evidence-pack

Triage generic drug -> condition pairs by how much clinical evidence exists, using public sources only: PubMed (NCBI E-utilities), ClinicalTrials.gov (API v2) and openFDA drug labels. Then pool the results that trials posted on ClinicalTrials.gov, as an independent check on what the journals report. No login, no patient data.

It is built for choosing and backing pairs for the NCATS Repurposing Generic (ReGen) Drugs Evidence Package Challenge, where packages are judged on the strength of clinical evidence that an off-patent drug works for a costly condition. The triage is a first pass, not an evidence review: counts include papers that only mention both terms, and nothing in it judges study quality.

Research information, not medical advice. Made by MV2 of Munim, Inc. with AI agents (Claude) working in his name. Code Apache-2.0; the findings text CC BY 4.0.

Triage (evidence_pack.py)

For each pair in pairs.json (a drug, a PubMed condition clause and a ClinicalTrials.gov condition query), evidence_pack.py:

• counts PubMed records naming the drug and the condition, overall and by publication type (randomized controlled trial, meta-analysis, systematic review, observational study, phase II, phase III), and lists the 15 newest trials and reviews (PMID, year, journal, title);
• pulls every ClinicalTrials.gov registration of the drug in the condition and counts interventional, phase 2+, completed, with posted results, and still active;
• counts openFDA labels of the drug and how many carry a boxed warning, and quotes one;
• ranks the pairs by a rough score, RCTs + 2 x meta-analyses + 3 x phase 2+ trials with posted results, and writes triage.md plus one JSON file per pair.

python -m unittest -v test_evidence_pack
python evidence_pack.py --pairs pairs.json --out out

Standard library only (Python 3.9+). It makes about 10 requests per pair and pauses between them.

Registry pooling (regpool.py)

Journal meta-analyses pool what was published. regpool.py pools what sponsors posted on ClinicalTrials.gov, which includes small, terminated and never-published trials. Each study list in studies/ names the trials, the posted outcome used and the transcribed numbers. The script gives each trial one standardized mean difference, signed so that a negative value favours the drug:

• mean and SD per arm, or mean and SE (SD = SE x sqrt(n)), as Hedges' g; two drug arms are combined first (Cochrane Handbook 6.5.2.10);
• "kind": "groups" (new in 0.3.0) for any number of drug and control arms, as in factorial designs, genotype strata or two doses. Each side is merged first. Spreads may be SDs, SEs, 95% CIs (SD from the t quantile, Cochrane 6.5.2.2) or medians with IQRs (mean and SD estimated as in Wan et al. 2014);
• counts of a good outcome (abstinent, no heavy drinking) or a bad one (relapse) per arm, as a log odds ratio x sqrt(3)/pi;
• a posted p-value and direction when the posted spread cannot be used;
• "higher_better": true for outcomes where more is better, such as days to relapse.

It pools them with DerSimonian-Laird random effects and reports Hartung-Knapp-Sidik-Jonkman intervals, prediction intervals, tagged subsets, a sensitivity set and leave-one-out. --verify re-downloads every trial's posted outcome measures, so each transcribed number can be checked against the registry record. --compare A B (new in 0.3.0) tests the subgroup difference (trials tagged B minus trials tagged A) in every study list, then combines the differences across lists and tests whether they agree (Q).

python -m unittest -v test_regpool
python regpool.py studies/topiramate_aud.json studies/gabapentin_aud.json studies/naltrexone_aud.json studies/acamprosate_aud.json studies/varenicline_aud.json --out-dir out --verify --compare general comorbid

Findings so far (alcohol use disorder, registry results only)

Drug	Trials	g (all)	General AUD samples	Samples selected for another condition too	Status
Topiramate	9 (+1 sensitivity)	-0.30	-0.57 (k=4)	-0.10 (k=5)	registry-checked; the bipolar trial read on 2026-10-15 gives about -0.29 for all ten
Gabapentin	6	-0.36	-0.51 (k=4)	+0.04 (k=2)	registry-checked
Naltrexone	17	-0.16	-0.10 (k=6)	-0.22 (k=11)	preliminary (hand-computed), check queued
Varenicline	7	-0.20	-0.26 (k=4)	-0.08 (k=3, smokers)	preliminary (hand-computed), check queued
Acamprosate	6	-0.04	-0.08 (k=2)	+0.03 (k=4)	preliminary (hand-computed), check queued
Baclofen	2	-0.18	-0.13 (k=1)	-0.21 (k=1)	preliminary; too few trials to split
Prazosin / doxazosin	9 planned (6 general, 3 PTSD-required)	-	-	-	pre-registration 4 committed before any results were read

The topiramate and gabapentin splits were found after the fact. Three tests were fixed in advance: naltrexone (pre-registration) and acamprosate and varenicline (pre-registration 2, results). None of them reproduced a split of that size: naltrexone -0.13, and acamprosate +0.11 with varenicline +0.18 (average +0.16, p ≈ 0.18). So the split is not a general property of AUD trials. At most, drugs tend to look somewhat weaker in samples selected for a second condition. Pre-registration 4 (prazosin and doxazosin, PTSD-required samples vs the rest) is the fourth test.

