 Reacting only to this session right here, although I would love to react to all of the discussions today. So with that in mind, I think there's sort of one core theme I want to focus on for my reaction, which is the patients. How does ultimately everything we're doing get back to the health and welfare of the patients? And so there's three sections I've been asked to address. I'll talk about novel sources of data with respect to patients, both how patients can generate data and mechanisms by which they can, and how deep learning can impact genomic medicine. So I'm assuming my mandate is broader than just emerged since genomic medicine I think is a very broad topic. So I'll talk about stuff that could be relevant to emerge. But basically any way that I see, both on the patient side and as a computer scientist, that deep learning could in some way impact genomic medicine. And then I'll talk about novel ways of measuring outcomes that really reflect the patient. So in terms of novel sources of data, I think that the patients are providing an awful lot these days that previously was sort of unconsiderable. I think the challenge is how do we take these sources of data and get them in a more structured way into the EMR. So I think one thing that's already beginning to happen is patients are doing an awful lot of self or what I would call peer-to-peer phenotyping. And we've got to find a way to start to capture this in some structured format. And there's sort of degrees of formality to this. So if you haven't done this, if you sort of search YouTube, you will find that patients are now frequently posting videos of themselves or their children saying, what is the symptom? Or is the symptom what I think it is? And so patients are, and you'll get tremendous amounts of feedback in the comments, sometimes from other patients, sometimes from professionals saying, yes, I think this is what you have. So patients are doing a lot of phenotyping on their own and there's gotta be a way to capture this in a structured way. And I think that some of the tools to do this are evolving now. So if you look at the Monarch Initiative and this new sort of patient phenotyping ontology that's been developed, there's a way to start to link what patients might be able to type in and figure out into the more structured terms that we're all more used to. So I think we have an opportunity to start with the patients and their own interpretation of their phenotype, with the language they understand and move that in an automated way into the HPOs that we can do really rigorous science with the data that they generate. There's also, this comes up a lot today, this question of what do you do with the U.S.? I think this is one of the central challenge in genomic medicine. Well, it turns out that if you turn the U.S. over to patients, sometimes they can do something with them and oftentimes what they do is they turn to the internet and say, what does this variant mean? And you get all sorts of answers when you do something like this. I've had some personal experience with this and I've even encapsulated some of that experience through the UDN in their participant web pages project where we actually will create web pages for individuals as long as they consent to it and we will do things like search engine optimization, we'll try to get things to go viral, we'll try to get cases out in front of other eyeballs and find a match to get an interpretation of a U.S. by finding another case. And it turns out this actually works remarkably well. So in about 25% of the cases that we post about a month later, we have a match of some kind. And this is much cheaper than doing functional studies or many of the other forms of interpretation that we might go about in the events of the U.S. Yeah, and then there's, I think there's, even sort of the next generation versions of this emerging where patients are even more empowered to do it. So there's sites like MyGeneTune now where patients can go and upload their own data in a very structured way and ends up searchable index. And I think it's even linked through matchmaker exchange now, so you're ending up in the other databases that are out there with systems like this. The other thing I started to consider was, well, how else other than sequencing might you arrive at genotypes? And I realized that there are other sources of data that give off clues as to what our genotype might be. So, for example, I've found on forums where people starting to infer their own SIP variants just by looking at how they interact with drugs. Like I say, well, how do you interact with Nyquil? How do you interact with Dayquil? What does this do to you? And you can start to maybe figure out what variants you have for something like SIP2D6. And so I think there are other ways of getting at genotypic data through questionnaires that might even be low cost or easy sort of prefilters before we actually do any kind of real solid genotyping. Of course, it's possible in some cases now to infer genotypes from images. So there's this FDNA project where they can actually diagnose rare diseases through nothing more than the images of the patients. And then I started to think really big and I wondered, are there more interesting ways to gather genotypes about patients? And I'm willing to bet that if you had access to somebody's full Google search history, you could start to infer certain SNPs. In fact, so if you saw people Google something like high cholesterol or difficulty swallowing, I'm sure you'd start to learn something about them genetically over time. And in fact, I think all of us should probably consider asking patients if they