AML drug-repurposing research is spread across many drugs, not concentrated in a few

The question

Acute myeloid leukaemia (AML) is an aggressive blood cancer, and older patients have poor outcomes. Researchers often propose repurposing: testing drugs already approved for other conditions (antiparasitics, antipsychotics, diabetes drugs, vitamins) against leukaemia cells. Reviews list dozens of such candidates. Valli et al. (2020) note that none of them has entered standard AML care.

We pre-registered one question: is this preclinical literature concentrated in a few compounds? The hypothesis, fixed before any analysis, was that the top 5 candidates account for at least 50% of AML-repurposing papers in Europe PMC, 2015–2025.

This is a study of publication patterns. It says nothing about whether any drug works in patients, and it is not medical advice.

Method

A single script queried the public Europe PMC API (preprints excluded, records de-duplicated). Of 34,248 AML records from 2015–2025, it built two corpora:

• Strict corpus: AML plus an explicit repurposing word (repurpos*, reposition*, "old drug") in the title, abstract or keywords. 154 hits, 147 kept after filtering.
• Loose corpus: AML and a named drug both in the title, with no repurposing word required. 237 papers kept after removing supportive-care topics such as antifungal prophylaxis and antiemetics.

Each paper was matched against a frozen dictionary of 125 approved drugs with no oncology approval (ATC codes outside L01/L02). Drugs with cancer approvals were excluded, including ATRA, arsenic trioxide, thalidomide analogues and methotrexate. A paper naming several drugs counts once for each.

The primary metric is the share of drug-assigned papers that involve at least one of the five most frequent compounds:

$S_1 = \frac{\#\text{papers mentioning a top-5 compound}}{\#\text{papers assigned any dictionary compound}}$

Uncertainty comes from 2,000 bootstrap resamples. The top 5 is re-chosen in every resample, because picking the "top" from the same data inflates the share. The decision rule was fixed in advance: $S_1 \ge 0.50$ in the strict corpus means supported, below means refuted. If the 95% interval contains 0.50, the result is labelled "not decisive".

Results

• The primary estimate is below the threshold (0.424), but its interval reaches 0.67. On its own, the strict result is not decisive.
• The larger loose corpus gives a tighter estimate (0.409), and its interval excludes 0.50.
• No single drug dominates. The most-studied compound accounts for 12% (strict) or 10% (loose) of assigned papers. The "effective number" of compounds (the inverse of the Herfindahl concentration index) is about 19 in the strict corpus and 22.5 in the loose one. That describes a long tail.

The drugs at the top also differ between the two corpora. Papers that explicitly say "repurposing" favour antipsychotics, anthelmintics and disulfiram. Papers that simply test a known drug against AML favour metformin, valproate, vitamin C, sirolimus and the antimalarial artesunate.

Sensitivity checks

Nine of the ten scenarios fall below 0.50. The exception is the loose corpus with ATRA counted. ATRA is an approved treatment for one AML subtype (acute promyelocytic leukaemia), so it is outside the scope of the question. ATRA alone appears in 50 of the 237 loose papers, which shows how much the definition of "non-oncology" can move this kind of count.

Caveats

• Most strict-corpus papers were not assigned a drug. 114 of 147 (78%) matched no dictionary compound. A spot-check found reviews, oncology drugs and papers that only use the word "repurposing" in passing. If many of these actually study approved non-cancer drugs the dictionary misses, the true picture could differ.
• Ties. Three compounds tie at rank 5 in the strict corpus, so the fifth slot is arbitrary.
• Keyword matching on titles and abstracts only. No full texts were read. A 50-paper precision sample has been written out but not yet reviewed by hand.
• Definitions matter. The dictionary, the boundary of "non-oncology" and the query wording all shape the counts. An earlier exploratory count with slightly different filters gave 37,518 AML papers and 147 repurposing hits. Those figures are context only and are not used in the test.
• Paper counts are not evidence strength. Many papers on one drug can reflect publication fashion, positive-result bias or duplicated cell-line work, not promise. Most of this literature uses cell lines and mouse xenografts.

What runs next

1. Manually review the precision sample and the 114 unassigned strict papers.
2. Extend the dictionary with what that review turns up.
3. Check recall against PubMed's controlled vocabulary (MeSH Drug Repositioning combined with Leukemia, Myeloid, Acute).
4. Re-estimate the top-5 share with a larger, cleaner drug-assigned set.

All numbers above come from analysis.py (seed 20261101). Three independent re-runs of the script reproduced them exactly.