Jimmy Lin is CSO of Freenome, which is developing blood-based tests for early cancer detection, starting with colon cancer. He pioneered computational methods for pulling insights from large-scale genomic data, and led the computational analyses for the first genome-wide sequencing studies in several cancer types across different tumor classes and datasets worldwide.
In this interview, Lin discusses the difficulties of carrying out a company mission that aims to unite machine learning with biological data. He describes the three kinds of people you need to hire to build a techbio company, the pitfalls to avoid, how to know when the union of two fields is or is not working, and the subtleties of adapting biological studies and machine learning to each other.
Why ML behaves differently on biomolecular data overall
Like many fields, there is plenty of excitement about applying machine learning to bio. Yet progress has seemed harder won. Is biomolecular data somehow different from the kinds of data usually used with machine learning, in practice and more generally than people expect at first?
Traditional machine learning data are broad and shallow. The problems machine learning often solves are the ones humans can handle in a nanosecond, like image recognition. To teach a computer to spot a cat in an image, you would need billions upon billions of images for training, but each image contains relatively little data. Biological data are usually the opposite. We do not have billions of individuals. We are fortunate to get thousands. But for each individual, we have billions and billions of data points. We have fewer examples, but each one is very deep. This means different modeling assumptions entirely.
At the same time, biological questions are less often problems humans can solve. We are working on things even world experts in this area cannot do. So the nature of the problems is very different, which means we need new thinking about how to approach them today.
Do the methods have to be built from scratch for biomolecular data, or can existing ones be adapted?
There are ways to featurize this deep information so you can use existing tools, whether statistical learning or deep learning methods. It is not a direct copy-and-paste, but there are many ways to transfer machine learning methods and apply them to biological problems even when the mapping is not one-to-one. That still leaves room for substantial reuse of ideas and workflows across domains in practice and in research.
Digging deeper into the data issue, biological data have a lot of variability—there is biological noise, there is experimental noise. What is the best way to create machine-learning-ready biomedical data in practice?
That is a great question. From the very beginning, Freenome has thought about how to generate the best data for machine learning. Throughout the process—from study design to sample collection to running the assays to data analysis—every step has to be handled carefully so the data are optimized for machine learning, especially when you have far more features than samples. It is the classic big-p, little-n problem for teams like ours in practice from day one onward.
First and foremost, we designed our study to minimize confounders. Many companies have leaned on historical datasets and worked hard to reduce cohort effects and strip out confounders. But is that really the best approach? No, the best approach is a prospective study in which you control for confounders from the start. That is why, even in our discovery work, we chose to run a large multisite prospective trial that gathers gold-standard data upfront, as in our AI-EMERGE trial from day one onward.
Fortunately, we have investors who believed in us enough to let us generate these data. That was actually a big risk, because these studies are very expensive to run overall.
Then once you get the data, what do you do with it?
Well, you need to train all the sites in a consistent way, and control for confounders across the different sites so the patients look as similar as possible. Then, once you run the samples, you need to think carefully about how to minimize batch effects, such as by placing the right mix of samples on different machines in the right proportions there.
This is very hard when you are doing multiomics because the machines that analyze one class of biomolecules may take hundreds of samples in a single run, whereas the machines that analyze another class of biomolecules may take only a few. On top of that, you want to eliminate human error. So we introduced automation almost immediately, at the stage of generating training data.
Also, when you have billions of data points per person, it becomes very, very easy to overfit. So we make sure our training generalizes to the populations we ultimately want to apply it to, using the right statistical corrections and many successive train-and-test holdout sets overall in practice too here
Why balancing ML and bio matters here
Combining machine learning with biomolecular data is something many biotech companies are trying to do, but often there is a lot of vagueness about how they will do it. What do you see as an essential feature of integrating them well?
At Freenome, we are blending machine learning and multiomics. To do that, you need to do both well. The key is having deep expertise in each, and then being able to speak both languages. You need to be bilingual in practice and in the lab as well every day.
There are many companies that are expert in one area and then add a thin layer of the other. For example, there are tech companies that decide they want to move into bio, but all they do is hire a handful of wet lab scientists. On the other hand, there are biology companies that hire some machine learning scientists and then declare that they are an AI/ML company now instead.
What you really need is deep bench strength in both. You need a deep biological understanding of the system, the different assays, and the features of the knowledge space. But you also need a deep understanding of machine learning, data science, computational methods, and statistical learning, plus platforms to apply all that in practice effectively there.
The three types of people you need
That is challenging because those two areas are often very siloed. When you are thinking about the people you are hiring for the company, how do you build bridges between these two domains?
I think there are roughly three types of people you want to hire to bridge tech and bio. The first two are the standard ones: domain experts in machine learning or biology. But they also need to be open and willing to learn about the other domain, or even better, have exposure and experience working in those additional domains as well too.
For machine learning specialists, we look for people who are not simply there to build the newest algorithm, but who want to take the newest algorithms and use them for biological questions.
Biology is messy. Not only do we lack all the methods needed to measure the various analytes, but we are also continuously uncovering new biomolecules and features. There are many confounding factors and sources of noise that must be considered as well. In general, these issues are more complex than standard machine learning problems, where both the problem and the knowledge space are much more clearly defined. ML experts who want to apply their skills in biology need humility to learn the complexity that exists in biology and must be willing to work under less than ideal conditions and with differences in data availability.
