I’m sitting in an MIT conference on AI in drug discovery/development as I write this. One of the speakers here (Mathai Mammen, J&J/Janssen) just made a good point – not a new one, but a solid one that deserves some thought. He called for “bilingual” people, by which he means people who have some fluency in data science and some fluency in one or more of the various fields that make up drug research.
That split has been a noticeable fact over my whole career. In my own field, there’s always been a gap (greater or lesser depending on the people and the circumstances) between the molecular modelers and the medicinal chemists. The cohort that really bridges the two was small when I started in the industry in 1989 – as well it might be – and although it’s larger now, it hasn’t grown as much as you might think. They’re still two separate fields, and their practitioners are still very capable of talking past each other. I remember hearing, years back, a prediction that over time every medicinal chemist would just naturally have computational chemistry as part of their tool kit, and that distinction between the tribes would disappear. Hasn’t happened.
Why not? There are several reasons. The sorts of people who get interested in these areas tend to be a bit different from the start, I think. People tend to get into both the subject matter of a given field and the tools and technology used to realize it. Lead guitarists tend to know an awful lot about the different brands of electric guitars, pickups, slides, amplifiers and so on. Oil painters develop a lot of opinions, backed up by experience, about different brands of paint, canvas, types of brushes, and all the rest. Meanwhile, medicinal chemists tend to be people who get into (or got into!) organic synthesis, biochemistry, chemical biology and such, and have both knowledge of and affection for the tools of those trades. And for their part, the folks who really know and practice molecular modeling tend to be people who really got into. . .coding. Programming, algorithmic optimization, data handling, with a strong side interest in computer hardware and various software packages and tools. These are different people. We need both types (and more besides!), but pretending that they’re not different people with different interests is not useful.
Another reason that these populations haven’t converged, I think, is that they haven’t felt the need to. For one thing, both groups have been content to let the specialists in the other amass the knowledge needed to do their jobs – there are only so many hours in the day and so many neurons in one’s head. And to be really honest about it, neither group has felt that a detailed knowledge of the other’s field is necessarily worth the effort. Most medicinal chemists, as mentioned, get into the area via organic chemistry. We know how to make molecules (and by “know” I mean both intellectually and through physical experience), and later on we discover how to apply that knowledge to work in drug research. A computational chemist will (rightly) not see the point in knowing the ins and outs of metal-catalyzed couplings, which purification techniques are the first to try, which sorts of assays tend to give more actionable data, all the bench-level stuff that is the foundation for most of their med-chem colleagues. Likewise, the med chem crowd has not seen any advantage to learning about the efficiencies of various sorting and sampling methods, how to configure a computational task to take advantage of GPU hardware, or the strengths and weaknesses of the various levels and flavors of quantum-mechanical approximations.
At an even higher level, one that’s even more uncomfortable to face, medicinal chemists would probably get more interested in learning the details if the computational approaches produced useful recommendations more often. Or perhaps if they hadn’t been burned as many times in the past. To be sure, the computational chemists would get more interested in the nuts and bolts of organic chemistry and assays if that knowledge had a chance of actually making their lives easier, too. As it stands, there’s a “Why should I” aspect that’s hard to get over.
But that’s the promise of machine learning/AI: that it might actually start providing some attention-getting answers to “Why should I”? If ML can start spitting out actionable predictions and insights that we wouldn’t have seen on our own, well, attention is going to get paid. We may well be on the threshold of that now – many ML/AI people would say that we passed that threshold a while back, but the real test is convincing a skeptical audience from outside your field. As much as I can sound like a skeptical member of that skeptical crowd, I will be very happy indeed if the convincers start coming. We’re about to find out.