None of us is as wise as all of us, when it comes to diagnostics. So it makes sense to involve deep learning in medicine wherever we can. Think of what IBM’s Watson can do today. Now imagine an AI capable of deep learning — one specifically built for medicine, programmed for diagnostics. Any one person isn’t going to win when pitched against the assembled diagnostic insights and clinical pearls of all the doctors, past and present. Put another way, deep learning could be an incredible force multiplier.
And we need all the force multipliers we can get, in this unending war against disease. Build a better mousetrap, as the saying goes, and Nature will build a better mouse. Sure enough, diseases evolve. Even viruses. As you may know, there are multiple strains of HIV? Group M is just one type of the “first” strain (HIV-1) responsible for the human AIDS pandemic. But there are about a dozen subtypes within group M that each hang out in their own bioregion.
One of the obstacles to treating HIV is its high genetic variability. It’s difficult to make antibody-based drugs fast enough to keep up with a virus that’s constantly shuffling around its genome. Trials for a vaccine are ongoing, but nobody has quite got it, partly because of this incessant mutation.
The fascinating thing about viruses is that under the hood they’re nothing but a wisp of genetic material, with a header and footer containing its duplication code, plus a few lines of metadata that might code for a protein or a lipid or two. Just like you can track changes in Word, you can track changes in a virus over time, given enough samples and sufficient investment of computational muscle. That’s how we know about all those subtypes.
As it happens, though, multivariate analysis is a particular strength of AI. The kind of sophisticated n-dimensional number crunching that could keep a team of dozens of scientists busy for years, Watson could eat for breakfast with its 16 terabytes of RAM. That’s how we get those beautiful predictive models of what galaxies will do when they collide, and how the aerodynamic performance will work on a car designed entirely with CAD.
It’s also what makes AI a powerful ally in the fight against HIV.
Many factors govern the spread of diseases. Beyond the pathogen’s own genetic sequence, and the virulence factors it codes for, there are still many other variables. Economic, political, social, and meteorological forces can change the movement of people, individually and en masse. There is a nationwide opiate and heroin addiction crisis, and in its wake there is a gathering storm of HIV infections via needle sharing. People move around the planet, and with those people travel the pathogens they host.
But we could use AI to construct a nuanced, informed assessment of many different such forces and factors, by plugging in that ridiculous volume of multivariate data to a program that can track all those changing rates at once. We could deploy deep learning and neural nets to suss out the patterns we can’t see, and then use those patterns to track and predict the spread and change of the many subtypes of HIV.
But the AI’s work isn’t done yet. Comparing the change in genetic code with infection rates and virulence factors could give us a better model for working toward a vaccine for this insufferable virus. And if we finally managed to program an AI that would tell us how it arrives at its conclusions, that would be a powerful collaboration indeed. Imagine an AI that evolves with the virus it tracks. A purpose-built artificial intelligence that could tell us how it’s making its decisions, if applied to epidemiology and virology, could advance the entire field.