Founder Stories: Mark DePristo, Co-Founder and CEO of BigHat Biosciences

Hello, I'm Mark DePristo, the founder and CEO of Big Hat Biosciences. Thank you for listening to my talk, A Guide to Deep Learning in Healthcare, for the European Society for Human Genetics meeting in June of 2020. This talk is divided into two parts. In the first section, I'll review the essentials of machine learning and deep learning needed to understand the upcoming detailed examples in the healthcare setting, which cover medical imaging, medical records, genomics, and finally end with a little discussion of molecular design with AI/ML. I have over 20 years of experience applying AI/ML techniques to the life— in the life sciences. I have a background in computer science and math as well as experimental biochemistry. I've used these types of techniques in a variety of different widely used software packages that I've written and have used them at both diagnostic and therapeutic companies that I've been involved with. I was the founding team member of the Genome Sequencing Analysis Group at the Broad Institute, as well as the genomics team in Google Health at Google's AI division. And I've written a number of different publications in this field in the last few years. So let's dive right into the bare essentials of machine learning and deep learning. Data is the key component for machine learning. All of machine learning algorithms start with data. In this situation, I'm talking about a famous concrete dataset, the Iris dataset from Ronald Fisher in 1936. Ronald Fisher collected a dataset that had the sepal lengths and petal lengths and widths for 3 different species of irises in the United States: the setosa, the versicolor, and the virginican. The columns of the dataset can be divided into the input features in green and the output label or class in purple. Each— the dataset has on the y-axis records or examples as they're often called in machine learning. In this case, I've shown 3 representative examples from a total of 150 rows in the full dataset. In machine learning parlance, we would say the n is 150 observations here, and there are p 4 features used and 1 output label which has 3, 3 different classes it can adopt: the setosa, the versicolor, and the virginican when I show the examples of the three different irises on the right. This dataset is used to— So the goal of most machine learning algorithms is to create a function that can accept as input the feature values for a record and emit an accurate prediction of what the true class is given those features. And the machine learning algorithms, if you think about it as f-hat, really are divided into two key phases. There's a training phase and an inference phase, and it can be— let's start with the inference phase first. Inference assumes that you have this, you know, a well-trained good model f-hat And what it does is it accepts novel examples. So this is an example where you haven't seen it before and you don't know its true label, but you have the feature values which you can collect. You can send those into f-hat as the values x and it will emit the y value, for instance here, varisicolor, that is the predicted label for the unseen or new example. And that's immensely useful. You know, it may be radically easier to correct, to collect the values in the features than the class, and so you can basically use inexpensive observations to compute a very expensive output. So the key question for most machine learning algorithms is how do you construct this function f-hat? What you're doing is essentially training the machine learning algorithm on a subset of your dataframe that I described earlier where you're feeding it pairs of the values X, here the vector of 4 values, and the known class that's been determined for that. So you would send these matrix of sort of n records by p features and an output label into the algorithm and it would adjust its internal parameters to maximize the predictive accuracy. And the fundamental difference among all machine learning algorithms is just what is the form of that function f-hat and how exactly do you train it. A key concept to understand in machine learning is that good predictors come from good features. So there's a large body of work around what is called feature engineering and selection, which aims to, in the selection case, remove irrelevant features that distract the training of the model, and to engineer more informative features that are easier to leverage into a good predictor. So I'll walk through examples of these in some detail here. The first one is really feature selection, and that's represented as eliminating this red column. And this is not true in this specific Iris dataset, but it's representative of the common challenge here. Suppose petal length is just not related to class in any way. You collected the data because you thought it might be useful, but it ultimately doesn't inform on the predictor, the desired class. So you can use statistical techniques to identify those features and eliminate them so they don't distract the training of the classifier. That's a very common thing you see in machine learning algorithms. The second thing you frequently see is feature engineering where the values in the dataset are transformed in some way. Here we do so-called z-score normalization where you're mapping a distribution that may be nearly Gaussian into the sort of offsets in a Gaussian distribution, the so-called z-scores. And this makes the values in that matrix slightly easier to to train on. And finally, there's the ability to create human synthetic features. So the example here is suppose you find actually that you want the sum of the sepal and the petal width as another feature, you simply create a function that sort of maps those two columns into this new synthetic column, and then that's added for training. And this is very common in almost all machine learning pipelines. So to make this challenge really concrete, let's walk through an example in some detail. So the goal that we're trying to get our machine learning algorithms to do here is to predict whether a flower is an Iris setosa given only the lengths and widths of its petals and sepals. Because it's hard to visualize a 4-dimensional space, let's just simplify the problem down to make it accessible. So in this case, let's look on the left. We're only going to consider 2 variables: the sepal length and the sepal width. And this is a plot in the left box with the blue and the red points that show the