Below are collected Twitter ramblings for August 2024, with commentary interspersed as usual. It’s a relatively short edition this month as I was away for two weeks on vacation, but still plenty of topics covered.
This month’s photo of the month is the Margerie Glacier in Glacier Bay National Park, Alaska. Vacation travels this year took us north, and the glaciers were a big highlight. We got to see this one calve too!
One of the things I’m most proud of about the time I’ve spent at my current company is the contributions we’ve been able to make to wresting oral absorption out of some pretty large molecules — ones that still make some medicinal chemists blanch, even today. At best, this idea was science fiction 20 years ago, and at worst, it was deemed laughably impossible. If one takes the attitude that things are impossible in the drug discovery world, you’ll probably be right 95% of the time. But you’ll also have a hell of a time discovering drugs while you’re busy saying no to everything.
Turkish pistol shooter Yusuf Dikeç captured the Internet’s imagination at the Summer Olympics in Paris this year. The sport has gone wild on high tech gadgetry, as witnessed by South Korean shooter Kim Yeji on the left side of the meme. She’s become an Internet sensation in her own right for having “main character energy,” and perhaps unsurprisingly has now landed a movie role as an assassin. Dikeç, in contrast, looks like he just rolled in off the street to compete in his t-shirt, hand in pocket, no gear — and still nabbed a silver medal.
Well, everyone raced to put this contrast in the context of their domain of expertise, and the memes were born. In a chemical series optimization, medicinal chemists often have a lot of ideas floating around about how to address the deficits of the current lead molecules. Designs to address the hypothesis, however, can vary from simple to exquisitely over-engineered. A common source of tension between medicinal and computational chemists is that the latter sometimes design such exquisite molecules which are precisely fit into some pocket in a model they built — and are also often nigh-on impossible to synthesize.
The intersection of the perfect design and an easy synthesis is always preferred, of course — when you can get it. But if I have to choose, I’ll pick the simplest structure that will answer the question. This is because most hypotheses won’t work out, and you can waste a lot of time here when you should probably just move on. I’ve seen people spend months doggedly pursuing a single compound this way, usually to their (and the project’s) detriment.
While it’s true that a good idea can happen anywhere — I, like so many, have had some good ones in the shower in the morning — you increase the chances of that idea emerging by increasing your interactions with your colleagues. Drug discovery is a big enterprise that many people with disparate expertise must contribute to for success, so no individual will ever know it all. It’s so common for a good idea to drop into my head while I’m talking with someone at work — even when discussing something that was only tangentially related to the eventual idea.
Good medicinal chemists are in some ways like archaeologists, constantly digging up experiences, memories, and learnings from past projects to apply to current ones. Sometimes it’s hard to marshal all of these scraps and fragments into a whole opus, laid down and temporally disconnected as they are in decades of drug discovery sediment. And sometimes all it takes is a few words with someone else to jog something loose in your brain. Long story short: you’ll be more successful in drug discovery, and in life, if you interact with folks instead of trying to go it alone.
You’d probably be surprised how I’ve made use of machine learning tools in my work over the years. Sometimes people only see me on Twitter taking a critical line with the AI proponents. But that misses an important nuance in the argument, because I’m not arguing against AI/ML approaches as an aid to drug discovery. I’m arguing against the idea — and those who espouse it — that we’ve suddenly found the cure to all our ills.
The drug discovery business frequently finds itself in the grip of mania when new ideas come along. There’s so much failure that people get desperate to do something — anything — that will move the abysmal clinical trial success rate off the peg it’s been stuck on since forever. Investors throw money at the New Thing, and big surprise, scientists and company executives alike pile in against all reason to chase and capture the funding flow. Meantime those of us who have been doing this work for a few decades wait for the chaos to subside, sift out the good ideas — the keepers — from the hype, incorporate them into our work, and move on.
Even in Alaska, you can’t really guarantee you’re gonna get a banger of a geomagnetic storm unless you stay up there for a while. The northern lights can be fickle. We got lucky that we saw this amazing show just two days into our trip. It only lasted for an hour or two, and we never saw them again after this. But this unplanned moment became the highlight of the trip for me. The photos do zero justice to how amazing it was to see in person. Maybe there are readers here in more northern latitudes who see the lights more often — I’d wager it doesn’t get old.
I’m as guilty as anyone of talking about drug discovery — especially in the Twitterverse — in an abstract and theoretical way. To be sure, there’s plenty of theory at the foundation of the science we do, and you’ll be better at discovering drugs if you take the time to build this foundation. But at some point the talk and the planning and the PowerPoints of proposals and screening cascades and timelines and checklists has to end. This is a business where there’s a huge premium on good execution. It’s an experimental science where we have to design things and then go make and test them. It’s engineering at a molecular level, and like all good engineers, we have to build stuff. Roll up your sleeves and get to work and make shit happen. The doers will steamroll the pontificators in drug discovery all day, every time.
Yes, I’m that old.
The training set question is why I remain deeply skeptical of some of the loftier visions of the AI proponents, such as “simulating” biology. Too many people seem to believe this is simply a problem of scaling infrastructure — that all that lies between us and the simulation is enough sufficiently powerful GPUs. Amazing business model if you’re in the GPU selling business.
In reality, the big problem is the data we have to interrogate biology as a system is, quite frankly, a mess. It’s incomplete, oversimplified, noisy, and lossy — all at once. That’s not for lack of trying, either — it’s just a massively hard problem to make inroads on, as anyone who does it for a living can tell you. And all of that sets aside the decades-long crisis in the reproducibility of the data we do have. And also sets aside that some people are actively polluting the data set by fabricating things out of whole cloth — as I’m writing this, Dr. Eliezer Masliah is in an ocean of hot water for apparent decades of data fabrication. There’s a lot of movement to omics the shit out of everything to pipe more data into the models, too, but we should then question if the quality of data we’re getting there is really going to be what’s needed for a simulation model.
Make no mistake: AI and ML models and tools will, in my opinion, make an impact on the drug discovery process — they already are. But I think we should prepare ourselves for an incremental future rather than a revolutionary one. That’s the track record of all the prior revolutions in this business.
Worth reading into the comments on this one for some terrifically bad puns, which all scientists seem to have a weakness for.
Not that I didn’t learn some useful skills at Pizza Hut about working in a team environment, dealing with a boss, managing money and time, and so on. I encourage all teenagers to get a part-time job if they can, even if it’s just a summer job for a few hours a week. Good lessons about the real world await.
I’ve been asked to write some guest editorials for ChemMedChem, along with some other medicinal chemistry voices, and the very first one (which I’m writing now, I swear!) is going to be about the need of in vitro potency, pharmacokinetics, pharmacodynamics, and in vivo efficacy to fit together as a whole. Too many med chem journals have too many stories that hang together by a thread, or not at all, simply because nobody seems to have taken a step back and asked “does all of this fit together well in a single framework?” Only a handful of questions are really needed to interrogate the framework, too. More to come!
Also actively working on this post. It’s been delayed repeatedly over the last month due to other commitments in the real world, but it’s like 95% done and just needs a conclusion and some editing. Optimistic this will drop sometime in October.