Data — I Beat Claude on TNFα for $677

Both arms of both targets, the per-engine scores behind the selection, and the binding kinetics from Adaptyv Bio. Released under ODC-BY.

File	What it is
`tnfa_designs.csv`	All 20 TNFα sequences. `Control = TRUE` is the arm with the sequence/structure co-optimization stage removed.
`pdl1_designs.csv`	All 20 PD-L1 sequences, same convention.
`engine_scores.csv`	Per-design scores from every co-folding engine used for selection: Boltz-2, AlphaFold2, Protenix, RFdiffusion3, OpenFold3, ESMFold2. ipSAE, RMSD, pLDDT, pTM, ipTM, interface PAE. Also the consensus rank each design was selected on.
`tnfa_binding_summary.csv`	Binding calls and fitted kinetics per design: per replicate and overall. `binding`, `binding_strength`, `kd_app` with CIs, `kon`, `koff`, fit quality, the fitted model, and the sequence.
`pdl1_binding_summary.csv`	Same for PD-L1.
`tnfa_kinetics/`	The fit-level tables behind the summary: `fits.csv`, `reads.csv`, `replicates.csv`, `blanks.csv`, plus the binding models — equations, parameters and units.

What is not here

The raw response curves and sensorgram plots — one per concentration per replicate, 46 MB. Ask and I will send them.

Caveats that matter when you read these numbers

The TNFα run is not closed. 19 of the 20 designs have both replicates. TNFA_OPT_08 currently has one. Its overall row is that single replicate.
kd_app is apparent, not monovalent. TNFα is a homotrimer and the binder is the immobilized partner, so avidity is built into this format. Several fits sit at the koff floor of 1e-5 /s, which means the off-rate was too slow to resolve in the dissociation window, not that it was measured to be that value.
Antigen: ACROBiosystems TNA-H4211, the same catalogue number MIT used for BoltzGen. Anthropic used TNA-H5228.
Measured by Adaptyv Bio, BLI, five-point concentration series, duplicate.

Anthropic’s data for the same target is at Anthropic/claude-protein-binder-design under CC-BY-4.0. Every figure in the post can be rebuilt from those two sources.

--- title: "Data — I Beat Claude on TNFα for $677" --- Both arms of both targets, the per-engine scores behind the selection, and the binding kinetics from Adaptyv Bio. Released under [ODC-BY](https://opendatacommons.org/licenses/by/1-0/). | File | What it is | |---|---| | [`tnfa_designs.csv`](tnfa_designs.csv) | All 20 TNFα sequences. `Control = TRUE` is the arm with the sequence/structure co-optimization stage removed. | | [`pdl1_designs.csv`](pdl1_designs.csv) | All 20 PD-L1 sequences, same convention. | | [`engine_scores.csv`](engine_scores.csv) | Per-design scores from every co-folding engine used for selection: Boltz-2, AlphaFold2, Protenix, RFdiffusion3, OpenFold3, ESMFold2. ipSAE, RMSD, pLDDT, pTM, ipTM, interface PAE. Also the consensus rank each design was selected on. | | [`tnfa_binding_summary.csv`](tnfa_binding_summary.csv) | Binding calls and fitted kinetics per design: per replicate and overall. `binding`, `binding_strength`, `kd_app` with CIs, `kon`, `koff`, fit quality, the fitted model, and the sequence. | | [`pdl1_binding_summary.csv`](pdl1_binding_summary.csv) | Same for PD-L1. | | `tnfa_kinetics/` | The fit-level tables behind the summary: [`fits.csv`](tnfa_kinetics/fits.csv), [`reads.csv`](tnfa_kinetics/reads.csv), [`replicates.csv`](tnfa_kinetics/replicates.csv), [`blanks.csv`](tnfa_kinetics/blanks.csv), plus [the binding models](tnfa_kinetics/binding_models.html) — equations, parameters and units. | ## What is not here The raw response curves and sensorgram plots — one per concentration per replicate, 46 MB. Ask and I will send them. ## Caveats that matter when you read these numbers - **The TNFα run is not closed.** 19 of the 20 designs have both replicates. `TNFA_OPT_08` currently has one. Its overall row is that single replicate. - **`kd_app` is apparent, not monovalent.** TNFα is a homotrimer and the binder is the immobilized partner, so avidity is built into this format. Several fits sit at the `koff` floor of 1e-5 /s, which means the off-rate was too slow to resolve in the dissociation window, not that it was measured to be that value. - **Antigen:** ACROBiosystems TNA-H4211, the same catalogue number MIT used for BoltzGen. Anthropic used TNA-H5228. - Measured by Adaptyv Bio, BLI, five-point concentration series, duplicate. Anthropic's data for the same target is at [Anthropic/claude-protein-binder-design](https://huggingface.co/datasets/Anthropic/claude-protein-binder-design) under CC-BY-4.0. Every figure in the post can be rebuilt from those two sources.