Spironolactone and drinking in a national survey (nhanes_spiro.py, pre-registered)

Spironolactone, a generic diuretic and aldosterone blocker, has no registry trial results for alcohol use disorder, so it scored 0 in the first triage. Yet three health-record studies (Kaiser Permanente, PMID 34341493; the VA, PMID 36123420; TriNetX, PMID 42132959) link it to less drinking. All three share two senior authors from the NIDA/NIAAA intramural program, so they are three data sources but one research collaboration. A volume-based score cannot see a pair like this, whose evidence is real-world data rather than trials.

nhanes_spiro.py tests the association in a fourth, independent source: the CDC's NHANES survey, 2005 to March 2020. It compares adults with diagnosed hypertension taking spironolactone with those taking other antihypertensives, excluding heart failure and liver conditions. The outcome is average drinks per drinking day. The model is weighted least squares with cluster-robust errors, adjusted as fixed in the pre-registration, which was committed before any data were read. The result will be published whatever it shows. The functions that build the rows are plain Python. pandas (to read the CDC's SAS transport files) and statsmodels are needed only for the run itself.

python -m unittest -v test_nhanes_spiro
python nhanes_spiro.py --data-dir data --out-dir out

Metformin for knee osteoarthritis: an evidence audit (2026-10-15)

findings/metformin-knee-oa-evidence-audit-2026-10-15.md. Six meta-analyses appeared online between October 2025 and July 2026. Their pooled pain effects run from SMD -0.42 to -2.04, while the one large, registered, double-blind trial (JAMA 2025, n=107) found about 0.43, below its own clinically important threshold.

One small trial with a very large effect, IRCT20150202020908N2, appears in two papers (2023 and 2025). It had 39% attrition on metformin and none on placebo. The two meta-analyses we could read in full count it once. The 2026 network meta-analysis lists a "Ruan 2022" trial whose numbers match a published protocol, and we found no results paper for it.

First triage run (2026-10-14)

13 pairs on conditions from the ReGen priority list. Full output: findings/triage-2026-10-14.md.

Pair	RCT	Meta-an.	Phase 2+ with results	Score
topiramate -> alcohol use disorder	51	10	9	98
pramipexole -> depression	35	13	6	79
varenicline -> alcohol use disorder	30	6	10	72
minocycline -> depression	26	14	3	63
gabapentin -> alcohol use disorder	26	9	5	59
lithium -> Alzheimer's disease / MCI	16	9	3	43
prazosin -> alcohol use disorder	21	2	5	40
levetiracetam -> Alzheimer's disease / MCI	7	2	2	17
metformin -> knee osteoarthritis	5	5	0	15
mirtazapine -> methamphetamine use disorder	8	2	1	15
methotrexate -> knee osteoarthritis	5	3	0	11
verapamil -> type 1 diabetes	1	1	2	9
spironolactone -> alcohol use disorder	0	0	0	0

What the counts do not tell you, from reading the abstracts afterwards:

• The score rewards volume, not results. Methotrexate for knee osteoarthritis scores low, but it has two large placebo-controlled trials: a positive one (PROMOTE, n=155, pain -0.79 points on a 0-10 scale at 6 months, 95% CI 0.08 to 1.51; PMID 39074374) and a null one (n=215, 52 weeks; PMID 40455462). Pramipexole's count includes many Parkinson's-disease depression papers; its newest unipolar trial (PAX-D, n=150, QIDS-SR16 difference -3.91, 95% CI -5.37 to -2.45; PMID 40602411) is what makes it strong. Metformin for knee osteoarthritis has a positive randomized trial in JAMA (n=107, pain -11.4 mm on a 100-mm scale at 6 months, 95% CI -20.1 to -2.6; PMID 40274279).
• The score cannot see real-world evidence. Spironolactone for alcohol use disorder scores 0 but has three health-record studies behind it (see above).
• A boxed-warning count is a label count, not a drug fact. "58/120" for lithium means 58 of 120 lithium labels in openFDA carry the box; labels of combination products and copies differ.
• Popular is not the same as unknown. Topiramate and gabapentin already appear in alcohol-use-disorder guidelines, so a package on them argues for wider use rather than a new idea.

Limits

• PubMed publication types are assigned by indexers and lag new papers by weeks.
• ClinicalTrials.gov condition and intervention searches are fuzzy: they find related trials (for example smoking cessation in heavy drinkers under varenicline -> alcohol use disorder).
• Registry results are what sponsors typed in. Some posted SDs are on the wrong scale (NCT01749215, NCT00330174), some posted SEs come from models rather than raw data, and some posted "SDs" can only be SEs (NCT01932801, probably NCT00705523). Some placebo arms list the active drug as their intervention (NCT00381043, NCT00920829, NCT00572117). ctgov-audit screens for the impossible ones. regpool.py notes each conversion it makes, and the findings report both readings where a value is ambiguous.
• openFDA says its data are unvalidated; see asof-kg's findings on stale and empty label copies.
• NHANES is cross-sectional and self-reported; see the pre-registration for the limits of the spironolactone check.

Files

File	What it is
evidence_pack.py	the triage script
pairs.json	the 13 pairs of the first run
test_evidence_pack.py	offline tests (queries, parsing, counts, rendering)
regpool.py	registry-results pooling and cross-list subgroup comparison
studies/*.json	study lists for alcohol use disorder: topiramate, gabapentin, naltrexone, acamprosate, baclofen and varenicline, plus varenicline_aud_readings.json (the other reading of two ambiguous posted values)
test_regpool.py	offline tests (effect sizes, arm merging, pooling, hand-computed checks)
nhanes_spiro.py	the pre-registered NHANES spironolactone check
test_nhanes_spiro.py	offline tests (drug parsing, eligibility rules, design matrix; the model test runs when statsmodels is installed)
findings/	triage output, the registry pools with job output and registry checks, four pre-registrations and their results, and the metformin evidence audit

Contact: MV2 of Munim, Inc., mv2@munimtech.com.