AI × Biology Papers Corpus

The definitive, machine-readable index of AI × Biology papers accepted at ICLR, ICML and NeurIPS.

Browse  ·  Figures  ·  Tags  ·  Data  ·  Load  ·  Methods

---

Why this exists

🎯 One canonical corpus Every AI × Biology paper accepted at ICLR / ICML / NeurIPS, from NeurIPS 2010 to ICLR 2026, in one place. 3,722 records, all with abstract, authors and PDF link.

🏷️ Machine-labelled, human-audited Two-stage keyword filter with 95.9 % recall against a 315-paper hand-curated truth set. Ten topical tags (protein, molecule, single_cell, …) assigned per paper.

📊 Ready to analyse Ships as structured JSON + CSV plus 33 foldable markdown files you can browse on GitHub. Reproducible: the entire tree regenerates from one JSON.

---

🚀 Quick start

git clone https://github.com/biotender/ai-bio-conference-papers.git
cd ai-bio-conference-papers

# Browse in your editor — every paper is a foldable <details> card
open papers/ICLR/2026.md

# Or load the data
python -c "import json; papers = json.load(open('data/papers_master.json')); print(len(papers), 'papers')"

---

📈 At a glance

Corpus overview — 3,722 papers, 2010 → 2026

Stacked bars by venue, one bar per year. NeurIPS dominates the early record (2010–2017, ICLR/ICML absent from the corpus); ICLR takes over from 2020 onwards. The 2026 bar reaches 1,235 papers even without NeurIPS 2026, which had not yet run at snapshot time. The AlphaFold2 reference line marks the 2021 inflection point.

SVG (vector, editable): figures/corpus_overview.svg

Topic trajectories — 7 tags, 2017 → 2026

Per-year paper counts for the seven densest tags. protein climbs from ~2 papers/year (2018) to 294 in 2026 — a textbook post-AlphaFold2 explosion. single_cell lifts late (25 → 125 across 2020–2026). genomics roughly doubles between 2024 → 2025 (59 → 148). medical_imaging and clinical grow more steadily.

SVG (vector, editable): figures/topic_trajectories.svg

Both figures are regenerated directly from data/papers_master.json; the plotting code lives in the companion biotender-ai-bio-corpus repo.

---

🗂 Browse by venue × year

Each cell links to a markdown file containing all papers for that venue/year, grouped by tag and sorted alphabetically. Numbers are paper counts.

Year	ICLR	ICML	NeurIPS	Total
2010	—	—	10	10
2011	—	—	5	5
2012	—	—	13	13
2013	—	—	7	7
2014	—	—	9	9
2015	—	—	6	6
2016	—	—	14	14
2017	—	—	18	18
2018	—	4	21	25
2019	31	2	27	60
2020	75	13	42	130
2021	80	10	42	132
2022	126	23	71	220
2023	217	39	141	397
2024	264	73	195	532
2025	501	134	274	909
2026	887	348	—	1,235
Total	2,181	646	895	3,722

Note: NeurIPS 2026 is not yet in the corpus — the conference had not run when this snapshot was taken. ICML 2019 has only 2 papers because the PMLR proceedings for that year yielded very few abstracts matching the biology filter (title-only signal for the pre-2020 PMLR floor is weaker).

---

🎴 Anatomy of a paper card

Every paper inside papers/*/*.md is a foldable <details> block. GitHub renders it natively — click the arrow to expand.

Collapsed view (always visible):

Title — Author1, Author2, Author3 et al. · page · pdf · tag1 tag2

Expanded view (unfolded on click):

Full abstract (verbatim, HTML-escaped).

Matched: clinical (13) · prognostic · healthcare

The Matched: … line lists which biology keywords the filter picked up in this paper (with per-term hit counts when a term appears more than once) — useful for auditing why a paper was kept.

Sorting rules inside each per-year markdown (click to expand)

1. Papers are grouped by their first _tags entry (or untagged if the list is empty).
2. Tag groups are ordered by descending paper count within that file.
3. untagged always appears last.
4. Inside each tag group, papers are sorted alphabetically by title (case-insensitive).

Multi-tag papers only appear once (under their first tag) so per-file counts stay additive to the venue-year total. The complete tag list is preserved in papers_master.json.

---

🏷 Tag universe

Tags are non-exclusive — a paper can carry multiple tags. Counts below are over the full corpus (a paper tagged both protein and single_cell contributes to both rows).