will consent to have their entire search history collected so that you could do FIWA studies where the phenotype is search history. I think that would be very interesting to see what pops out of that. I mean, I actually don't know, but I bet you something would happen. You'd find something. So let's talk about deep learning and how this might interact with genomic medicine. Well, I think there's one actually high impact way and it will certainly act very early on. And this is what I would call diagnostic trajectory recognition. In fact, I've got a project with Ken Mandel and Ben Rabie where we're doing exactly this, where we're saying when you have particularly rare disease, you end up with these sort of diagnostic trajectories where for any given disease, there's often a set of prior diagnoses for which you're misdiagnosed first. So you might start off with no diagnosis, then you go to AT. Then some sort of lysosomal storage disease then AAA, then RAT for ultimately some day landing on a disease like anglo-1 deficiency. And when I talk to patients in the rare disease community, they all have these stories. And each community has its own set of prior diseases for which they were considered. So if you look at the negative genomic data in the arm, that is the set of genetic tests for which a patient was negative that could be informative about where they're ultimately headed. And maybe you can short circuit some of these audiences when you see a negative test for this and a negative test for that, a negative test for that. Maybe you can say, well, once you've had these three negative tests, two or three years ago, you're likely to have a positive test here. So maybe you can short circuit that and just jump all the way down the line. And of course, this is learnable. If you turn deep learning loose on an EMR with genetic data, it's gonna find these trajectories automatically for you. It certainly is capable of doing that. And then of course, we come to variant interpretation. And I know I think that the bias in this room is certainly towards the statistical approach to interpreting variants where we have large amounts of genomic data over large populations to infer pathogenicity and things like that. But I want to remind you that the structural interpretation is important too. And if we're asking about what deep learning can do, then deep learning can definitely help in a situation like this. Certainly deep learning can help us go from variants that are reported to actual structures. So right now when we do this, there's a lot of hand tweaking where we, there's very little ab initio protein folding in these days. But oftentimes you start with a known protein structure and do some homology modeling to get you into the right place. But I think going forward, a lot of what's done by hand right now, particularly the homology modeling part, maybe even entirely new ways of doing ab initio protein folding will entirely be done by deep learning. I think deep learning is gonna be fantastic for predicting structures. And then structures of course are very useful when interpreting these VUSs. You can start to say, well, where in the protein is this actually occurring and what might it really do from a structural perspective? So I think deep learning will really accelerate our interpretation of variants. Then of course it has a role to play in accelerating us towards therapies. So that's ultimately what this is all about at the end of the day. We wanna match patients to existing therapies or if we can develop new therapies entirely. So I think if we expand genomic data to include the transcript and of course there's lots of interesting techniques where deep learning can play a role already in terms of matching patients to potential drug cocktails. But I think we can also accelerate this process in other ways including more traditional acceleration of the sort of standard drug development process but scaled down to the level of an individual. And so one area where this can help already is in something like docking simulations where you've got some sort of molecular target you wanna shoot at and the question is, well, what small molecules would actually hit that? Well, there's already research in the space happening that shows that deep learning is significantly better than traditional means of doing docking simulations in terms of predicting which small molecules might interact with a particular drug target. So there's in fact, there's even a startup called atom-wise already doing stuff like this where you can send them structures and they will send you listed of interactors for that structure. And it goes hand in hand with course predicting toxicity. So right now toxicity is kind of a black art. You get a bunch of medicinal chemists in a room and say, how do you think this is gonna react? And they all sort of have this intuition about how it'll happen. And anytime you have a process where you have a whole bunch of experts using their intuition, I think you know right away you have a candidate for replacement with deep learning. That's sort of how this goes. So I think ultimately, you know, TOX is the kind of thing that we're gonna start to do more and more with deep learning instead of just getting a bunch of smart people in a room and trying a bunch of stuff in a lab. And this ultimately could lead us even into structural pharmacogenomics where we could use deep learning to predict, you know, with a little more resolution than we have today about what some of the US might actually do in terms of our interaction with drugs. You know, as we might even make novel predictions someday where we see a variant