Tell us about your background.

I always wanted to contribute to the understanding of how diseases work, or, better yet, contribute to developing new therapies for treating them. At the same time, I was never interested in the purely academic route, as I was always more attracted to the translation from new academic findings to products that can be used in practice and affect the lives of patients.

Having said that, my journey to data science in biotech and big pharma – and the AI/ML space – was a somewhat circuitous one!

I was interested in several areas of science, like physics, chemistry and biology, as well as math and computer science. All my interests came together rather neatly during my PhD at the National Institutes of Health and University of Edinburgh, where my area of focus was Computational Biophysics.

That led me to join Novartis as a Presidential Postdoctoral Fellow, and then join Pfizer for eight years or so, and, ultimately, Korro Bio, where I head up data science. Each of these experiences was distinct, but AI/ML and statistical data analysis have been unifying factors.

What got you interested in artificial intelligence and machine learning?

It began with a fascination with computers in general. As a kid, I spent countless hours taking computers apart and rebuilding them. That glorious blue BASIC prompt that I saw every day between ages three and ten on various family computers, from our ZX Spectrum, then Atari, and Commodore 64, is firmly etched into my memory. I did a good amount of coding as a hobby, predominantly in elementary school, so these experiences were a good foundation for where I’ve ultimately found myself all these years later.

Using computation and AI/ML for scientific inquiry was a no-brainer for me, although when I started my doctoral work in 2008, AI wasn’t nearly as popular or accessible as it is now, and far less common to use for modeling and analysis in biology. Still, I was able to use Bayesian modeling in a couple of my PhD projects and during a short postdoc at the University of Cambridge. Those initial positive experiences piqued my interest and propelled me into using increasing amounts of ML and eventually AI methods in my further work. I was, and continue to be, struck with the wide applicability of AI/ML.

And how is this applied to drug discovery / development?

AI has become extremely widespread, and drug discovery is no exception. At Novartis, I was able to use ML to understand and improve the ways that compound libraries can be selected for assays to optimize the exploration, or exploitation, of already existing knowledge depending on the goals of the study. At Pfizer, the emphasis was much more on which targets are responsible for a given pathology, as well as what pathways should be perturbed by a specific therapy.

Now at Korro Bio, we use ML as a component in our design process to identify oligos that help with the recruitment of ADAR, the enzyme that carries out endogenous RNA editing. It has been fascinating to lead the group at Korro and see the progression from more standard (but still powerful) ML models, like XGBoost, through deep learning models, like convolutional neural networks, to newer model types, like graph neural networks. Not only did the accuracy of our models increase dramatically, but we have also increased our ability to create models that generalize much better to new chemical modifications.

There are many other possible areas of application. At Korro, we have been using AI to assist us in understanding the structure of our synthetic oligos and their interaction with the RNA targets as well as the ADAR enzyme. ML is a tool that can support us in correlating various types of outcomes with input data or for unsupervised exploration of available data distribution (think data clustering). These and other general categories of ML can be applied to several different tasks in drug discovery and beyond, and indeed, we have been seeing a true explosion of applications of AI in big pharma and biotech.

What sparked your move to Korro Bio?

First, Korro’s RNA editing platform is truly fascinating. Our technology is based around designing oligonucleotides that help direct the ADAR enzyme to the desired target to effect A-to-G editing with high specificity and efficacy. The opportunity to contribute to a platform like this was a key reason behind my decision. As you can imagine, at a large company like Pfizer there are numerous opportunities to contribute to great scientific work, but there is less emphasis (and velocity) around developing new technologies.

Second, developing the Korro product immediately struck me as a high complexity task that could benefit from the application of machine learning. It’s not “just” a regular problem that’s amenable to ML, but in fact a number of interconnected problems at the intersection of chemistry and biology where computational analysis can be truly enlightening. The problems we’re tackling at Korro have been very intellectually stimulating and rewarding from the ML perspective, so I have not been disappointed.

Finally, the Korro team is extremely smart and motivated. I immediately had the sense that not only was Korro working to solve important problems, but that the company has what it takes to get this potentially transformative technology through the finish line to commercialization. After talking to our CEO, Ram Aiyar, I immediately “got it” and knew that Korro was a place I wanted to be, and that I wanted to help them make the editing technology a reality.

How does Korro’s focus on RNA editing differ from other technologies such as antisense RNA or mRNA?

RNA editing can lead to either increased or reduced function, which sets it apart from many other therapeutic modalities.

