Explore POISK: antibody sequences and functional metadata from the global patent literature.
Read the paper ↗·Accepted at NeurIPS 2026
The compact search index loads once. Full records load as you explore.
Try a shorter identifier, another target name, or fewer filters.
This table scrolls sideways. With it focused, use the left and right arrow keys to reach every column.
One row represents a binding record. A single antibody may appear in multiple records.
POISK brings antibody sequences and functional metadata from patent literature into a structured research dataset.
Free text searches record IDs, patent IDs, antibody identifiers, target protein and gene names, UniProt IDs, and the discovery status, source organism, and platform columns. Words are matched together, ignoring case, punctuation, and Greek-letter spelling, so “PD-1” also finds “PD1”, and “TGFβ” finds “TGF-beta”. Use quotation marks for a phrase. When no column sort is chosen, matches are ordered by quality, with whole-word hits on a target or antibody name first.
Choose “VH/VL sequence (exact)” and paste a chain to compare it against the — distinct VH and VL sequences in this release. Only letters are compared, so spacing, numbering, and case do not matter — but the comparison is exact and whole-chain. A fragment, a single-residue variant, or a chain shorter than 20 residues will not match, and an empty result is not evidence that a sequence is unpatented. Alignment, substring, and approximate sequence matching are not provided. Antibody type, target type, organism, assay method, epitope type, and audit verdict are filters rather than free text; audit reasoning and target sequences appear in a record but are not searched.
Combine filters with your search. Each filter accepts several values at once, and the count beside an option follows your current results. Organism and assay labels preserve the original dataset wording, so variants such as “Human” and “Homo sapiens” remain separate options — select both to see both. Type inside a filter to narrow its labels, then select every match at once; long lists render the first 50 matches, so type to reach the rest. Filters and column sorting are paused during an exact sequence lookup.
Records were extracted using language models, then audited against the patent. Every record carries one of three verdicts. CONFIRMED means every field was explicitly supported by the patent text. PLAUSIBLE means nothing in the patent contradicted the record, but support was partial, indirect, or inferred. REVIEW means the record appeared incorrect, or its correctness could not be established and a person should examine it — it is not an assertion that the record is false. These are extraction audit verdicts, not independent experimental validation. Each record retains its audit explanation and patent reference.
Affinity values retain their original units, qualifiers, and measurement descriptions. They are not normalized into a single numeric scale. Missing values are shown as a dash. The data-available filters exclude empty values and common placeholders; ticking both chain boxes selects records that report a paired heavy and light chain.
Cite the POISK paper and the dataset version you used. Export matching metadata, complete records, or the VH and VL chains as FASTA. The original CSV, release information, and applicable dataset license are available on Zenodo. This explorer does not assign a new license to the data.
Please cite the POISK paper when using this resource. Accepted at NeurIPS 2026 ↗
Tupikina, D., Roncoli, A., Bujotzek, A., & Abanades, B. (2026). Can Folding Models Tell Binders from Bluffers? Evidence from POISK: The Patent-Derived Antibody Dataset. bioRxiv. https://doi.org/10.64898/2026.10.24.761513
Also cite the dataset version you used so others can reproduce your work.
The versioned dataset citation is available with the Zenodo release.
License details are provided with the Zenodo release.
For the complete, unmodified release, download the original CSV from Zenodo.