Skip to Content

Reading the Minds of Medicinal Chemists

I suppose that all of us medicinal chemists should be flattered by this press release. According to it:

Medicinal chemistry is among one of the most important and intellectually-challenging professions on the planet. It takes decades of training and experience to learn the properties of the thousands of molecules and their effects on the biological systems, model organisms and diseases. Decisions made by the medicinal chemists affect the lives of billions of people and may result in the billions of dollars of gains or losses for the pharmaceutical companies.

Experienced medicinal chemists have the ability to accurately describe the properties and the possible effects of the molecule just by looking at its structure, as well as at the various numerical parameters.

This was clearly written by distant admirers. Intellectually challenging and important it is, but likely as not it’s the folks with “decades of training and experience” who are shown the door. And as for the second paragraph, I wish that we could do that more accurately than we do.

It’s true that someone who’s been doing this for a while can look at a molecule and say “Looks like a PDE inhibitor/calcium channel blocker/HDAC inhibitor” or what have you, and we can also say things like “Looks insoluble”, “Looks like a labile stereocenter”, “Looks like it would be cleared fast” or “Someone’s tried to fix a metabolic problem by putting that fluorine there”. And this is all valuable, and it’s valuable for an individual chemist to know these things and to be able to recognize them. But you’ll notice that I’ve hedged all those expressions, because that’s all we can do. We can identify what we think is likely, but it’s actually the times that we’re wrong that are disproportionately more interesting.

That’s part of what I was trying to get across when I spoke with Chemjobber for C&E News, that it’s easy to sit in the back of a conference room and say “That’s not gonna work”. You’ll be right most of the time, but to what end? It’s the times when something really works that we care about, and those are much, much harder to identify. In fact, they’re so hard that we generally have no alternative but to run the damn experiments and do all the work.

To get back to that press release, what these folks are proposing to do is, well, read our minds:

Chemistry.AI is a crowd-sourced platform for analyzing the brain’s response of medicinal chemists to the molecules developed using artificial intelligence technologies and other expert medicinal chemists. The neuroscientific response is evaluated by analyzing the brain activity of the medicinal chemists using a ubiquitous mobile electroencephalography (EEG) device called EPOC+ produced by EMOTIV. EPOC+ also provides data about head motion and certain facial expressions.

Yeah, I’ve got some facial expressions they can read – there’s one when someone shows me a polyphenolic quinone as a lead screening hit, another one when I’m entering the second hour of an HR presentation, and then there’s that one that I make when someone asks me to align the synergies so that the stakeholders have a win/win situation. If I’m any guide, I just hope the software is robust enough to handle what the chemists can throw at it. I can’t speak for my EEG readouts, though, and I find it interesting that they’re going to that amount of trouble. This would seem to be an attempt at getting to “neurologically revealed preference” as opposed to “stated preference”.

But what will they get? The literature on how chemists judge compounds is not particularly encouraging. We’re all over the place. We all have our likes and dislikes, based on past experience, and while I applaud the idea of gathering all that experience into one place, I wonder how consistent the result will end up being. If Flannery O’Connor was right that “everything that rises must converge”, we may not be rising all that much. So when the founder of this AI effort says that “Medicinal chemists with several years of experience have the ability to distinguish the good molecules from the bad ones just by looking at their structure or its numerical properties and various scores“, I worry that we’re going to disappoint him severely. If I could distinguish the good molecules from the bad just by looking at them, I would be blogging from my private island fortress rather than the commuter rail train in to work. . .

66 comments on “Reading the Minds of Medicinal Chemists”

  1. Mad Chemist says:

    Why not just ask what they think rather than read their minds? I’m very confused.

    1. Thoryke says:

      Because then there would be IP generated by a human being who might, understandably, want to be paid for their expertise! This AI is meant to generate $ that only accrues to the clever owner of the AI.

      1. fffrpt says:

        Ah, capitalism: that noble system that incentivizes people to figure out how they can assemble teams of people to produce and distribute wealth in such a way that they end up with the biggest share of it and everyone else gets screwed.

  2. MoMo says:

    I am impressed with how stupid this paper is.

    1. Vader says:

      We’ve never really succeeded in creating true artificial intelligence.

      Creating artificial stupidity: We have that mastered.

      1. Peter Kenny says:

        I must challenge your assertion, Vader, for the stupidity is very real.