distribution of those 2 properties. Across all of the Iris data. And we've binarized the problem, meaning that we're trying to predict whether something is a setosa, which are all the blue points, or just not a setosa, which is the other two classes in the red points. And that's why there's roughly twice as many red points as blue points. And the points are further divided into a critical concept of test versus training data. So the blue points. And the red points that are in circles are considered training data. We will show that to the machine learning algorithms to let it learn how to divide the, the, the two datasets. But we'll keep back some of these points. We'll keep back the testing points in crosses here, and we'll use those to evaluate the actual performance of the model. And since it didn't see those points, it's a relatively unbiased estimate of how accurate predictor really is. And we give the example on the right now figure of using very simple one-layer neural network to learn how to separate these two datasets. And conceptually, it produces this dividing line which you see in white here between the blue region where it thinks everything is a setosa and the red region where it thinks everything is not a setosa. And that white line separates the two groups, and we know that the accuracy is quite good of that separating plane because we've held back these test data. And they, when you feed that through the algorithm and compare the predicted state of the labels with the actual setosa/non-setosa labels that we know, you see that, you know, your accuracy is like 98%. So that's a very encouraging result if you're trying to predict the true setosa, not setosa label, given only these data. And ultimately, it provides a very clean example of the real challenge of machine learning. Machine learning algorithms ultimately kind of try to find these dividing lines between groups of points in space, and those dividing lines are essentially what the predictor is using to divide, say, the setosa group from the other setosa group. So you can think of the underlying challenge here as is dividing space into these buckets that correspond to different classes in your prediction problem. And, you know, the thing that makes it extremely interesting as a technology is that it's a very general-purpose approach for data-driven modeling. So here is an example. I just take a couple of other features as Xs and Ys. You know, in the upper left, it's now petal width instead of sepal width. And, you know, in the bottom left it's petal width by petal length. And you can see that, like, the, the separation is, is pretty good in most of the figures, but, um, it's also clear that, you know, you could draw a separating line between the two. Like, the challenge is learning what that line is. But as you can see, with two different types of neural networks, one has more kind of flexibility in it, which is the, the two hidden layer versus the The one, you can see that actually the line that does separate the two is different in each one of them, like in material ways. You know, the orientation of the plane that separates the two groups is totally different. It's like overall shape and curvature and everything is very, very different. And so, you know, the thing that's exciting about machine learning as a technology is that We know that it's very good at figuring out how to solve such problems, so you don't really have to worry about, you know, many of the inner details of exactly how like petal length and petal width relate to each other. You don't have to actually build a model, something that you yourself understand intuitively, out of sort of abstract mathematical form, and then use that to sort of estimate, you know, probabilities in some kind of likelihood-based framework, you can just here bring the machinery of a neural net and you throw, you know, the input data, whatever it is. You don't even really have to think about the, you know, the details of mathematically and how they really relate, and it can just figure out how to do the separation. So it's, you know, immensely practically useful. And so ultimately, you know, here we see on this slide about how neural networks really work and why it's so important to think about this plane and linear plane separators is that the thing that makes neural networks sort of work under the hood is you have these different layers, is the layers act as sort of warpers. They warp the underlying, you know, grid on the left, the sort of the true representation in a uniform two-dimensional grid. And it basically is, you know, compressing and twisting that space in a mathematically derived way based on the the exact properties or exact values in these hidden layer matrices. And by warping the, really, the two dimensions here, what you're essentially doing is transforming and learning how, you know, these two variables relate to each other in the input with respect in particular to the ability to discriminate between the two classes here in orange and green. And the longer that, that And sort of the deeper that network gets and the more values it can capture in each step, and there's all sorts of complex sub-modules, it's a very detailed and very interesting research area. But fundamentally, you have more of these layers holding more information and they can become better at representing and twisting space in a very useful way. And that lets them solve more and more problems. And that's super important The reason that nonlinear separation is so important is that many distributions in the real world can be, you know, obviously separated visually in, in two dimensions, but can't be done with a single straight line through the plane. And this is particularly true if you look at the top two distributions here with the, the half moon and the embedded circle. It's clear that blue and red can easily be separated, but No line can actually do that. And so the key challenge for many machine learning algorithms is to, is to figure out how to represent these types of distributions in a way that makes drawing a single separating line between the two groups straightforward. And that's really accomplished with neural networks as described by warping the dimensions until you can separate the two groups in a very simple way. I won't go into detail in this slide, but despite what you may have heard about how inaccessible deep learning tech is and how high-scale and demanding the compute is, I did want to emphasize that with modern Python libraries like