PAINS: a question of applicability domain?

<< previous || next >>

As most readers of this blog will know, analysis of large (often proprietary) data sets is very much a part of modern drug discovery. Some will have discerned a tendency for the importance of these studies to get 'talked up' and the proprietary nature of many of the data sets makes it difficult to challenge published claims. There are two ways in which data analysis studies in the drug discovery literature get 'talked up'. Firstly, trends in data are made to look stronger than they actually are and this has been discussed. Secondly, it may be suggested that the applicability domain for an analysis is broader than it actually is.

So it's back to PAINS with the fifth installment in the series ( 1 | 2 | 3 | 4 ) and, if you've found reading these posts tedious, spare a thought for the unfortunate person who has to write them. In the two posts on PAINS that will follow this one, I'll explore how PAINS have become integrated into journal guidelines for authors before concluding the series with some suggestions about how we might move things forward. But before doing this, I do need to take another look at the Nature PAINS article (Chemical con artists foil drug discovery) that was discussed in the first post of the series. I will refer this article as BW2014 in this post. I'll use the term 'pathological' as a catch all term in this post to describe any behavior by compounds in assays that results in an inappropriate assessment of the activity of those compounds.   

BW2014 received a somewhat genuflectory review in a Practical Fragments post. You can see from the comments on the post that I was becoming uneasy about the size and 'homogeneity' of the PAINS assay panel although it was a rather intemperate PAINS-shaming post a couple of months later that goaded me into taking a more forensic look at the field. I'd like to get a few things straight before I get going. It has been known from the mid-1990s that not all high-throughput screening (HTS) output smells of roses and the challenge has been establishing by experiment that suspect compounds are indeed behaving pathologically. When working up HTS output, we typically have to make decisions based on incomplete information. One question that I'd like you think about is how would knowing that a catechol matched a PAINS substructure change your perception of that compound as a hit from HTS?

So before I go on it is perhaps a good idea to say what is meant the term 'PAINS' which is an acronym for Pan Assay INterference compoundS. In the literature and blogs, the term 'PAINS' appears to mean one of the following:

1) Compounds matching substructural patterns disclosed in the original PAINS study

2) Compounds that have been demonstrated by experiment to behave pathologically in screening

3) Substructural definitions such as, but not necessarily, those described in the original PAINS article, claimed to be predictive of pathological behavior in screening

4) Compounds that matching substructural definitions such as, but not necessarily, those described in the original PAINS article

5) Compounds (or classes of compounds) believed to have the potential to behave pathologically in screens.

There is still some ambiguity within the categories and, in the original PAINS study, PAINS are identified by frequent-hitter behavior in an assay panel. Do you think that is justified to label compounds that fail to hit a single assay in the panel as PAINS simply because they share substructural elements with frequent-hitters? Category 5 is especially problematic because it can be difficult to know if those denouncing a class of compounds as PAINS are doing so on the basis of relevant experimental observations, model-based prediction or 'expert' opinion. I'd guess that those doing the denouncing often don't know either. Drug discovery suffers from blurring of what has been measured with what has been opined and this post should give you a better idea of what I'm getting at here.

This is a good point to summarize the original PAINS study. Compounds were identified as PAINS on the basis of frequent-hitter behavior in a panel of six AlphaScreen assays for inhibition of protein-protein interactions. The results of the study were a set of substructural patterns and a summary of the frequent hitter associated with each pattern. The original PAINS study invokes literature studies and four instances of  'personal communication' in support of the claim that PAINS filters are predictive of pathological behavior in screening although, in the data analysis context, this 'evidence' should be regarded as anecdotal and circumstantial. Neither chemical structures nor assay data were disclosed in the original PAINS study and the data must be regarded as proprietary.

The PAINS substructural patterns would certainly be useful to anybody using AlphaScreen. My criticism of the 'PAINS field' is not of the substructural patterns themselves (or indeed of attempts to identify compounds likely to behave pathologically when screened) but of the manner in which they are extrapolated out of their applicability domain. I would regard interpreting frequent-hitter behavior in a panel of six AlphaScreen assays as pan-assay interference as a significant extrapolation?

But I have droned on enough so now let's take a look at some what BW2014 has to say:

"Artefacts have subversive reactivity that masquerades as drug-like binding and yields false signals across a variety of assays [1,2]. These molecules — pan-assay interference compounds, or PAINS — have defined structures, covering several classes of compound (see ‘Worst offenders)."