With both antisense RNA and mRNA therapeutics, continuous delivery is required. Much of the exogenous mRNA can get degraded, hampering translation. ADAR editing is a fully endogenous mechanism that uses synthetic oligonucleotides only to hone its specificity and efficacy towards a selected target.

RNA editing also has distinct advantages over CRISPR-based gene editing, especially with regards to safety. CRISPR gene editing can still frequently lead to full or partial chromosome loss! To me, that’s a scary prospect, and, as you can imagine, it may also lead to severe side effects that stem from an irreversible change introduced at the DNA level. In RNA editing, the ADAR machinery is much more precise, leading to single point changes, with minimal off-target effects. Moreover, any off-target effects are reversible, as the DNA stays intact, so even if we accumulate off-target editing in the subject RNA, it will eventually get cleared from the target system, and the original unedited RNA will be transcribed from the intact DNA again.

What kinds of diseases might be appropriate targets for this type of approach?

RNA editing provides a wide range of possibilities and allows us to address a wide spectrum of diseases. The potential of the platform is truly spectacular – it can be used in a straightforward situation where a specific variant is associated with a given disease and we can simply “repair” it, but also in the more complex scenarios, where we can introduce missense mutations that alter the function of the resulting protein, ultimately leading to a different phenotypic outcome. So, the applicability is very wide, spanning simple monogenic diseases as well as much more complex phenotypes. At Korro, we are currently focused on difficult-to-treat diseases affecting the liver and central nervous system that are driven by disease-causing sequences that would be responsive to the safe and targeting editing of RNA.  

How do you apply AI/ML to the discovery/development process?

At each position in the oligo, we can have insertions or deletions, or chemical modifications at the level of the linker, sugar, or base – very quickly, that leads to a combinatorial explosion of the design space, which makes it a perfect problem to navigate with machine learning. Korro has generated a treasure trove of data that allows one to correlate the structure (and structural modifications) of a given oligo with its level of editing. We have been able to build well-performing deep learning models that leverage that data, along with the structural information about the oligo, its molecular interactions with the target, and the ADAR enzyme. What has been especially interesting to see is that the relatively new graph neural networks have thus far provided us with the most effective deep learning models correlating the structures of oligos with the editing that they enable.

What sort of savings does this lead to in terms of the development process? Are there experiments you can avoid by applying ML?

Absolutely, ML has been a true money and time saver.

We have been able to create very accurate graph neural network models that can even generalize to chemistries that haven’t been previously tested. With that, we can create whole libraries of oligonucleotides and prioritize unusual and previously untested chemistries, as well as pick designs that maximize the editing.

This has, in fact, truly allowed us to push our programs forward. Without this effort, we would have needed many more rounds of iteration to arrive at our lead oligos. What we are currently exploring is how translatable the modification choices will be to other targets, and if other types of chemical modifications are needed in those other targets, ML combined with molecular modeling will help us understand that translation.

Because of the role that ML plays in translation of our findings to new targets, it has become an integral part of our platform.

What hurdles does AI/ML still face in its application to drug discovery / development?

There are a couple of hurdles. Most AI/ML systems are quite literal in their capabilities. What I mean by that is that, like most computational systems (or computer programs in general), ML models do exactly what they’re told (or in this case, taught): given an input, they produce the corresponding output. So, ML models that are trained based on in vitro data on interaction between small drugs and their target protein will at best be able to predict very well how strongly new drug molecules will interact with that target. It will not necessarily translate well to a biochemical cell assay that measures a similar readout, and there can be various reasons behind that, including compound stability and the toxicity of that compound. Similarly, moving into an in vivo system can exacerbate those differences further, for example due to target differences between the protein used in vitro, and the one expressed in the animal model. That will be the same when moving into the higher species, and finally into the clinic. Technically, the correlation between in vitro and various stages of in vivo testing could also be learned by a more involved model, but that data is typically more sparse. So, the biggest problem that I still see in the way that ML relates to drug discovery, or more specifically to helping a new therapy make it to the market, is the lack of all necessary data.

Even with recognizing the challenges that are ahead, in our 2023 Cell and Gene Therapy Survey, respondents were bullish on the continued impact of AI in drug discovery. Where do you see the field going? Where might it be in five years?

We are at a turning point of AI/ML. Large language models (LLMs) like ChatGPT, OpenAI and Bard are being accessed by millions of individuals, most of whom are not ML experts. I think that turn towards generative AI will dominate the landscape, in conjunction with interconnected services, where we will rely on something akin to a GPT model that will be able to use chains of other computational tools (this is already being realized for example by LangChain) to augment its responses. I think that combination of specialized approaches and analyses with a system that integrates them will be a powerful direction over the next five years.

‍