The other side is hiring biologists who think about their problems in terms of larger-scale quantitative data generation, design studies to optimize signal-to-noise ratios, and understand the caveats of confounders and generalizability. It is about more than simply being able to speak and think in the language of code. Many of our biologists already code and have a solid statistical background, and are willing and eager to grow in these areas. In fact, at Freenome, we actually have training programs for biologists who want to learn more about coding so they can develop their statistical reasoning.
Even more important is that study design, and the questions we are able to ask, look different when they are designed in the context of big data and ML.
What’s the third type?
The third type of person to hire is the hardest to find. These are the bridgers – people who have worked fluently in both of these areas. There are very few places and labs in the world that sit right at this intersection. Finding people who can translate and bridge both areas is very, very important. But you do not want to build a company made up only of bridgers because often these people are not the experts in one area or the other, because of what they do. They often have a more general understanding. Still, they do the essential work of connecting the two fields.
So having all three groups of people matters. If you have only one type of domain expert specialist, you’ll only be strong in one area. Or, if you do not have the bridge builders, then you end up with silos of people who cannot talk to each other. Ideally, teams should include each of these three types of people to support a deep understanding of both ML and biology as well as an effective synergy between these two fields.
Do you see differences in how specialists in tech or computation attack problems versus how biologists approach problems?
Yeah. At one extreme, we definitely have people who come from statistical and quantitative backgrounds and they speak in code and equations. We need to help them take those equations and explain them clearly so that a general audience can understand.
Biologists have strong imagination because they work with things that are invisible. They use many illustrations in presentations to help show what is happening at the molecular level, and they have strong intuition about mechanisms and complexity. Much of this thinking is more qualitative. This creates a different way of thinking and communicating.
So, the way people communicate is going to be very, very different. The key is – we sort of jokingly say – we need to communicate in a way that even your grandma can understand.
It takes real mastery of your knowledge to be able to simplify it so that even a novice can understand. I think it is actually great training for someone to learn how to communicate very difficult concepts without relying on the usual shortcuts, jargon, and technical language.
What has inspired your particular viewpoint on how to marry machine learning and biology?
So, the problem is not new, but rather the latest version of an age-old problem. When the fields of computational biology and bioinformatics were first created, the same problem was there. Computer scientists, statisticians, data scientists, or even physicists entered the field of biology and brought their quantitative thinking to it. At the same time, biologists had to begin modeling beyond describing genes as up-regulated and down-regulated, and start approaching the data more quantitatively. The digitization of biological data has now simply expanded exponentially in scale. The problem is more acute and broader in scope, but the fundamental challenges are still the same.
What do you see as either the success metrics or red flags that tell you whether or not the marriage is working?
If you look at companies that are trying to combine fields, you can very quickly see how much they invest in one side or the other. So, if it is a company where 90% of the people are lab scientists, and then they only hired one or two machine learning scientists and they are calling themselves an ML company, then that is probably more of an afterthought.
Why humility is key
Is there one take-home lesson that you have learned in this whole process of marrying biology and machine learning?
I think intellectual humility, especially when it comes from the tech side. With something like solving for search, for example, all the information is already in a text form that you can easily access, and you know what you are looking for. So it becomes a solvable problem, right? The problem with biology is that we do not even know what datasets we are looking for, or whether we even have the right flashlight to shine on the right areas.
So sometimes when tech experts jump into bio they fall into a trap of oversimplification. Let’s say, as an example, for next generation sequencing they might say, “Wow. We can sequence DNA. Why do not we just sequence lots and lots of DNA? It becomes a data problem, and then we solve biology.”
But the problem is that DNA is one of dozens of different analytes in the body. There is RNA, protein, post-translational modifications, different compartments such as extracellular vesicles, and differences in time, space, cell type, among others. We need to understand the possibilities as well as the limitations of each data modality we use.
What's exciting about applying ML to biology
While it may be hard to believe, biology is still a field in its infancy. We sequenced a human genome a little over two decades ago. Most of the time, we cannot access individual biological signals so we are still taking measurements that are a conglomerate or average across a lot of signals. We are just beginning to measure one cell at a time. There is still a lot to do and this is why it is an exciting time to go into biology.
But with that infancy comes great potential to solve problems that will have huge impacts on human health and wellbeing. It is a pretty amazing time because we are opening new frontiers of biology.
What sorts of frontiers? Is there a branch of biology or medicine where you are especially eager to see computation used?
Yes—everything! But let me think. In cancer, I think that within our generation the new therapies and early detection work now emerging will turn cancer into a chronic disease that is no longer so frightening, as we have done for HIV. And we can probably apply very similar kinds of methods to disease detection and prevention more broadly. The main thing I’m excited about is that we can begin finding out whether the disease is already present before symptoms appear.
Beyond cancer diagnostics, what’s also really exciting is the shift toward building with biology rather than only reading and writing it. I’m excited about the parts of synthetic biology where we’re using biology as a technology, whether that’s CRISPR or synthetic peptides or synthetic nucleotides. Using biology as a tool opens up broad possibilities to totally reshape traditional resource-generating industries, from agriculture to energy. This is truly an incredible time to be a biologist!