Tag	Papers	Definition
🧪 molecule	1,005	small-molecule drug discovery, molecular property prediction, generative chemistry, SMILES/graph modeling
🩺 clinical	925	clinical prediction, electronic health records, healthcare, prognostic / diagnostic modeling
🧬 protein	925	protein structure prediction, protein language models, protein design, folding
🧠 neuro	916	neural signals / spikes / EEG / fMRI, brain imaging, neuroscience
🧬 genomics	480	DNA / RNA sequence models, variant calling, gene expression, transcriptomics
🩻 medical_imaging	465	radiology, histopathology images, medical segmentation, X-ray / CT / MRI
🔬 single_cell	401	single-cell RNA-seq, cell type prediction, single-cell foundation models
⚛️ md_structbio	286	molecular dynamics, structural biology, force fields, energy prediction
🦠 oncology	188	cancer-specific modeling — subtypes, tumor imaging, drug response
🛡 immuno	91	immunology — antibody / TCR / B-cell modeling, MHC binding, vaccine design
— untagged	132	Passed the biology filter but did not hit any tag keyword — kept for completeness.

---

📁 Repository layout

ai-bio-conference-papers/
├── README.md              ← you are here
├── LICENSE                MIT (scripts, README, CHANGELOG)
├── LICENSE-DATA           CC BY 4.0 (paper records + markdown + figures)
├── CHANGELOG.md
├── .gitignore
├── assets/
│   ├── hero_banner.png              README hero image
│   ├── biotender_logo.png           original logo (black on white)
│   ├── biotender_logo_white.png     white-on-transparent for dark bg
│   └── biotender_logo_black.png     black-on-transparent for light bg
├── papers/
│   ├── ICLR/2019.md .. 2026.md      (8 files)
│   ├── ICML/2018.md .. 2026.md      (9 files)
│   └── NeurIPS/2010.md .. 2025.md   (16 files)
├── data/
│   ├── papers_master.json           canonical, 3,722 dicts (~8.2 MB)
│   └── papers_master.csv            flattened subset (~1.3 MB)
├── figures/
│   ├── corpus_overview.{png,svg}    stacked bars, 2010–2026 by venue
│   └── topic_trajectories.{png,svg} 7-tag lines, 2017–2026
└── scripts/
    ├── generate_papers_md.py        regenerate 33 markdown files
    └── build_readme.py              regenerate this README

Total: 33 per-year markdown + 2 data files + 4 figures + 4 assets + 2 scripts + 5 root files = 50 tracked files, ~18 MB working tree.

---

📚 Data files

data/papers_master.json — canonical

A JSON array of 3,722 dicts. This is the source of truth; every markdown, count and table in this repo is derived from it.

Field dictionary (click to expand)

Field	Type	Description
title	str	Paper title, verbatim from the source.
abstract	str	Full abstract; never empty in this corpus.
authors	list[str]	Full author list, ordered as printed on the paper.
keywords	list[str]	Author-supplied keywords (OpenReview only; empty for PMLR / papers.nips.cc).
venue	str	One of ICLR, ICML, NeurIPS.
year	int	Publication year.
source	str	Upstream source: openreview / pmlr / nips_cc.
primary_area	str	OpenReview submission track; empty for non-OpenReview sources.
forum_id	str	OpenReview forum id when applicable.
paper_page	str	URL to the paper landing page.
pdf_url	str	Direct PDF URL.
_tags	list[str]	Topical tags assigned by the BioTender filter.
_hit_terms	list[str]	Biology keywords that the filter matched, order preserved.
_hits	dict[str,int]	Per-term hit counts across title + abstract + keywords.
_keep_reason	str	Which filter branch kept this paper (e.g. text_high_spec_focused, title_strong_bio).
_area	str	Coarse area bucket, or no_area.

Fields prefixed with _ are added by the filter; the rest are pulled verbatim from the upstream source.

data/papers_master.csv — flattened subset

For quick inspection in Excel / pandas. Drops several fields for size — see the mapping below. Use the JSON if you need the abstract or per-term hit counts.

CSV column mapping (click to expand)

CSV column	Corresponds to JSON field
venue, year, title, source, primary_area, forum_id, paper_page	same (verbatim)
authors	authors — list joined by ;
tags	_tags — list joined by ; (underscore dropped)
hit_terms	_hit_terms — list joined by ; (underscore dropped)
keep_reason	_keep_reason (underscore dropped)

Not in the CSV: abstract, keywords, pdf_url, _hits, _area. Load papers_master.json for those.

---

💻 Loading & usage

Python — JSON

import json

with open('data/papers_master.json') as f:
    papers = json.load(f)   # list of 3722 dicts

# Example: all ICLR 2026 protein-tagged papers
hits = [p for p in papers
        if p['venue'] == 'ICLR'
        and p['year'] == 2026
        and 'protein' in p['_tags']]
print(len(hits), 'papers')

# Example: papers matching a specific keyword
cryo = [p for p in papers if 'cryo-em' in p['_hit_terms']]

Python — CSV

import pandas as pd

df = pd.read_csv('data/papers_master.csv')

# Reconstruct list-valued columns
for col in ('authors', 'tags', 'hit_terms'):
    df[col] = df[col].fillna('').apply(lambda s: s.split('; ') if s else [])

# Group by venue-year
df.groupby(['venue', 'year']).size()

Regenerate the markdown

The 33 per-year markdown files and this README are both regenerable from data/papers_master.json:

python scripts/generate_papers_md.py    # rewrites papers/**/*.md
python scripts/build_readme.py          # rewrites README.md

Neither script has external dependencies (Python 3.10+ standard library only).

---

🔬 Methodology

Data sources

Venue	Years	Source	Access pattern
ICLR	2019–2026	OpenReview API v2	api2.openreview.net/notes?content.venue=…
ICML	2018–2026	PMLR proceedings (+ OpenReview for 2026 only)	proceedings.mlr.press/v<vol>/
NeurIPS	2010–2025	papers.nips.cc (+ OpenReview 2023 onward)	papers.nips.cc/paper_files/paper/<year>

All fetching uses public endpoints with polite rate-limiting. Only accepted papers are included — rejected / withdrawn submissions and workshops are out of scope.

Filter design (summary)

Two-stage filter, applied to title + abstract + keywords:

1. Keyword hit count against a curated 44-term biology lexicon (protein, ligand, patient, EEG, transcriptomic, cryo-em, single-cell, …).
2. Tag assignment against 12 topical groups (10 stayed active in this corpus — snn and chem-pure-materials are used only as exclusion branches).

A paper is kept if its combined signal exceeds a per-branch threshold. False-positive branches (snn_dropped, chem_pure_materials_dropped, brain_false_positive, neuro_no_neurobio, chem_bio_no_bio) explicitly drop papers that match biology keywords but are actually about spiking-neural-networks, computational chemistry of pure materials, or brain-inspired ML without any biology content.

The full filter code (with all lexicons, thresholds and branch definitions) lives in the companion biotender-ai-bio-corpus repo under scripts/filter_pipeline.py. This repo intentionally ships only the output.

Filter validation

Against a 315-paper hand-curated truth set (ICML 2026 papers manually labelled as "in scope for AI × biology"):

Metric	Value
Truth set size	315
Filter kept	348
Overlap (true positives)	302
Recall	95.9 % (302 / 315)
Missing (false negatives)	13 — all borderline chemistry / materials science with weak biology signal
Extras (kept beyond truth)	46 — audited as legitimate AI × biology, retained after review

The 46 extras include: bio-inspired representation-learning papers (e.g. hippocampal place-field models, biologically-plausible connectivity), a handful of protein / genomic foundation-model papers where the reviewer had originally been strict about "pure ML methods on bio data", and several medical-imaging papers with heavier ML machinery. All 46 were audited during the validation pass and kept.

---

⚠️ Known limitations

1. PMLR floor. ICML abstracts before 2020 are partly truncated on PMLR — filtering leans more on titles for those years. This is why ICML 2019 shows only 2 papers.
2. Accepted-papers only. Rejected / withdrawn submissions are not included. Workshops are not covered.
3. Editorial subjectivity. The 46 legitimate extras vs. 13 borderline chemistry/materials exclusions are judgement calls. See the filter code for the exact rules.
4. Tag first-match rule. In per-year markdown files, multi-tag papers appear under their first tag only. The complete tag list is preserved in papers_master.json.
5. NeurIPS 2026 not included. Snapshot taken before the conference ran.
6. Snapshot date. This is a point-in-time cut; papers added / retracted after the generation date shown at the top of this README are not reflected.

---

📖 Citation

If you use this corpus, please cite it as:

BioTender (2026). ai-bio-conference-papers: 3,722 AI × biology papers
from ICLR, ICML and NeurIPS (2010–2026). https://biotender.online

Paper titles, abstracts and author names remain the copyright of the original authors and their publishers. This project claims copyright only over the compilation, tag assignment, filter rationale (_tags, _hit_terms, _hits, _keep_reason, _area) and the markdown organisation.

---

⚖ License

This repository is dual-licensed:

• Scripts & documentation — MIT (see LICENSE). Applies to scripts/, README.md, CHANGELOG.md.
• Paper records, markdown, figures & assets — CC BY 4.0 (see LICENSE-DATA). Applies to papers/**/*.md, data/papers_master.{json,csv}, figures/*.{png,svg}, assets/*.png.

Users redistributing this repository must respect the underlying copyright of the paper text itself, which is held by the original authors and publishers.

---

_{Part of the BioTender · OSINT curation for AI × Biology}