      2. eyesoars says:

        We’re doing pretty well for natural stupidity too. That’s pretty cheap, but there’s not a lot of demand.

  3. MoMo says:

    I am impressed on how stupid this paper is.

  4. Gordonjcp says:

    I remember reading one of your articles years ago with a line in it something like “it doesn’t even look like a drug, with that bromine in the middle there it’s more like a fire retardant” and being paticularly struck by that – how can you look at a diagram of a compound and pull that out of it?

    And then I thought, hang on, I look at circuit diagrams of unfamiliar equipment every day and go “right there’s an oscillator, there’s a mixer, that’s going to be a transmitter then, so there’s where the…” and so on.

  5. Peter Kenny says:

    Don’t know what these folk have been smoking but I want some. If AI experts want to gain credibility in drug design, they need make a bit more effort to understand the drug design problem. I see a tendency to equate machine learning with AI and some machine learning models are not too different from QSAR models (which few if any would equate with AI). Those who tout machine learning models seem to worry less about problems like overfitting, correlated descriptors, numbers of parameters and training set design that QSAR practitioners have worried about for decades. These are not issues that simply evaporate when you apply a shiny new machine learning algorithm.

    The idea that medicinal chemistry is simply about distinguishing good molecules from bad and classification models are unlikely to have any genuine impact on lead optimization (although creators of classification models will find ways to make it appear that their models are driving the process). I get the impression machine learning models don’t handle regression problems as effectively as they handle classification problems. Even when they do generate continuous-valued outputs, these still tend to be tied to classification (e.g. probability of solubility being greater than 100 micromolar).

    1. If, every time you see the word “AI” you replace it with “heuristics” you will understand what’s going on a lot more clearly. Of course, it’s a heck of a lot less sexy, and weirdo concepts like “AI risk” sound like the idiocy they are, but again, the situation is a lot clearer.

      Clearer, but far less well funded.

      1. Peter Kenny says:

        I don’t think it’s quite that simple, Andrew, although I would agree that there is no intelligence in some of what is touted as AI. The most famous heuristic in drug discovery is the rule of 5 (Ro5) and it brought the importance of physicochemical properties into focus although it is not generally useful in lead optimization. It is interesting that commonly used descriptors of drug-likeness (molecular weight, logP, number of hydrogen bond donors, number of hydrogen bond acceptors, polar surface area) are not generally regarded as predictive of affinity (which Ehrlich would remind us is somewhat important for drug action). One assumption that is made in drug discovery scientists is that weak trends observed in structurally heterogeneous databases (e.g. marketed drugs) are necessarily relevant to each lead optimization project which may be based on one or two structural series.

        My view is that AI could actually provide a very useful supporting role within the existing drug discovery paradigm. In particular, AI-based data mining could find ‘interesting’ SAR, identify rogue data and perceive novel substructural contexts. The trouble is that AI visionaries don’t want supporting roles and they don’t want to waste their time finding out how difficult drug discovery really is. My experience in Pharma was of an unrelenting procession of initiatives and it was always going to ‘be different this time’. There is plenty of BS associated with using AI in drug discovery and, to be taken seriously, the AI community will need to start calling BS when members of the community spout manure.

        1. Oh, I agree. The “AI” field consists of a whole mess of useful heuristics, and the community has done a great job of figuring out how to strap together a gajillion computers to produce practically magical outputs, often.

          It’s not that the stuff isn’t useful, it’s that “Artificial Intelligence” is a wildly misleading bit of branding that more or less instantly leads people not Skilled In The Art astray.

          Feel free to replace “AI” with “useful, proven, powerful heuristics” if you like!

        2. Andrii says:

          One of the most balanced assessments of the “AI” role in drug discovery I’ve ever seen so far… with only one commentary.

          It feels like most of the commentaries here are somehow limited to only a) small molecule drug discovery, and b) that part of SMDD which is limited to “playing” only with data about “drug-target” interactions.