I don't think that it is correct to equate artefacts with reactivity since compounds that absorb or fluoresce strongly or that quench fluorescence can all interfere with assays without actually reacting with anything. My bigger issue with this statement is claiming "a variety of assays" when the PAINS assay panel consisted of six AlphaScreen assays. Strictly, we should be applying the term 'artefact' to assay results rather than compounds but that would be nitpicking. Let's continue from BW2014:

"In a typical academic screening library, some 5–12% of compounds are PAINS [1]."

Do these figures reflect actual analysis on real academic screening libraries? Have these PAINS actually been observed to behave pathologically in real assays or are they simply been predicted to behave badly? Does the analysis take account of the different PAIN levels associated with different  PAINS substructures?  Continuing from BW2014:

“Most PAINS function as reactive chemicals rather than discriminating drugs. They give false readouts in a variety of ways. Some are fluorescent or strongly coloured. In certain assays, they give a positive signal even when no protein is present. Other compounds can trap the toxic or reactive metals used to synthesize molecules in a screening library or used as reagents in assays.”

“PAINS often interfere with many other proteins as well as the one intended."

At the risk of appearing repetitive, it is not clear exactly what is meant by the term 'PAINS' here. How many compounds identified as PAINS in the original study were actually shown by experiment to function as "reactive chemicals" under assay conditions? How many compounds identified as PAINS in the original study were actually shown to "interfere with many other proteins"? How many compounds identified as PAINS in the original study were actually shown to interact with even one of the proteins used in the PAINS assay panel? This would have been a good point to have mentioned that singlet oxygen quenchers and scavengers can interfere with the AlphaScreen detection used in all six assays of the original PAINS assay panel.

BW2014 offers some advice on PAINS-proof drug discovery and I'll make the observation that there is an element of 'do as I say, not as I do' to some of this advice. BW2014 suggests: 

“Scan compounds for functional groups that could have reactions with, rather than affinity for, proteins.”

You should always be concerned about potential electrophilicity of screening hits (I had two 'levels' of electron-withdrawing group typed as SMARTS vector bindings in my Pharma days although I accept that may have been a bit obsessive) but you also need to be aware that covalent bond formation between protein and ligand is a perfectly acceptable way to engage targets. 

The following advice from BW2014 is certainly sound:

“Check the literature. Search by both chemical similarity and substructure to see if a hit interacts with unrelated proteins or has been implicated in non-drug-like mechanisms.”

This is a good point to mention that singlet oxygen quenchers and scavengers can interfere with the AlphaScreen detection used in the six assays of the original PAINS assay panel. I realize it is somewhat uncouth to say so but the original PAINS study didn't exactly scour the literature on quenchers and scavenger of singlet oxygen.  For example DABCO is described as a "strong singlet oxygen quencher" without any supporting references. 

BW2014 makes this recommendation:

"Assess assays. For each hit, conduct at least one assay that detects activity with a different readout. Be wary of compounds that do not show activity in both assays. If possible, assess binding directly, with a technique such as surface plasmon resonance."

Again this makes a lot of sense and I would add that sometimes pathological behavior of compounds in assays can be discerned by looking at the concentration response of signal. Direct (i.e. label-free) quantification is particularly valuable and surface plasmon resonance can also characterize binding stoichiometry which can be diagnostic of pathological behavior in screens. However, the above advice begs the question why a panel of six assays with the same readout was chosen for a study of pan assay interference.   

I'll finish off with some questions that I'd like you to think about. Would you consider a compounds hitting all assays in a panel composed of six AlphaScreen assays to constitute evidence for pan assay interference by that compound? Given the results from 40 HTS campaigns, how would you design a study to characterize pan-assay interference? How would knowing that a catechol was an efficient quencher of singlet oxygen change your perception of that compound as a hit from HTS?

So now that I've distracted you with some questions, I'm going to try to slip away unnoticed. In the next PAINS post, I'll be taking a close look at how PAINS have found their way into the J Med Chem guidelines for authors. Before that, I'll try to entertain you with some lighter fare. Please stay tuned for Confessions of a Units Nazi... 

Yet more on ligand efficiency metrics

In this post, I'll be responding to a couple of articles in the literature that cited our gentle critique of ligand efficiency metrics (LEMs). The critique has also been distilled into a harangue and  readers may find that a bit more digestible that the article. As we all know, Ligand Efficiency (LE), the original LEM was introduced to normalize affinity with respect to molecular size. Before getting started, I'd like to ask you, the reader, to ask yourself exactly what you take to mean by the term 'normalize'.

The first article which I'll call L2016 states:

Optimisation frequently increases molecular size, and on average there is a trade-off between potency and size gains, leading to little or no gain in LE [42,52] but an increase in SILE [52]. This, and the nonlinear dependence of LE on heavy atom count, together with thermodynamic considerations, has led some authors to question the validity of LE [76,77], while others support its use [52,78,79].

