| Baell and Holloway (2010) | Launched drugs in MDDR 2008.1 | 1,808 | 5.0% |
| Baell and Holloway (2010) | Australian Medicines Handbook 2008 | 743 | 48 (6.5%) |
| Capuzzi et al. (2017) | Small-molecule drugs in Drugs@FDA | 1,460 | 87 (6.0%) |
| ProteinIQ (2026) | Approved small molecules in ChEMBL 37 | 2,104 | 122 (5.8%) |
The flagged drugs rarely carry more than one problem: 114 of the 122 match a single filter, and only 31 of the 480 filters match any approved drug at all. If we keep only families A and B, as Baell and Nissink recommend for triage, 103 drugs (4.9%) remain. Oral drugs are flagged slightly less often than the rest, 5.4% against 6.4%.
The alerts cluster in a few corners of pharmacology. Antineoplastic agents have the highest rate among the larger drug classes, 22 of 177 (12.4%), followed by cardiac therapy with 6 of 51, mostly catecholamines such as dobutamine. Not one of the 35 drugs for diabetes or the 36 antiepileptics triggers an alert.
](/data/guides/pains-filters/summary.json).')
## Which approved drugs contain PAINS substructures?
Catechols lead the list with 24 drugs, and most of them are catechols because their targets are. Levodopa, carbidopa, methyldopa, dopamine, norepinephrine, epinephrine, dobutamine and isoproterenol all mimic or feed the body's catecholamines, and entacapone, tolcapone and opicapone inhibit the enzyme that degrades them. The antibiotic cefiderocol carries a catechol for a different reason: it binds iron, so bacteria pull the drug in through their iron-uptake channels.
Quinones come third with 16 drugs, and here the alert and the pharmacology overlap. The anthracyclines doxorubicin, daunorubicin, epirubicin, idarubicin and valrubicin are quinones, as are mitomycin, mitoxantrone, atovaquone and vitamin K. Capuzzi and colleagues found doxorubicin active in about 85% of the more than 4,000 PubChem assays that tested it, so a drug can be both a genuine medicine and a genuine pan-assay interferer.
Other matches are deliberate chemistry. Sulfasalazine, balsalazide and olsalazine are azo compounds because the azo bond is the delivery system: gut bacteria cleave it to release mesalazine in the colon. Epalrestat, an aldose reductase inhibitor approved in Japan for diabetic neuropathy, is the only ene-rhodanine in the set. Eltrombopag, flagged here as an imine ketone, began as an azo-containing screening hit.
Some flags say more about the filter than the drug. Zidovudine matches the azo filter because the pattern also fits its azide group. The 23 dialkylanilines include linezolid, ketoconazole, posaconazole and mifepristone, where the flagged nitrogen is an ordinary aryl morpholine, aryl piperazine or dimethylaniline. Baell and Nissink suggest that aniline-based PAINS may interfere with AlphaScreen specifically, probably by reacting with the singlet oxygen the assay uses to make its signal.
None of this means a PAINS hit from a screen is safe to pursue. Baell and Nissink point out that the PAINS-containing drugs were mostly found by observing efficacy downstream, not by target-based screening, and that their existence "has no relevance to low micromolar screening hits". The logic cuts both ways: an alert does not condemn levodopa, and levodopa does not clear a rhodanine that inhibits five unrelated enzymes.
## Has the share changed over time?
Drugs first approved in the 1950s and 1960s trigger PAINS alerts most often, 7.9% and 8.4%, and the share drifts down to 5.6% in the 2000s and 2010s and 4.6% for 2020 to 2024. None of the 43 drugs first approved before 1950 is flagged. Across the 1,805 drugs with a recorded approval year, 6.5% of those approved before 2000 carry an alert, against 5.4% of those approved since.
](/data/guides/pains-filters/summary.json).')
The decline is small, and the decades are too small to prove it. The 95% intervals for the two halves, 5.2% to 8.1% before 2000 and 3.9% to 7.4% since, overlap. The drugs approved since 2010 also came from programs that started before the filters existed, so the curve cannot yet show whether screening with PAINS filters changed what reaches the clinic.
What did change is the kind of alert. Catechols and quinones account for 30 of the 75 flagged drugs approved before 2000 but only 7 of the 35 approved since. The newer flags are mostly dialkylanilines, 13 of the 35, in kinase inhibitors such as gilteritinib and antimicrobials such as linezolid and delamanid: the family most likely to reflect AlphaScreen rather than the molecule.
## Why PAINS filters over- and under-flag
The filters inherit the limits of the experiment behind them: one library, one assay technology and one class of target. Capuzzi and colleagues counted that 328 of the 480 filters, 68%, were derived from four or fewer compounds, and 190 from a single compound. A pattern seen once in an AlphaScreen campaign is a hypothesis, not a rule.
When they tested the filters against public data, the alerts did not predict promiscuity. In PubChem's AlphaScreen protein-protein interaction assays, 97% of the compounds carrying PAINS alerts were infrequent hitters, and the alerts showed no rise in activity across luciferase, β-lactamase or fluorescence assays either. They also found 109 of the alerts among 3,570 compounds in Dark Chemical Matter, a set of molecules tested extensively and never active.
Jasial, Hu and Bajorath reached a similar view from more than 23,000 extensively tested compounds with PAINS substructures. Their median hit count was two to five across hundreds of assays, and the same substructure appeared in consistently inactive and in frequently active compounds, so the surrounding structure decides whether the motif causes trouble. Even in the original data, 86 of 370 quinones were clean.
The core of the filters holds up better than the tail. Of the 16 family A substructures, 13 were confirmed as problematic in AstraZeneca's screening data and the public Badapple promiscuity tool. Over-flagging is mainly a problem of family C and of filters fitted to AlphaScreen.
The filters also miss real PAINS. Baell and Nissink stress that "a compound not recognized as a PAIN may still be a PAIN". The original library excluded epoxides, aziridines and nitro compounds, so no filter covers them. Classes identified later, such as β-aminoketones, isothiazolones and toxoflavins, are absent, and because the original assays contained detergent, colloidal aggregators were largely suppressed and are not captured.
Broader filters trade the opposite way. The [Brenk filter](/app/brenk-filter) applies 105 rules that Brenk and colleagues wrote to keep mutagenic, reactive, pharmacokinetically poor and assay-interfering groups out of a screening library for neglected diseases. On the same 2,104 drugs it flags 982 (46.7%), led by long aliphatic chains in 222 drugs, isolated alkenes in 116 and N-O single bonds in 110. It also catches 84 of the 122 PAINS drugs.