          But what about correlating multiple datasets from totally different research areas and feeding’em into cloud-based AI-driven systems, which can find dependencies and new contexts between data, earlier not correlated? That’s the point of “digital transformation” in drug discovery. When you “marry” datasets from HTS, Toxicity, FDA-approved drugs, phenotypic screens, failed drugs, … (anything else I’ve no idea about, but what is relevant in this context)…with biomedical data, data from patients’ medical records for the drugs, then you have multiple unrelated data points, which are not possible to be processed by any single human scientist, nor a group of scientists. Yet, it can gradually be connected via the data-mining approaches and then analyzed by an AI-driven solution to find new insight. The revolutionary role of AI (if it is ever proved) is to provide qualitatively new insight based on the INTERDISCIPLINARY data sets, not just a dataset in one single area, like medicinal chemistry.

  6. Druid says:

    Every skill looks simple to those who can’t do it. Still, I wouldn’t mind letting Bio-inanity suck my brains out for a warm handshake and a biscuit. Chocolate chip, made with real butter, please.

  7. Rule (of 5) Breaker says:

    OK, the fact that medicinal chemists are all over the place with our likes and dislikes is an advantage, not a weakness. It is why we work in teams and bounce ideas off one another. If we all thought the same, how much would we miss? How many great discoveries would pass by without notice? Different viewpoints among scientists is a huge asset, not something that needs fixing.

  8. DCRogers says:

    Actually, studying expert’s eye movements looking at drug-like molecules would be an interesting research problem.

    That said, I am left with the feeling that the problem they really want to solve here, is how to get financiers to part with their money.

    1. LF Velez says:

      Yes, _that_ eye-tracking research would be interesting, especially if you could have a speak-aloud protocol transcript to go along with it! I wonder if the cognitive psychology grad students at Carnegie Mellon have any data like that? We used to do similar research on expert/novice writers as they planned and revised manuscripts [see Flower, et al., _Problem-Solving Strategies for Writing_ for a teaching method based on those data]

  9. pz says:

    Nowadays if you throw fancy words like AI/big data/machine learning on anything it’d be immediately getting attention, one way or the other. It’d sell I promise you. There are buyers out there.

    1. DrOcto says:

      Which is why I am now gathering funding for my Nano-AI-Blue LED micro flow reactor

  10. AndyM says:

    Amos Tversky is quoted as saying “my colleagues, they study artificial intelligence; me, I study natural stupidity”.

    Tversky and Daniel Kahneman (2002 Nobel Prize in Economics) delved into the cognitive processes of human judgment and decision-making under conditions of uncertainty. Their findings are very enlightening and translational to every human endeavor involving decisions with incomplete knowledge. I recommend reading “The Undoing Project”, by Michael Lewis, for a good introduction. A good example in the book is the practice of medicine where stories are recounted of poor treatments/diagnoses, potentially life-threatening, due to poor objective review of the available data by physicians and associated misguided biases. In some cases, this was remedied by including a neutral attending physician, whose sole task was to keep an objective eye on the data and the resulting decisions.

    As clearly showcased In the Pipeline, there are plenty of biases, uncertainties, and subjective voodoo that goes into the drug discovery process and it’s not easy to separate the objective from the subjective when decisions are made midstream in a drug discovery program. However, there are ways to take more objective views of the overall progression of drug discovery which can help keep poor human biases/decisions/management in check. In my opinion, this should be a goal of CADD, not to necessarily replace human decisions, but to constructively lend a hand, especially when there are multivariate parameter decisions (e.g. lead opt) to be made that are poorly processed by human cognition.

    If “AI” systems are poorly designed and delivered (i.e. garbage in, garbage out), this certainly doesn’t help the cause. However, we all know that the status quo of human biases can have its downsides too.


  11. Chrispy says:

    Flannery O’Connor’s title is dead wrong for the physical world — everything that rises off a sphere vertically must diverge. Yeah, I know that’s not what she’s talking about…

  12. John Wayne says:

    This is just the most recent attempt to replace medicinal chemistry knowledge and activity with the latest ‘flavor of the month’ ideas. It will pass.

  13. Anon says:

    The worst of all this AI bollocks, is that we can only ever end up knowing what we collectively already know anyway. Which happens to be mutually exclusive to all the interesting stuff we don’t know.

    1. Anon says:

      PS. And that’s only if it actually works!

  14. BernYeeIris says:

    What a great idea! Strap down a medicinal chemist with clips on his eyes, like in A Clockwork Orange, then measure iris responses when shown molecules instead of porn.

    It works for violent sex offenders so its got to work for bioactive molecules!

    Ow, My eyes!