scikit-learn and TensorFlow and PyTorch, machine learning, deep learning tech is actually surprisingly accessible, and most problems can get started at a pretty small scale. Scales. And this is a piece of code that I won't walk through, but is actually a fully functional system to load the Iris dataset I've been talking about, to transform the input dataset, to divide it up in a way that lets us do the training and testing in a fair way, and then fits one of those single-layer neural networks to the data and produces a very accurate classifier. And this all runs you know, essentially instantaneously on my laptop these days. So I wouldn't let the general narrative of the community that machine learning and deep learning tech requires huge amounts of investment to discourage you from getting started because it's actually surprisingly accessible. So this is really the first slide that introduces the concept of deep learning and really deep learning is is conceptually pretty simple. It's an extension of that shallow neural network, which is here. You can see the input layer. This is where the image gets fed in, and then these hidden layers, 1 or 2, which is what I used earlier to model the iris. But the thing with the deep neural networks is it just keeps getting— goes on more and more layers. And modern neural networks have you know, for image recognition might have 50 layers or 100 depending on how you count them. And they're very complicated. They have all sorts of different properties, you know, that I'm not going to go into right now. But the fundamental thing that's exciting about it is represented here. There's just very deep network. It has a lot of learning capacity. And because it's able— because we have algorithms today that, you know, work very well to train these out, these sorts of networks against in particular when you have lots and lots of input data labeled with the sort of outputs you want. It can take hugely diverse data types, things from imaging of the chest to time-based assays like an EKG. It could take genomics data, it could take all sorts of things. We have machines that can basically consume that data in a pretty raw form, learn important attributes about it for predicting different, you know, for instance, medical aspects of like, do you have, you know, tuberculosis in a chest X-ray? You know, do you have an A, you know, do you look like you might have AFib risk in the, you know, the EKG or multimodal input? You could also be looking for SNPs that are associated with the trait. And because it's very deep, it can really learn how to take the very raw inputs and turn them into those representation in a higher order that separates the two groups that you might be learning, and this allows you to make quite complicated predictions pretty effectively. And you know, when we say deep networks learn to construct the features of their input data, what we're really saying is, and you know, if you think back to that two-dimensional version of the iris, you know, here we just have many, many dimensions. The inputs are some large image, maybe 300 by 300 in size, so it's already enormously large number of values, not just 2. And it's trying to learn from that just input how to extract basic information to say learn how to discriminate a cat versus a dog versus a motorcycle, which is what you see in this ImageNet. And GoogleNet was actually a pretty old now deep learning-based image classification model. It's very, very deep. You can see this schematic representation above. And when it's fully trained on something like an ImageNet challenge, which is millions of images among thousands of labels, it actually learns at these different layers things that are clearly interpretable— edge detectors, things that represent different types of textures or different types of patterns. Different parts of things, and it all kind of goes up into this sort of object-level stuff, which is ultimately like the thing that is fed into the final discriminator, which is just sort of a simple sum of the values in this final matrix that goes in. And that allows it to learn very effectively how to, you know, recognize images and recognize objects and images, you know, identify draw bounding boxes around specific objects and images. It's an amazingly powerful technology and it does this by basically learning these sorts of features dynamically from the data, which is incredibly powerful. With that background in machine learning and deep learning under our belt, let's now turn to how those technologies are being applied in the healthcare setting. The growing excitement for applying AI/ML technologies into healthcare really arises from the observation that there's quite a lot of things we want to predict in healthcare. You know, your risk of dying of a heart attack in 5 years, whether or not a pathology slide says that you have cancer, where there's a very complicated relationship between the things you can observe and the outcome you actually truly want to predict. Which makes the problem very challenging. And therefore, our previous approaches before the big breakthroughs in deep learning in particular made that— meant that we only had pretty limited systems that were not particularly accurate, especially when compared to, you know, the performance of human beings doing those predictions. And so, we're in a situation where the advances in machine learning that we see in technology application area where there's been large improvements in prediction quality. We have reason to believe that when we apply those into the healthcare space, we'll have equally large improvements in those predictions. And that hope that we'll be able to— we're at the cusp of a major increase in the quality of medical predictions that we can make and coupled with the large and growing body of healthcare data means that we see a future for healthcare where there's a large number of very powerful AI/ML predictors throughout the system doing their best to make the most of the growing body of data that we're collecting to help guide health decisions for individuals. Medical imaging really is the killer application of AI/ML tech in healthcare. Before I get into the, the why it is the killer app, let's just talk about what we're learning here. And in the image here, you can see really the key aspects. There's a variety of different types of medical imaging from, you know, pictures of skin issues to pathology slides, you're using deep convolutional neural networks like GoogleNet that I talked about