7 ms·
This paper shows what can be done when you carefully run an ML program alongside a wet lab experimental program tailored to feed back into the ML program. The r
by entee 7y ago
This paper shows what can be done when you carefully run an ML program alongside a wet lab experimental program tailored to feed back into the ML program. The results end up far more interesting than some recent "ML aided drug discovery" papers in that they actually discovered a drug that functions very differently than known antibiotics (AB).
Even though the structures that came out look AB-like, they work differently than known ABs, probably by disrupting the pH gradient across the cell membrane. Other ABs might do the same as part of their activity, but work better under different conditions than this one. The result is an innovative structure, and a molecule that can hit resistant strains.
Combining ML and wet lab is the real way we'll get new drugs. You need to regularly check in with a high content ground truth or you'll come up with either uninteresting or useless results. I'm a bit biased though, that's what we do at my company ;)
- chrisgd 7y agoCan you help me understand what is the difference between what you are doing and what people like Certara and Simulations Plus are doing? The simulation software market is really fascinating
- entee 7y agoSure! To a first approximation they both provide software packages to pharma who then apply those to their efforts. In our case we are building an integrated system with a biochemical component and a computational component. We will then seek out molecules to difficult targets ourselves because we believe the integration gives us a competitive advantage.
- spitfire 7y agoNot just in drug discovery but in most interesting industries. Using ML as either a human/cyborg aid or ML+real world ground truth is a secret superpower that I'm surprised more people don't know about. I'm glad they don't.
- timClicks 7y agoThey do. It's just easier to add extra CPU cores to a ML pipeline than another person.
- IAmEveryone 7y agoAs a less cynical observation than a sister comment, I think this is partly due to how current tools work, or possibly even inherent to the technology. Just look at the number of ML examples where you never even get to see a single example of the data you are working with. Interactivity is even rarer, certainly because it doesn’t scale, but maybe also for the somewhat pedestrian reason that interactive UIs doesn’t come natural for either the tools and/or the people using them. Then, there’s this issue of “explainability”: if you want to direct some generator you need to find out how the concepts you want to work with are encoded in intermediary layers. To be fair, all this isn’t much of a secret, and there are quite a few projects doing interesting things. Magenta comes to mind, or GANBreeder.
- sgt101 7y agoOne challenge here is seen in the types of articles researhers in comp sci have published over the years. I believe that there has been a massive decline in case studies and field trials of software and a rise in analytic and empirical algorithm development. To unlock the super power for people we will need more good hci and the wheel will have to turn in research practice once more.
- xzel 7y agoPeople have been doing this exact thing for two decades at least but obviously with less computing power. There's literally nothing new about the idea. The real trick is being incredibly lucky and finding something that actually works in humans after multiple trials. I'm sure you know this based on your comment and this isn't really directed at you (truly wish you best of luck, I really hope the computing power and skill we have saves lives) but I absolutely hate these types of articles, which I've seen becoming more frequent the past year. AI + scientific challenge + possibility = puff article. This is the type of semi-hysterical reporting I expect from a local news station.
- entee 7y agoI share some of your skepticism, but I think this particular paper is worth a little less cynicism. It’s true that statistical analysis has always helped inform medicinal chemistry efforts. They used to be called QSAR ;) What’s different here is the scale of information that was processed. As you point out the computational power and therefore the algorithms we can apply has enabled different experiments which I very much hope will yield better and more abundant medicines. Papers like this are a sign of these methods bearing fruit!
- xzel 7y agoI totally agree, and I think my comment could be summarized as such from another angle. I think the recent Coronavirus + AI = cure spam has me a little jaded. I'll add this, this paper is really well done and like I said I really hope this works in humans!
- headmelted 7y agoI’m usually also skeptical of getting excited about successful animal trials for new medications, but with this I’m actually leaning more on the side of optimism. With cancer treatments and antivirals in mice, we’re not so much targeting the pathogen as targeting the host immune system in the hopes it ends up nerfing the intended target (tumor/virus/whatever). Given that the compound seems effective against C. Difficile (even if it’s in mice), I’d expect it to work elsewhere. Of course, I’m not a doctor and have no idea what I’m talking about so grain of salt required.
- naresh_xai 7y agoExactly why we need causal reasoning/causal proof alongwith ML
- entee 7y agoIn this case I’m not sure if the reasoning for what molecules work would make sense even if we had an “oracle” to explain it. Why a molecule works is a complicated interplay of fundamental physics and emergent properties. The explanation is likely not human interpretable without a ton of equations in the first place. That said in broader medical cases of ML where particular symptoms and measurements drive a diagnosis for example, I tend to agree.
- IAmEveryone 7y agoI’m not sure about the exact mechanisms of drug efficiency. But there are well-known patterns in molecular biology, auch as specific sequences in proteins reliably forming specific 3D structures. If these models produce novel but (somewhat) effective structures, it must be because they pick up on less obvious patterns in the data. To be able to describe these would seem to super effective.