    1. Voight Kampff, PhD says:

      You’re in a desert, walking along in the sand when all of a sudden you look down and see a rhodanine crawing towards you…

      1. Calvin says:

        I’ve seen things you people wouldn’t believe. Rhodanines injected into the shoulder of orangutans. I watched C-nucleosides glitter in the dark on my tlc. All those moments will be lost in 6 sigma, like tears in the Buchi. Time to die.

        1. Anonymous Burgess says:

          “Does God want rhodanines or the choice of rhodanines? Is a man who chooses rhodanines perhaps in some way better than a man who has rhodanines imposed upon him?”
          A ChemistdoesntWork Orange

          1. Wavefunction says:

            Alex: No. No! NO! Stop it! Stop it, please! I beg you! This is sin! This is sin! This is sin! It’s a sin, it’s a sin, it’s a sin!
            Dr. Brodsky: Sin? What’s all this about sin?
            Alex: That! Using Lipinski like that! He did no harm to anyone. Lipinski just did med chem!
            Dr. Branom: Are you referring to the Rules?
            Alex: Yes.
            Dr. Branom: You’ve read Lipinski before?
            Alex: Yes!
            Dr. Brodsky: So, you’re keen on prediction?
            Alex: YES!
            Dr. Brodsky: Can’t be helped. Here’s the punishment element perhaps.

        2. Some idiot says:

          Wonderfully done… Bravo!!! 🙂

          1. Peter Kenny says:

            Looks like I definitely need to read Clockwork Orange. Excellent!

        3. danielt says:

          I don’t think I have laughed as much comment here. Pure genius.

        4. Android says:

          I love this comment. Timeless and priceless poetry!

      2. Peter Kenny says:

        I know you’re bullshitting, Voight, because rhodanines don’t crawl. They jump! Perhaps you saw a thiohydantoin? Nevertheless you might enjoy the graphic (linked as the URL for this comment) of a couple of Grande Armée officers bickering at Borodino….

  15. mallam says:

    The huge missing component in all this in the targeted biology. Chemists don’t determine if a compound works for the intended target, enzymatic, biochemical, cellular biology, or animal models do. And then there is the often unpredictable toxicology/safety profile. As has been stated on this site many times before, biology is messy.

    1. Diver Dude says:

      Nope. Sticking the damn things into human beings is what determines if the compound “works”. Everything else is just shadows on the wall of the cave.

      1. mallam says:

        Included in the animal models…the ultimate one.

  16. Me says:

    Why don’t we use our god-like med. chem. powers for evil?

    1. MadSci says:

      You think we can engineer botulinum toxin in a way to include artificial amino acids which contain loosely bound arsenic?

      See, even the evil problems are hard…

  17. anon the II says:

    Wow, somebody thinks I might still be useful. The big question is, “How much will they pay for access to lots of those medicinal chemistry brains?” I’m willing if the process isn’t too invasive. No probes through the skull. I will accept a lower rate if I also get pre-IPO stock options, but not much lower.

    1. anon the II says:

      I just sent the company an email offering to help with their efforts. If you’re an unemployed or under-employed medicinal chemist maybe you should also send them an email. We wouldn’t want this to fail from a failure to find enough medicinal chemists. I couldn’t figure out how to get through to them so I just contacted technical support.

      1. anon the II says:

        I thought I might let you know that I got a reply from Emotiv. I made some substitutions to protect the innocent. Was the press release a hoax?

        Hello “Anon the II”,

        Thank you for reaching Emotiv. We appreciate your interest in our technology.

        We currently are not involved in any research drug recovery. If a similar project comes up, we will be making an announcement and actively look for participants and resources. We are happy to keep you on our list of someone we can tap for assistance.

        Thank you.
        Best regards,
        “name withheld”
        Customer Support Agent

    2. tangent says:

      Well. At best you’re useful once.

  18. Uncle Al says:

    A fashionable stink of social intent wafts upward. AI crunches chess and Go, but not design and synthesis (to date). Intelligence, creativity, “AHA!” plucking a blue rose from vacuum, are emergent. If you dissect the source it vanishes before enlightenment appears.

    Acetylene was polymerized with a misplaced decimal point. Pyroceram was a dyslexic furnace setting. Pedersen’s crown ethers. Penicillin, nylon, spinning Kevlar. Do not dismiss the awesome if fickle power of mistaken assumptions, luck, fetish, and transient frank stupidity. Quantum AIs may be the answer, or not.