earlier and learning to predict things about from those images. The example would be in the pathology slide, you would be looking at these pictures of cells and trying to decide if cancer was present in any of the cells in the image. And this has been a hugely impactful area and is really the leading application of AI tech, and that's really driven by three key issues. One is deep learning research itself arose in the field of image classification, which meant that there was quite a lot of expertise in the computer science field that could be leveraged as we headed into the medical application area. So there were systems built already to train these models, to use them for inference, there was just deep expertise in how best to approach that. And, and that could all, uh, give the medical imaging field a leg up. Um, imaging is also a very common modality in healthcare. You know, there's 4 different types of, of images here, but there's easily, uh, many, many more. And each of those, uh, specific data types resulted in a, you know, research group focused on, on making the most of that data and, you know, had associated commercial activities, and that just grew the field very quickly. There were many people who were able to productively do work in parallel. And finally, the imaging datasets are actually relatively accessible. Obtaining large amounts of retinal images or pathology slide images was actually much easier than some of the other data types in healthcare, and that really accelerated the field. So now let's look at electronic medical records. Certainly an equally exciting application area as medical imaging, but as I noted earlier, significantly less advanced due to the sort of structural advantages imaging has. In this scenario, what you're trying to do is look at all of these relatively unstructured and free-form, in some cases, medical records across many institutions, learn something about the temporal sequence of events in those in those records, as well as from the exact sort of values in, in lab measurements that might be present, to answer all sorts of questions about the future for individual patients. And you, if you can do a good job at predicting from those outcomes, you know, what is likely to happen in the future, for instance, what's the chance that they'll be readmitted If we discharge them today, what's the chance that they'll have a heart attack in the next 5 years? This allows you to answer sort of higher-level questions that are extremely valuable and important and, and honestly quite difficult to do today. You might be able to ask questions like, what is the value if we were to intervene and give this person a statin to reduce their risk of heart disease? You'd be able to look at their medical record and potentially identify diagnostic tests that would be, you know, maximally useful to guide treatment or guide patients toward less risky trajectories. And you can imagine a whole bunch of tools built on top of that to improve the experience of doctors. And so there's quite a lot of growing applications in this area. And I would read more details in our review article if you'd like to see the individual applications, but it is a very exciting opportunity. Community. So that's a very exciting area. Another big application area for deep learning tech in healthcare is in the genomics data. This is in part because again, there's so much rich data in genomics from the explosion of sequencing in the last 10 years. So we have all sorts of of deep learning tech from gene expression analysis and learning the regulatory code in the genome, which determines, you know, which proteins are produced when and in which cell types, to tools to analyze the actual output of these instruments much more powerfully and efficiently and automatically than we've been able to do in the past, as well as learning to identify regions in the genome of the conferred risk to disease. And these are all done again using these deep neural networks trained on a variety of different types of problems. You know, as I was saying, all the way from experimental data interpretation to identification of pathological mutations and things like recommending different drug therapies based on different variants in your genome. And the final area that I wanted to highlight at the end of the talk, which is really the main area of my startup, Big Hat Biosciences, which is using AI technologies to help us build better molecules. And in this world, what we're doing is we're teaching machines the relationship between the chemical code, it could be the small molecule code, if you're a startup like Atomwise or a tool like DeepChem, to the area that we work in, which is proteins and antibodies where the mutations are in DNA. And what you're trying to do is evolve, you know, you're trying to mutate a gene here to have, you know, better properties. You know, it might be more stable, might bind more efficiently, might catalyze a reaction faster. And traditionally, this has been done through physical, so-called directed evolution approaches in the lab. And that can be, typically it can be very finicky and challenging to do. And new AI approaches and coupled with synthetic biology is really allowing us to really see if these AI approaches can yield finally the ability to rationally and reliably design molecules, biological molecules that have properties that we pre-specify and can with confidence hit through a really heavily computational approach. I'd like people to leave with 4 big takeaways from this talk. One is that machine learning is a very powerful tool for relating input data to output labels even when that relationship is extremely complicated. And although classical machine learning already required limited human experience, that's been supercharged by deep learning with its ability to automatically learn features from particularly complicated data. And because of that, the data richness and complexity of healthcare problems looks like an excellent area to apply those techniques over the coming years. And Since it's going to take quite a long time to bring these to, all these technologies to bear on all the different problems in healthcare, it's a great time to get involved. So we're just at the start of this journey. If you have an interest in this field, I couldn't recommend more strongly to find a problem, grab Python, and jump right in. So I'd just like to end with a thank you to the organizers of the conference for inviting me to speak. I hope everyone has enjoyed the talk.