This statement is misleading because the "thermodynamic considerations" are that our perception of efficiency changes when we change the concentration units in which affinity and potency are expressed.  As such, LE is a physicochemically meaningless quantity and, in any case, references 52 and 78 precede our challenge to the thermodynamic validity of LE (although not an equivalent challenge in 2009). Reference 78 uses a mathematically invalid formula for LE when claiming to have shown that LE is mathematically valid and reference 79 creates much noise while evading the challenge. I have responded to reference 79 (aka the 'sound and fury article') in two blog posts ( 1 | 2 ).

This is a good place for a graphic to break up the text a bit and I'll use the table (pulled from an earlier post) that shows how our perception of ligand efficiency changes with the concentration units used to define affinity. I've used base 10 logarithms and dispensed with energy units (which are often discarded) to redefine LE as generalized LE (GLE) so that we can explore the effect of changing the concentration unit (which I've called a reference concentration). Please take special of note how a change in concentration unit can change your perception of efficiency for the three compounds. Do you think it makes sense to try to 'correct' LE for the effects of molecular size? 

Another article also cites our LEM critique.  Let's take a look at how the study, which I'll call M2016, responds to our criticism of LE (reference 69 in this study):

The appeal of LE and GE is in the convenience and rapidity with which these factors can be assessed during lead optimization, but the simplistic nature of these metrics requires an understanding of, and appreciation for, their inherent limitations when interpreting data.[67,68,69,70] The relevance of LE as a metric has been challenged based on the lack of direct proportionality to molecular size and an inconsistency of the magnitude of effect between homologous series, both attributed to a fundamental invalidity underlying its mathematical derivation.[65,67] These criticisms have stimulated considerable discussion and provoked discourse that attempts to moderate the perspective and provide guidance on how to use LE and GE as rule-of-thumb metrics in lead optimization.[68,69,70]

To be blunt, I don't think that the M2016 study does actually respond to our criticism of LE as a metric which is that our perception of efficiency changes when we change the concentration unit with which we specify affinity or potency. This is an alarming characteristic for something that is presented as a tool for decision making and, if it were a navigational instrument, we'd be talking about fundamental design flaws rather than "limitations". The choice of 1 M is entirely arbitrary and selecting a particular concentration unit for calculation of LE places the burden of proof on those making the selection to demonstrate that this particular concentration unit is indeed the one that is most fit for purpose. 

The other class of LEM that is commonly encountered is exemplified by what is probably best termed lipophilic efficiency (LipE).  Although the term LLE is more often used, there appears to be some confusion as to whether this should be taken to mean ligand-lipophilicity efficiency or lipophilic ligand efficiency so it's probably safest to use LipE. Let's see what the M2016 study has to say about LipE:

LLE is an offsetting metric that reflects the difference in the affinity of a drug for its target versus water compared to the distribution of the drug between octanol and water, which is a measure of nonspecific lipophilic association.[69,12]

If I knew very little about LEMs, I would find this sentence a bit confusing although I think that it is essentially correct. We used (and possibly even introduced) the term 'offset' in the LEM context to describe metrics that are defined by subtracting risk factor from affinity (or potency). This is in contrast to LE and its variations which are defined by dividing affinity (or potency) by molecular size and can be described as scaled. There is still an arbitrary aspect to LipE in that we could ask whether (pIC50 - 0.5 ´ logP) might not be a better metric than  (pIC50 - logP).  Unlike LE, however, LipE is a quantity that actually has some physicochemical meaning, provided that the compound in question binds to its target in an uncharged form. Specifically, LipE can be considered to quantify the ease (or difficulty) of moving the compound from octanol to its binding site in the target as shown in the figure below:

Let's see what M2016 study has to say:

However, care needs to be exercised in applying this metric since it is dependent on the ionization state of a molecule, and either Log P or Log D should be used when appropriate.

This statement fails to acknowledge a third option which is that there may be situations in which  neither logP nor logD is appropriate for defining LipE. One such situation is when the compound binds to its target in a charged form. When this is the case, neither logP nor logD quantifies the ease (or difficulty) of moving the bound form of compound from octanol to water. As an aside, using logD to quantify compound quality suggests that increasing the extent of ionization will lead to better compounds and I hope that readers will see that this is a strategy that is likely to end in tears.