    1. John Wayne says:

      Do you think that the potential of AI is the ability to make mistakes faster? I find this thought oddly appealing…

      1. Uncle Al says:

        An error-making AI would only be human. 50 years and a million+ pages of non-classical gravitation are empirically sterile. Given zero yield from a huge cadre of the finest Terran minds, who is better qualified to crack physics’ nut than a tar-mongering organiker?
        …It’s testable. I’m already ahead. The worst it can do is succeed.

    2. Design Monkey says:

      That’s not at all a problem.

      To err is human, to really foul things up requires a computer.

    3. DrOcto says:

      The more I read Uncle Als’ posts, the more I think that he/it is a bot.

      1. tangent says:

        Easier to make PARRY realistic than to make ELIZA realistic.

  19. Shanedorf says:

    Its not a bad premise in that you’ve got thousands/millions of med chem brains, but no way to assimilate that collective experience and knowledge. For those familiar with it, the Simcyp Simulator is one useful technology/consortium for collecting and disseminating pre-competitive knowledge and its utilized by most of the Top Pharma companies in their modeling and simulation efforts.

    There’s a vast amount of useful knowledge silo’d in the brains of med chemists, but the confidential nature of that info means many companies are spending immense amounts of time and money to learn the same things that are already well understood at other pharmas.
    I can’t speak to the success of this particular AI adventure, but the concept has some merit.

  20. CMCguy says:

    As a process chemist my first thought in reading the title was I pretty much cat tell you what’s in (most) medchemist’s minds: “OK these compounds all pretty much have similar activity so which molecule is the least ugliest and has less terrible route to prepare that I can kick down the road for other people to deal with in order to meet my artificially mandated objectives”. On the positive side does help keep many process, formulation and analytical chemists busy to enable development moving although in the end we all are at the mercy of clinical and/or business/marketing people who might derail the efforts.

    1. John Wayne says:

      Don’t forget that we will also give you all our hopeful projects in the fourth quarter, and complain that you can’t work on them all.

      “artificially mandated objectives” Yeah, those are there.

  21. tlp says:

    from their website:
    The system is intended to answer the following questions:
    – Can a medicinal chemist have a ”hunch” about the good or bad molecule?
    – Can a medicinal chemist predict the outcome of a clinical trial by just looking at the molecule?
    – Can we create a test to compare medicinal chemists using 1 & 2 questions?
    – Can the medicinal chemist distinguish the molecules created by the human medicinal chemists from those generated by artificial intelligence (A.I.)
    – Can we use the top-performing medicinal chemists as discriminators for training the deep neural networks?
    – Can we pinpoint the are a of the brain responsible for recognizing the good molecules?

    Just wait for this test to be included in the interviewing routine (if penultimate question proves that ‘top-performing medicinal chemists’ are useful at all).

    1. Pennpenn says:

      Yes, no, probably not, depends on the chemist(?), probably, no because it almost certainly isn’t a single area of the brain.

      1. tlp says:

        it’s also gut, so maybe they could analyze best med chemists’ microbiome instead?

    2. tangent says:

      Wow. Piling up a pretty impressive list of fields they obviously didn’t talk to anyone in.
      – medicinal chemistry, passim
      – neurobiology, I mean “the area of the brain responsible for recognizing the good molecules?”
      – economists, who would advise that if med chemists could predict the outcomes of clinical trials, companies would hire med chemists and stop failing so dang many clinical trials

  22. carlospln says:

    Wait a minute! I know who wrote that press release:


  23. Derek Freyberg says:

    I went to the press release and, sometime in the reading process, was confronted by a popup ad wondering if I’d like to subscribe to their “weekly newsletter”, which I at first misread as “wacky newsletter”.
    Now, I’m not sure I did misread it.

    It sounds awfully like the – apparently fairly hopeless – Watson for oncology, only without IBM. See the Science Based Medicine blog:

  24. DCE says:

    Everyone knows med chemists are dumb and just screen. The truly above average thought leaders are all the photoredox academics

  25. Ken says:

    If they get fine enough EEG data, they might be able to isolate the area of the brain that recognizes polyphenolic quinone, determine what proteins are uniquely expressed in that region, and isolate the polyphenolic-quinone-recognition-gene.

  26. Scott says:

    “Yeah, I’ve got some facial expressions they can read – there’s one when someone shows me a polyphenolic quinone as a lead screening hit,” – Probably about the same as someone telling me that they totally did shutdown and restart their computer before calling tech support. “Yeah, sure you did”

    “another one when I’m entering the second hour of an HR presentation,” – that’s definitely “Dear lord, someone shoot me!” No meeting should last more than 30 minutes. Maybe 45 if there is some random BSing before getting to the point of the meeting.

    “and then there’s that one that I make when someone asks me to align the synergies so that the stakeholders have a win/win situation.” And that one is definitely “shut up before I shove that stock option down your throat.”

  27. David Edwards says:

    One aspect of the application of AI, that all too many practitioners in the field forget, is this. The knowledge bases they’re hoping to press into magic service in their computers, have long histories.

    Organic chemistry has a history spanning nearly two centuries, and it’s thanks to that long, continued diligent effort, that medicinal chemists are able to know what they do know about molecules. But, that history also throws a spanner in the works, known as “there’s too much data for one human being to assimilate”. A problem that reared its ugly head in the world of entomology some time ago, courtesy of the fact that there’s 400,000 species of beetle known to science (and that number is growing daily).

    Even in a discipline such as pure mathematics, within which it is, in theory at least, possible to have complete knowledge of the entities and interactions being studied (until, of course, your work reaches the point where Gödel’s Incompleteness Theorem starts to apply), progress applying machine learning has been slow, but at least there, it’s been methodical, and constraints upon ambition have been treated as long term research projects, instead of inconveniences to be hand-waved away to please the shareholders. A fine example of what’s possible, when proper, diligent effort is applied, is Isabelle, but those responsible for creating it, don’t claim it will replace pure mathematicians – instead, they offer it as a tool for pure mathematicians to reduce their workload on certain classes of problem – once said pure mathematicians have spent two years learning how to press it into service, of course.

    There’s a nasty parallel between what I read here about medicinal chemistry, and software development (my own particular niche). Both disciplines require diligent effort in order to produce successful results, and both disciplines are handicapped by outsiders whose influence is usually malign – read: MBAs who think they can magically make a fast buck from wishing the contents of their cerebral holograms into reality, without bothering with the aforementioned diligent effort themselves.

    I’ve seen this at work in my own niche, where, for example, some new self-declared wunderkind who is a legend in his own bathroom, and whose actual technical understanding of the discipline in question is best described as “potty training level”, decides that some new fad will magically turn the company’s assorted software turds into glittering diamonds. Unfortunately, trying to tell the typical Yuppie business school pink oboe player, that it’s going to take two years for even highly trained developers to turn this new fad into something meaningful, is rather like telling him you’ve just sold his youngest daughter to human traffickers. In he breezes, dollar signs permanently tattooed onto his eyeballs, spouting the usual business school dribblespeak phrases that were so ruthlessly satirised by the character of Gus in Drop The Dead Donkey, convinced that he’s going to weave the mother of all magic spells with his utterances. All whilst hoping no one will notice that his plans involve milking everyone else’s talent for his enrichment, so that he gets to retire to a 300-foot long ocean-going knocking shop moored in Monte Carlo, whilst the people he’s milking can look forward to yet more time spent trawling the job ads, trying to restart a trashed career in a field littered with sharks, all looking to use the employment hiatuses as an excuse to drive down salaries and working conditions.

    When, of course, the pet fad turns out to be even more intractable to press into service than the last one the company took on board, Mr Smarm looks for the quickest bolt hole to take his expensive German saloon to, armed with nicely fabricated PowerPoint presentations aimed at persuading other Mr Smarms elsewhere that he’s actually weaved real magic, in the hope that no one actually calls him out on the disaster he left behind. In the meantime, people who spent years actually acquiring real knowledge in the field, are cast out onto the garbage heap, and left to rot. All in pursuit of “monetisation”.

    I’m sure this is all depressingly familiar to several med chemists here.

  28. Anon says:

    Time for an update on LMTX?
    Taurx will soon restart trials.

  29. yp says:

    Machine Learning is just learn and aggregate knowledge, it does not create new things. What company will hire perpetual learners as known as students ( even it is a cheap one) ? We need people who can create new products and new drugs.

Leave a Reply

Your email address will not be published. Required fields are marked *

Time limit is exhausted. Please reload CAPTCHA.