posters.science → biomarker discovery (BOSC 2026 / CollaborationFest)

An end-to-end pipeline that mines the posters.science corpus (31,417 machine-actionable scientific posters) for biomarkers — including novel / unseen candidates, catalogues them with conditions and identifiers, and flags which are absent from a known biomarker database.

This workflow and project was completed as a part of BOSC CoFest 2026.

The design goal throughout: do not perturb the search for unseen biomarkers. No keyword pre-filtering is used to decide what is "about" biomarkers — posters are scoped by research field (a model, not a vocabulary), and biomarkers are pulled from full text by an LLM asked to include implicated/candidate markers, not just canonical ones.

Pipeline at a glance

posters-science-export.zip (31,417 posters, DataCite 4.7 NDJSON)
        │
        │  filter to publicationYear >= 2024                         → 11,028 posters
        ▼
[01] classify_fields    paper-to-field (BioM-ELECTRA) → OpenAlex field + [CLS] emb
[02] assign_topics      [CLS] nearest-neighbour vs topic_embeddings → subfield/topic
[03] merge_annotations  fold field/domain/topic into each record; split by domain
        │
        │  keep Health Sciences + Life Sciences domains              → 2,930 posters
        ▼
[04] extract_biomarkers DeepSeek full-text, BROAD prompt (mentioned OR implicated)
[05] enrich_catalogue   DeepSeek re-run: acronym, synonyms, conditions, specimen,
                        role, direction, method, population per biomarker
[06] assemble_catalogue collapse names (char → fairly semantic → alias dict) and
                        match each canonical entity vs the known BiomarkerKB list
                        → exact | partial | none   ('none' = candidate novel)
[07] seeded_pass        DeepSeek re-run seeded with our finds; pulls standard IDs
                        (HGNC/UniProt/CHEBI) + ontology conditions → re-match
[09] extract_conditions DeepSeek conditions for the biomarker-negative posters
[10] condition_frequency corpus-wide condition frequency (cross-conference)
[11] master_record      per-poster join of every layer
[12] rematch_extended   re-match vs BiomarkerKB + MarkerDB + PRGdb (HGNC/Ensembl alias)
[13] classify_none      methodical none-parse: NOT_ENTITY / KNOWN_UNCATALOGUED /
[13b] adjudicate_none      NOVEL_CANDIDATE / UNRESOLVED (deterministic + LLM)
[14] mine_known_uncatalogued  reference-DB coverage-gap audit

Models & data

Component	What	Where it runs
jimnoneill/paper-to-field	BioM-ELECTRA-Large, 26 OpenAlex fields + topic NN	local GPU (fp32; Pascal)
DeepSeek deepseek-chat	full-text biomarker extraction / enrichment	API
fair-ly-accurate	synonym-lustre name normalization (gte-large + HDBSCAN)	local CPU
BiomarkerKB (biomarker_list.csv) + MarkerDB 2.0 + PRGdb 4.0	curated biomarker reference sets (match targets)	local
HGNC + Ensembl Plants	gene-alias resolvers (normalization only, not match targets)	local

Name normalization (the three collapse tiers)

Biomarker surface forms are collapsed to canonical entities in three tiers, so counts are accurate and novelty is judged per-entity rather than per-string:

1. Character — NFKC + casefold + strip punctuation (IL-6 = IL6).
2. Semantic — fair-ly-accurate synonym-lustre: embed distinct names (gte-large), HDBSCAN cluster, with the acronym/short-token holdout so IL-6 and IL-8 never merge.
3. Alias dictionary — the acronym ↔ full-name ↔ synonym map that the DeepSeek passes emit is the controlled vocabulary the holdout deliberately leaves open (IL-6 ↔ interleukin-6).

Outputs

Written to ~/Downloads (see .gitignore — data is not committed):

File	Contents
posters-2024plus-annotated.ndjson	all 11,028, each record + _paper_to_field
posters-2024plus-biomarkers.*	the 2,930 biomedical posters + _deepseek_biomarkers
posters-2024plus.biomarkers-counts.csv	one row per canonical biomarker + match_status
posters-2024plus.biomarker-catalogue.csv	long: poster × biomarker × condition, full enrichment
posters-2024plus.biomarkers-index.json	machine-readable summary + novel (none) list

Running it

See docs/WORKFLOW.md for the full end-to-end runbook (environment, exact commands, host notes). Scripts in pipeline/ are numbered by stage; paths are currently set for the build host (parameterize via the config block at the top of each). DeepSeek needs DEEPSEEK_API in the environment.

Provenance / reproducibility notes

• The corpus dump and export are transferred out-of-band (too large for git).
• paper-to-field runs fp32 on the GTX 1070 — Pascal fp16 is 1/64 rate.
• Topic-embedding row order is validated (nearest-topic field vs classifier field) before topic labels are trusted.
• Every DeepSeek stage is resumable (skips ids already in its output file).

Team

Posters-are-cool (BOSC 2026 CollaborationFest): Jamey O'Neill, Sujeet Kulkarni, Hudson Smith, Olaitan Awe, Bhavesh Patel.

Citation

If you use this software, data, or results derived from it, please cite:

O'Neill J, Kulkarni S, Smith H, Awe O, Patel B. Potential Novel Biomarker Nomination from Recent Conference Posters. Posters-are-cool, BOSC 2026 CollaborationFest, 2026. https://github.com/fairdataihub/posters-science-bosc-2026-cofest

A machine-readable citation is in CITATION.cff.

Funding

Poster Sharing and Discovery Made Easy with Posters.science. This work was supported by The Navigation Fund. https://doi.org/10.71707/rk36-9x79