alex2awesome/lipinski-tacit-knowledge

Lipinski Rule of 5: papers and design-rationale extractions (private)

Medicinal-chemistry and drug-discovery papers that cite or mention Lipinski's Rule of 5, or whose compounds are in ChEMBL 37, with the design decisions their authors state, extracted by Gemma 4 31B with a LoRA adapter (lora_r6g), and each paper's stance toward the Rule of 5. Part of a study of tacit knowledge in science (Stanford; project code in tacit-knowledge-in-science).

Private research dataset. Do not make public. Elsevier rows carry derived labels only (see below), and many open papers are not under a redistributable license (license column).

Contents

config	rows	what
papers	78,178	one row per paper: ids, venue, year, license, ChEMBL roles and Rule-of-5 counts, Ro5 stance, number of decisions
decisions	495,978	one row per extracted design decision
fulltext_open	26,303	open papers only: JATS XML and the rendered text the extractor read

source	papers	with at least one decision	decisions
open (Europe PMC full text)	26,303	21,566	157,104
Elsevier (ScienceDirect TDM)	51,875	49,150	338,874

Elsevier rows

Elsevier full text was obtained under a TDM agreement that keeps article text on Stanford servers. This dataset holds no Elsevier text: no XML, no rendered text, and reason_quote / ro5_quote are null. Every decision instead carries quote_spans ([start, end) character offsets into norm(rendered text), with norm and the renderer in promptopt/common.py and promptopt/elsevier_render.py) and quote_sha1, so quotes can be restored on Stanford servers from the original XML. Derived labels may be kept; published quotations must be short snippets (at most 200 characters) with a DOI link. Elsevier downloads are deleted at the end of the project.

Paper roles (ChEMBL 37)

A compound breaks the rule with 2+ of: MW > 500, ALogP > 5, HBA > 10, HBD > 5 (ChEMBL num_ro5_violations). D1_rule_breaking: at least 5 compounds and at least half break the rule. D4_step_away: the paper's main single-protein target (CYP, hERG and P-gp counter-screens excluded) had under 25% rule-breaking compounds in at least 50 compounds published earlier. role (for ChEMBL-linked papers): step_away (D4), rule_breaking (D1 not D4), inside_rule.

role	open	Elsevier
step_away	291	1,576
rule_breaking	688	3,624
inside_rule	7,639	35,606

Decision fields

change (what the authors changed), compounds (in-paper labels, e.g. "12a"), reason_category, reason_quote (verbatim, checked against the source), mentions_physchem_consequence, framing, evidence, section, paragraph (index into the rendered text's paragraphs), quote_spans, quote_sha1.

reason_category	decisions
potency_or_affinity	112,801
binding_site_interaction	91,142
sar_exploration	64,277
synthetic_accessibility	41,619
adme_pk	38,789
physicochemical_or_druglike	38,513
none_stated	33,768
mechanism_or_modality	30,222
safety_or_toxicity	16,715
selectivity	16,554
novelty_or_ip	11,317
INVALID	261

Extractor quality (held-out, decision-level F1 against reconciled hand labels): 0.71 to 0.73.

Ro5 stance (all papers)

stance	papers
not_mentioned	52,165
applied_as_filter	18,638
acknowledged_violation	4,877
argued_exception	2,498

Open-paper licenses

license	papers
cc-by	11,114
author_manuscript	7,549
none_stated	3,536
cc-by-nc-nd	2,297
cc-by-nc	1,356
cc-by-nc-sa	203
license-type:open-access	142
license-type:editor-choice	55
cc0	33
cc-by-nd	6
license-type:acs-free-to-read	6
cc-by-sa	2
license-type:openaccess	2
cc-by-ncsa	1
license-type:us-gov	1

Sources and licenses

Compound counts and roles: ChEMBL 37 (CC BY-SA 3.0). Open full text: Europe PMC; each paper keeps its own license. Elsevier: ScienceDirect TDM, derived labels only.