap.