Distribution across the audit rubrics defined in the project schema (§9–§14, §10 badges). All bars normalize to campaign count (n=60 sub-campaigns for E/D/A/H/badges/domain; n=103 papers for peer review; n=101 for model openness).
Transparency about limitations. Where a paper's funnel data is incomplete or a claim requires expert judgment, this is surfaced rather than hidden.
Purpose. BioTender AI Biology Wet-Lab Evidence Tracker prevents misleading equivalence between computational prediction, primary screening signal, confirmed binder, functional molecule, and in vivo validated candidate. It reconstructs each campaign's full path from computational generation to experimental reality.
V1.0 scope. 103 AI-for-biology papers spanning protein binder design, antibody design, enzyme design, small-molecule discovery, RNA aptamer design, 5'UTR / mRNA design, DNA regulatory (STARR-seq + lentiMPRA), variant effect prediction, prime editing pegRNA design, semantic gene design via genomic LM, LLM-designed CRISPR effectors, fully autonomous closed-loop protein engineering, peptide antibiotic optimisation, and atom-level enzyme active-site scaffolding. Each paper is decomposed into one or more campaigns (n=60 sub-campaigns total). Multi-campaign papers (e.g., Watson 2023 RFdiffusion, 4 sub-campaigns) are split rather than averaged to preserve target-specific hit rates.
Curation rules. Every quantitative anchor in this tracker is transcript-verified against the original paper's methods/results/supplementary text. No fabricated PDB IDs, no invented candidate counts, no "dozens → 12" translations. Where a specific accession or count is not extracted, the field explicitly states this (e.g., "mark for SI cross-check").
Evidence scale (E0–E5) follows the project schema §9. E5 requires in vivo, translational, or clinical validation — company pipeline claims alone do not qualify. E4 requires functional or orthogonal validation beyond single-assay hit calling.
Denominator transparency (D0–D3) per §12. D3 requires most funnel stages reported; D2 major stages; D1 tested-and-hits only. Denominator gaps are surfaced in the Data Quality view.
AI contribution (A1–A4) per §13. A1 = scoring/ranking. A2 = candidate generation. A3 = generation + selection. A4 = fully autonomous closed-loop (introduced in V0.2 via P026 SAMPLE, expanded in V1.0).
Human intervention (H0–H4) per §14. H4 flags campaigns where post-AI human optimization dominates the final result (e.g., INS018_055 medicinal chemistry).
Voice convention. Audit-first neutral English. Each BioTender note (≤400 characters) states what the paper actually demonstrates, not marketing framing. Warnings are surfaced explicitly rather than embedded in narrative prose.
Openness posture. Model cards flag license status separately from code availability. Model weight restrictions are recorded but do not gate inclusion — landmark carve-out applies to field-defining tools regardless of license (e.g., AlphaProteo, AlphaMissense).
How to use this file. This is a single self-contained HTML file. It works offline under file:// with no server, no CDN, no external assets. All data (103 papers, 60 campaigns, 101 models) is embedded directly in the file. External DOI/PDB/NCT links are provided but not required for the interface to function.
What this is not. This is not a comprehensive review — it is a targeted set of landmark campaigns. Future updates will add virtual-cell / perturbation prediction, more preclinical clinical-stage assets, and additional prime-editor / base-editor benchmarks. This is also not a "model winner" ranking — targets, assays, and denominators across campaigns are not directly comparable, and the project schema §37 explicitly forbids overall success-rate aggregation.
Session: v1.0
Extraction date: 2026-07-13
File format: Single self-contained HTML with inline CSS + inline JS + embedded JSON data payload (~268 KB of curated data). Opens offline via file://.