for the first time and say, well, we've never seen this variant but we think it actually gives you a very high risk of having this adverse reaction right here like Steven Johnson's or something like that just because, you know, we've modeled, we should went through and modeled the interaction with this variant and we think it's potentially pathogenic for you. All right, so novel ways to study impact. Well, again, just sort of focusing on the patients. I think, you know, we can look at the sort of standard things like, you know, if an intervention has been made based on the basis of genomic information did we end up changing the prescription in some way? If we change treatment, then I think that counts as a whim. Of course, there's metrics like improved survival and improved quality of life. But then the question is, well, how do you really measure some of these things? I mean, survival, I think we're pretty good at measuring but improved quality of life is a little bit more difficult to do. But we have other sources of data now. So we've got wearables of course and this gives you sort of unprecedented even intrusive levels of precision into somebody's quality of life. And, you know, you can start to use all this information taken together to infer whether or not some intervention is really making a difference for them. The other thing we could consider and is if patients will consent to giving us access to their social media feeds, you know, companies like Facebook and Twitter have already figured out that you can do sentiment analysis just from what people are, how they're interacting with their social media. So you can figure out if somebody's happier or sad and you can tell if on net somebody's become happier or sadder as a result of some intervention. Yeah, and of course the, one of the reasons we know this is the few, if you remember this study, this is the one where they were not only measuring sentiment but determining if they could interfere with sentiment as well, which is definitely kind of, we don't want to interfere like that. We just want to just use it as a measurement. Of course there's metrics like financial impact. You know, I think in many cases going to the root cause of the disease should hopefully, you know, limit the costs that patients are facing once they finally get the answers that they're looking for. But I think to close, I want to talk about ways that patients can really engage with science. And I think this is something that is somewhat unique to genomic medicine or at least some more prolific in genomic medicine. This whole notions of patients getting more engaged with addressing or understanding or addressing the root cause of their disease. I think part of it is because genomic medicine is one of the few places where that actually happens, where you really do, in many cases, get an answer about what the root cause of your disease is. So I think this is a metric we ought to look at. And part of that is because, you know, even if there's not some sort of medication you can give a patient right away, if you get them engaged with science it's still something that they're doing and it should count as actionable under our def, I think there are any reasonable definition of actionable. So just as some examples of levels of engagement that are out there now for patients, I bring up this project Mark to Cure. So Mark to Cure allows laypersons, they don't have to have any background whatsoever in biology. And I'll just mention this is a partnership that I've got going with with Andrew Sue, which allows them to annotate abstracts in PubMed that might be related to a condition that they're interested in. In this case, the condition is Anglai-1 deficiency. And there's some very interesting findings coming out of this work, which is that if you take, and this is probably not so good for the NIH, but if you take six lay people and you average their annotations together, what you get is one PhD. So I don't know what that means for the future of PhDs in biology, but it's great because it means you can harness the interest that the patient community has and you can get them to structurally annotate abstracts and you can start to generate these minimal knowledge graphs that you otherwise couldn't have, which sometimes actually give you very meaningful information you otherwise would not have seen that take you closer to understanding your treatments for diseases. In some sense, it's sort of mining out the unknown. The things that we actually do know, but we didn't know that we know. You can do that all based on the willingness of patients to engage in a scientific way around their disease. And the other thing I think we should watch is the development of communities as an outcome. Our communities springing up in response to the introduction of genomic information to patients. I think in many cases they are. Patients are becoming linked through their genes. And as a result of that, they're forming communities that are advancing understanding. And what many of these communities are doing, and I think this is the final metric I'll offer for measuring outcomes, is trying to develop therapies. They start off with understanding, but ultimately move over time towards developing therapies. I talk more and more to patients and patient communities that are actually engaging in the science together and doing things like drug screening. This is stuff that I think was almost inconceivable at the scale that's happening right now, just five or 10 years ago, and yet it's happening. So we need a way to track that this is going on and measure it as an outcome for all genomic medicine. So thank you.