BioFormBench: Two Open Benchmarks for Biologics Formulation Research

This release contains two datasets, kept separate because they answer different questions and neither should be read as a substitute for the other.

Data integrity note (please read before using either dataset)

An earlier internal draft of BioFormBench-Real contained 67 rows. Auditing source_title/source_id provenance found that 18 of those rows had no citation and used placeholder identifiers (protein_1–protein_5) with round, unverified descriptors (50.0 kDa / pI 7.2 / Tm 65.0 C) repeated across otherwise-different molecules. Those rows have been removed. The 49 rows in this release all trace to a PubMed Central ID and were checked against their source text. If you have a copy of the 67-row file, discard it and use this one.

1. BioFormBench-Real (49 rows, 13 proteins)

Literature-mined formulation-outcome measurements (DSF Tm-shift, SEC % aggregation, turbidity) for 13 real proteins across 11 primary papers, each row traceable to its source. This is the dataset to use for any claim about stability-optimal formulation, but at n=13 proteins (9 with enough per-protein data for leave-one-protein-out evaluation) it is underpowered for a definitive protein-conditionality verdict — see the companion paper's Discussion for what scaling this further would require.

2. BioFormBench-Marketed (165 rows, 80 antibodies) — new in this release

165 formulation presentations for 80 FDA-approved antibody therapeutics, deterministically parsed from FDA Structured Product Labeling (DailyMed) and linked to real VH/VL variable-domain sequences from Thera-SAbDab. Every parsed field keeps its verbatim source sentence (source_text) for audit. To our knowledge this is the first open dataset connecting therapeutic antibody sequence to marketed formulation composition at this scale.

This dataset encodes what was chosen for an approved product, not what is necessarily stability-optimal — marketed choices are also shaped by manufacturability, prior platform investment and regulatory precedent. Used at full sample size (n=80) with proper multiple-comparison correction and leave-one-protein-out validation, protein sequence identity does not detectably predict the chosen pH or buffer beyond a constant "platform" baseline (all sign-flip p > 0.8) — see the companion paper for the full statistical analysis, including a worked example of how a naive n=27 subsample of this same data produces a spurious, non-reproducible positive result (Spearman rho=-0.464, p=0.015, uncorrected) that a full-sample, corrected, out-of-sample analysis overturns.

Use cases

1. Test protein-conditional formulation hypotheses at adequate statistical power, with the selection-artifact and platform-baseline results in the companion paper as a required comparison point, not merely a possible one.
2. Train/evaluate predictive or generative formulation models against a documented platform baseline rather than only random/uniform baselines.
3. Audit other mechanistic formulation simulators using the diagnostic described in the companion paper (dead-slot spread, interior-argmax fraction) against these real formulations.
4. Sequence-to-formulation representation learning (BioFormBench-Marketed includes Fv-derived descriptors so no additional sequence processing is required to get started).

Files

• bioformbench_real.csv — 49 rows, BioFormBench-Real (see schema above).
• bioformbench_marketed.csv — 165 rows, BioFormBench-Marketed (see schema above; full computed-descriptor and label-parse columns included).

Known limitations

• BioFormBench-Real: n=13 proteins (9 LOPO-eligible) is small; treat any per-protein result as illustrative, and prefer the aggregate statistics with their reported confidence intervals.
• BioFormBench-Marketed: pH is extracted with a confidence flag (ph_is_range/composition-sentence provenance in the full release columns); the low-confidence stratum is retained rather than silently dropped, and the companion paper shows the main finding (no sequence effect) holds in the high-confidence stratum alone. Only 27–28 of 80 molecules have a fully resolved buffer species and concentration; buffer-level analyses should use that subset and are correspondingly lower-powered.
• Neither dataset should be used to generate or select actual formulations for laboratory or clinical use without independent verification; see the companion paper's Ethics statement.

Citation

@dataset{kumar2026bioformbench,
  title={BioFormBench: Two Open Benchmarks for Biologics Formulation Research},
  author={Kumar, Bonthada Sravan},
  year={2026},
  publisher={Hugging Face Datasets},
  url={https://huggingface.co/datasets/Sravankumarbonthada/BioFormBench}
}

Sources

BioFormBench-Marketed is built from public FDA Structured Product Labeling (DailyMed) and Thera-SAbDab (Raybould et al., Nucleic Acids Research 2020). BioFormBench-Real is mined from 11 open-access PubMed Central articles, each cited per-row via source_id.

Contact

Bonthada Sravan Kumar — sravansaijohn@gmail.com — issues via GitHub (Maheshbonthada/bioform-lm)

License

CC-BY-4.0

---

Last updated: September 8, 2026