antibody-literature-search

Antibody literature and database search

Getting from a question to sourced evidence, with every citation verified.

Resolve before writing

An identifier that has not been fetched in this session does not get written down. This applies to PMIDs, DOIs, accessions, PDB IDs, clone names, catalogue numbers and trial IDs alike.

The discipline, in order:

1. Search, do not recall. Antibody and immunology citations are the easiest thing in this field to generate plausibly and wrongly: the right-looking author, journal and year attached to a paper that does not exist.
2. Fetch the record for every identifier that will appear in the output.
3. Match the record to the claim. A resolving PMID proves the paper exists, not that it says what the answer attributes to it. Read the abstract at minimum; read the methods when the claim is about how something was measured.
4. Carry provenance into the output — identifier, title, year, and what that source actually supports.

If a source cannot be resolved, write the claim as unsourced or drop it. Stating "I could not find a reference for this" is a correct answer.

Database records carry provenance too. A record is cited with its accession plus, where the resource versions its data, the release or version and the date it was retrieved — UniProt P04626 (retrieved 2026-11-13), PDB 1N8Z, Thera-SAbDab trastuzumab, IEDB epitope ID 12345. Database content changes between releases, so a bare accession with no version is not reproducible. Prefer the primary source (the deposited record) over a secondary summary of it.

Pick the resource from the question

Question	Go to
What is published about this target, antibody or method?	Europe PMC or PubMed
What is this protein, its domains, isoforms, PTMs, sequence?	UniProt
Which epitopes are known or mapped for this antigen?	IEDB
Is there a structure of this complex?	RCSB PDB, then SAbDab (SAbDab-nano for VHHs) for antibody-specific annotation
What do I need to know about this target before designing against it?	the target dossier: references/target-dossier.md
Which designed binders were tested for this target or by this method, and what happened?	Proteinbase
What mutational binding or developability data exist to calibrate against?	SKEMPI 2.0, AB-Bind, Ginkgo Datapoints
What is the sequence and format of a named therapeutic antibody?	Thera-SAbDab, IMGT/mAb-DB
Has this antibody sequence been published or patented before?	PLAbDab
Which germline genes and alleles exist?	IMGT/GENE-DB, OGRDB
Has this repertoire or clonotype been seen in a study?	OAS, iReceptor, VDJbase
Which trials exist and what phase?	ClinicalTrials.gov, EU CTIS, WHO ICTRP
Which antibodies are approved or under review?	The Antibody Society's table, YAbS
Has this sequence been claimed in a patent?	Google Patents, Lens.org, patent sequence databases
Which commercial antibodies target this protein, and are they validated?	antibody-reagent-selection

Details, URLs and API notes: references/resources.md. Query syntax and worked recipes: references/search-recipes.md.

With the immuno MCP

When the immuno MCP server is connected (get_server_capabilities shows it), prefer it to hand-written queries for these steps; every result carries provenance to cite.

• resolve_identifier confirms a PDB id, UniProt accession, PMID, PMCID or DOI before it is written.
• search_antibody_structures queries SAbDab2. Search by UniProt accession, not by protein name: a name such as "PD-1" or "Programmed cell death protein 1" with the VHH filter returned no rows on 2026-11-14, while the accession Q15116 returned 51. The name filter matches 25 characters, does not check species, and misses entries that name the antigen differently. Then get_structure for the entry, its CDRs and its antigen chains.
• search_literature then get_passages read and cite the text; a title alone is not evidence, and passages are abstract-only unless coverage says full text.
• query_binding_data returns SKEMPI 2.0 rows (ddG, Kd, the source PMID). The dataset dates from 2018.
• Datasets the server lists as planned (OAS, AB-Bind, GDPa1, Adaptyv) are not queryable yet. Say so rather than citing them.

Search breadth

A single query answers almost nothing in this field. For any question beyond a direct lookup, run several searches in one batch and vary the vocabulary, because the same object has many names:

• the target: gene symbol, protein name, CD number, former names, species.
• the antibody: INN, code name (often ABC-123), brand name, clone ID, originator company.
• the concept: the method name plus its common abbreviation.

Stop when the same papers keep returning from different vocabularies. That saturation is the completion criterion, not a fixed number of searches.

Summarise with the evidence attached

For each paper that earns a place in the answer, extract the fields in references/reading-a-paper.md — target, antibody and format, what was measured and how, model system, controls, effect size, limitations — and distinguish throughout:

• what was measured versus what was concluded,
• in vitro, animal, and human evidence,
• peer-reviewed versus preprint versus conference abstract versus company press release.

Rank claims by that hierarchy rather than by how recent or emphatic they are. A press release reporting a positive phase 2 is weaker evidence than the published trial, and both should be labelled as what they are.

Fetching politely

Public APIs here are free and rate-limited. Batch requests, obey the documented limits, and cache results in the scratchpad rather than re-fetching.

NCBI's E-utilities ask callers to identify themselves with tool and email parameters. Send a generic client name, and ask the user before putting their own email address in a request to any external service.