PAINS, REOS and structural alerts

PAINS is the most cited and most misused filter in medicinal chemistry. This module is about where the substructures came from, what they can honestly support, and what to do with a flagged compound other than delete it.

The provenance is the whole story, so it is worth stating exactly.

Baell and Holloway ran a particular library of about 93,000 compounds through six AlphaScreen assays against protein-protein interaction targets. They looked for substructures that appeared far more often in hits than chance allowed, and published 480 of them as Pan-Assay Interference Compounds.

Every word in that sentence is a limit on what follows:

• One library. Substructures absent from it could not be found. The set is not a catalogue of all problematic chemistry.
• One assay technology. AlphaScreen is a bead-based luminescent proximity assay with specific vulnerabilities — singlet oxygen chemistry, light interference — that other formats do not share.
• One target class. Protein-protein interactions have shallow, hydrophobic sites and are notoriously hard, so the hit rate for genuine binders was low and the artefact fraction correspondingly high.
• A frequency argument, not a mechanism. The method identifies substructures that hit often. For many of them the mechanism was later worked out; for some it was not.

None of that makes PAINS wrong. It makes it evidence of a particular kind, and treating a frequency observation from one assay format as a chemical verdict is where the misuse begins.

It helps to know that these sets are not all trying to do the same thing.

• REOS (Rapid Elimination Of Swill) — a set of substructure and property filters aimed at removing reactive, unstable and generally unpleasant compounds before screening. Its intent is library curation, applied before any assay is run.
• Brenk — substructures associated with toxicity, instability and poor pharmacokinetics, assembled for fragment libraries.
• In-house alert sets — every large organisation has one, encoding its own painful history. These are often the most useful and the least transferable.

So a compound flagged by REOS is being called chemically unattractive; one flagged by PAINS is being called likely to hit in assays for the wrong reason. Different claims, different responses.

This is the part that makes the filter usable, because each mechanism has an experiment attached. A flag without a mechanism is a rumour; a flag with one is a hypothesis you can test in an afternoon.

• Colloidal aggregation. The compound forms particles above a critical concentration and sequesters enzyme non-specifically. Probably the single commonest cause of false positives in screening. Test: add 0.01% Triton X-100. Aggregation-based inhibition disappears.
• Redox cycling. The compound plus a reducing agent generates hydrogen peroxide, which inactivates the enzyme. Test: add catalase, or remove the reducing agent.
• Fluorescence interference. The compound absorbs or emits at the assay wavelength, or quenches the signal. Test: change the readout, or measure the compound's own spectrum.
• Singlet oxygen production. Specific to AlphaScreen and related chemistry. Test: run an orthogonal assay format.
• Non-specific covalent reactivity. The compound modifies the protein. Test: mass spectrometry, or check for time-dependent inhibition.
• Chelation. The compound sequesters a metal the enzyme needs. Test: vary the metal concentration.

Notice that all six are cheap. That is the argument for investigating rather than deleting: an afternoon of orthogonal experiments settles what a filter can only guess at.

The filter errs both ways, and both errors are common.

Flagged compounds that are fine. Many compounds carrying a PAINS substructure behave perfectly well in a different assay format, or against a different target class. The substructure was associated with promiscuity in one context; yours may not be that context.

Approved drugs carry PAINS substructures. Analyses have found a meaningful fraction of marketed drugs matching one or more PAINS filters. Curcumin is the standing example of the opposite kind — a compound that trips several alerts, has been studied in enormous numbers of papers, and for which the alerts were essentially correct all along.

Clean compounds that are trouble. The filter sees the structure you drew. It cannot see what metabolism will make of it — module 5's paracetamol point — and it cannot see aggregation, which depends on physical properties rather than substructure.

There is a substantial literature criticising naive PAINS use, and the criticisms are worth knowing because they are largely correct.

• The set was derived from one library and one assay technology, and is applied far outside that domain.
• Many of the 480 substructures matched very few compounds in the original analysis, so the statistical support for individual filters varies enormously.
• Papers routinely cite a PAINS match as though it were proof of interference, without running a single orthogonal experiment.
• Reviewers reject papers on a PAINS match alone, which has the effect of removing whole chemotypes from the literature.

Baell himself has written repeatedly that the filters are meant to trigger scrutiny, not rejection. The disagreement in the field is mostly about how people use the tool, not about the tool.

A workable procedure, in order.

• Look at which substructure matched, and where. Some PAINS filters are chemically compelling — a quinone, a catechol, a rhodanine. Some match innocuous fragments of larger molecules.
• Ask which mechanism it implies, then run that specific test. Detergent for aggregation, catalase for redox, an orthogonal readout for interference.
• Check the dose-response curve. A steep Hill slope, or inhibition that appears abruptly above a threshold concentration, is characteristic of aggregation.
• Check whether there is a structure-activity relationship. Real binders have one: small changes produce sensible changes in potency. Artefacts usually do not.
• Get an orthogonal measurement. Surface plasmon resonance, calorimetry or a crystal structure settles it.
• Then decide — and write down why.

The flag is the beginning of an investigation, not the end of one. That is the entire module, and it is the same sentence that closed module 5.