Let's take a look at LEMs from the perspective of folk who are working in lead optimization projects or doing hit-to-lead work. Merely questioning the value of LEMs is likely incur the wrath of Mothers Against Molecular Obesity (MAMO) so I'll stress that I'm not denying that excessive lipophilicity and  molecular size are undesirable. We even called them "risk factors" in our LEM critique. That said, in the compound quality and drug-likeness literature, it is much more common to read that X and Y are correlated, associated or linked than to actually be shown how strong the correlation, association or linkage is. When you do get shown the relationship between X and Y, it's usually all smoke and mirrors (e.g. graphics colored in lurid, traffic light hues). When reading M2016 you might be asking why can't we see the relationship between PFI and aqueous solubility presented more directly (or even why iPFI is preferred over PFI for hERG and promiscuity). A plot of one against the other perhaps even a correlation coefficient? Is it really too much to ask?

The reason for the smoke and mirrors is that the correlations are probably weak. Does this mean that we don't need to worry about risk factors like molecular size and lipophilicity? No, it most definitely does not! "You speak in more riddles than a Lean Six Sigma belt", I hear you say, "and you tell us that the correlations with the risk factors have been smoked and mirrored and yet we still need to worry about the risk factors".  Patience, dear reader, because the apparent paradox can be resolved once you realize some much stronger local correlations may be lurking beneath an anemic global correlation. What this means is that potencies of compounds in different projects (and different chemical series in the same project) may respond differently to risk factors like lipophilicity and molecular size. You need to start thinking of each LO project as special (although 'unique' might be a better term because 'special projects' were what used to happen to senior managers at ICI before they were put out to pasture).  

Another view of LEMs is that they represent reference lines. For example, we can plot potency against molecular size and draw a line with positive slope from a point corresponding to a 1 M IC50 on the potency axis and say that all points on the line correspond to the same LE. Analogously, we can draw a line of unit slope on a plot of pIC50 against logP and say that all points on the line correspond to the same LipE.  You might be thinking that these reference lines are a bit arbitrary and you'd be thinking along the right lines. The intercept on the potency axis is entirely arbitrary and that was the basis of our criticism of LE. A stronger case can be made for considering  a line of unit slope on a plot of pIC50 against logP to represent constant LipE but only if the compounds bind in uncharged forms to their target.

Let's get back to that project you're working on and let's suppose that you want to manage risk factors like lipophilicity and molecular size. Before you calculate all those LEMs for your project compounds, I'd like you to plot pIC50 against molecular size (it actually doesn't matter too much what measure of molecular size you use). What you now have in front of you is the response of potency to molecular size.  Do you see any correlation between pIC50 and molecular size? Why not try fitting a straight line to your data to get an idea of the strength of the correlation? The points that lie above the line of fit beat the trend in the data and the points that lie below the line are beaten by the trend. The residual for a point is simply the distance above the line for that point and its value tells you how much the activity that it represents beats the trend in the data. Are there structural features that might explain why some points are relatively distant from the line that you've fit? In case you hadn't realized it, you've just normalized your data. Vorsprung durch technik! Here's a graphic to give you an idea how this might work. 

The relationship between affinity and molecular size shown in the plot above is likely to be a lot tighter than what you'll see for a typical project. In the early stages of a project, the range in activity for the project compounds will often be too narrow for the response of activity to risk factor to be discerned. You can make assumptions about the response of affinity (or potency) to risk factor (e.g. that LipE will remain constant during optimization) in order to forecast outcome but it's really important to continually monitor the response of activity to risk factor to check that your assumptions still hold. If affinity (or potency) is strongly correlated with risk factor then you want the response to risk factor to be as steep as possible. Could this be something to think about when trying to prioritize between series?  

So it's been a long post and there are only so many metrics that one can take in a day. If you want to base your decisions on metrics that cause your perception to change with units then as consenting adults you are free to do so (just as you are free to use astrological charts or to seek the face of a deity in clouds). A former prime minister of India drank a glass of his own urine every day and lived to 98. Who would have predicted that? Our LEM critique was entitled 'Ligand efficiency metrics considered harmful' and I now I need to say why. When doing property-based design, it is vital to get as full an understanding as possible of the response of affinity (or potency) to each of the properties in which you're interested. If exploring the relationship between X and Y, it is generally best to analyse the data as directly as possible and to keep X and Y separate (as opposed to looking at the response of a function of Y and X to X). When you use LEMs you're also making assumptions about the response of Y to X and you need to ask yourself whether that's a sensible way to explore the response of Y to X. If you want to normalize potency by risk factor, would you prefer to use the trend that you've actually observed in your data or an arbitrary trend that 'experts' recommend on the basis that it's "simple"?

Next week, PAINS...