Enveda Bio: Viswa Colluru
"Smart, ambitious people like working on hard things."
“Bad writers have influences. Good writers steal” [Stirling, 1953]. Perhaps the same can be said of drug developers.
Roughly one-third of FDA-approved medicines have structures taken directly from nature. These “natural products” (NPs) have evolved over billions of years in the microbial, plant and animal kingdoms: “[Natural products] gave us morphine, aspirin, warfarin, metformin, quinine, penicillin, cannabidiol…drugs that have touched billions of lives,” explains Viswa Colluru, CEO of Enveda Biosciences. NPs often have greater structural diversity, complexity and biological activity—promising traits when hunting for a novel therapeutic.
Stealing from nature seems like a good idea, but is easier said than done: “the most interesting chemistry is that which has spent 4 billion years evolving with life systems…if you can figure out where to search, and how pull out the needle in the haystack,” says Colluru. The difficulty lies in the vastness of the problem—with a universe of billions of compounds, how does one catalogue and evaluate each for bioactivity? The largest existing NP repository (COCONUT) has structures of only 400,000 molecules: “currently the hope for most patients comes from medicines that represent a tiny, tiny fraction of chemical space—a space that a small number of human chemists have fashioned in a lab. This is not at all inspiring, and actually anxiety inducing,” explains Colluru.
Is there a better way to search nature’s universe of small molecules to find new medicines?
Co-founded by Viswa Colluru (CEO) and Pieter Dorrestein in 2019, Enveda tackles this question head on. The company is using a variety of machine learning-enabled approaches (including transformer models) to identify promising natural products directly from complex mixtures, rather than isolating compounds one-by-one. While AI-powered metabolomics is the focus of the company’s platform, they are also steadily progressing their pipeline of immunology-focused assets towards the clinic. As Colluru explains: “The value of any drug company lies not in the approach, not in the story, but really in the pudding, which in this case are drugs.”
Enveda is currently progressing 3 molecules into clinical trials slated to start in 2024. The first is a novel gut-restricted inflammasome/NLRP3 inhibitor to treat Irritable Bowl Syndrome, which functions through a unique mechanism of action (MOA). The second program inhibits the aberrant neutrophil migration that may underlie many inflammatory diseases, but will be evaluated first in atopic dermatitis. In April of this year, Enveda raised a $51M series B1 financing to advance their pipeline and platform, led by new-investor Kinnevik with participation from Henry R Kravis, Dimension Capital, and FPV.
The son of a pharmacist from Visakhapatnam, India, Dr. Viswa Colluru is uniquely positioned to lead Enveda. He has a deep appreciation for the power of natural products, and his training in immuno-oncology (UW Madison) and computational drug discovery (Recursion), have given him the right tools to captain the ship: “I felt that I was in a great position, given my cultural upbringing, my life experiences, my drug development expertise and naivete combined, to give this a real shot.”
Viswa and team have a tall order ahead: they have a pipeline full of competitive indications also being pursued by larger biotech and big pharma. Though they have made promising first steps, characterizing NPs for drug discovery using transformer models is an area that is still developing. Lastly, there is a dogma in the biotech industry that NPs are difficult to work with when compared to synthetic molecules (e.g. they have challenging chemistry to synthesize and thus manufacture).
Imbued with an explorer’s mentality, Viswa leans on a lesson learned at Recursion Pharma: that “smart, ambitious people like working on hard things.” The drug development community, and a universe of billions of undiscovered natural products, are waiting to see what the team at Enveda finds.
1. What first got you interested in science and medicine?
Growing up, medicine and biology was not on my radar. At the top of the pecking order very firmly sat computer science. But there were a couple of inflection points in my life that got me interested in learning about molecules that may help suffering patients.
My dad owned the city's oldest pharmacy, which he inherited that from his maternal uncle. The pharmacy was established in 1949 and had been running continuously since. It was next door to one of the state's largest public government funded hospitals, which is where most of India's masses get treatment. We would see a lot of patients come in, and I would hang out frequently with my parents in the pharmacy—daycare wasn't a given back then. I remember distinctly thinking as a child: what is the difference between cetirizine and cetirizine hydrochloride? Different companies would mark the drug name differently. There was no Google back then, so I was a very annoying kid, especially when my dad couldn't answer these questions. I always had this background, but never thought I would make a career, let alone a life out of it.
It hit home how much this work [drug development] matters when my mother got diagnosed with leukemia in the late 90s. In 1998, she got diagnosed with Philadelphia chromosome positive, chronic myeloid leukemia. There was no treatment for CML back then--she got interferon alpha injections and taxane chemotherapy. But in 2001 she got expanded access to imatinib as it was getting approved in the US. A three-month course put her into complete remission. My mom won many, many battles, but ultimately lost the war—but reflecting on her journey, Imatinib had given her a new lease of life and extra years. I grew up to see her flourish from being a homemaker to being an entrepreneur that started multiple homebased businesses. I decided that if I can deliver the same kind of hope that we got from Imatinib to at least one family, I would have lived a phenomenal life. I started pivoting my career away from computer science, and into thinking about biological research—trying to walk down the sacred path of making new medicines.
2. What led you to graduate school and then Recursion?
When I started studying biology, the loss of my mother was fresh in my mind. My favorite textbook in undergrad was a book called The World of the Cell by Wayne Becker. My second favorite book was Lehninger’s Principles of Biochemistry. They're both written by professors at the University of Wisconsin in Madison. So I thought: “let me go to Madison to learn biology and biological research.”
The last chapter of The World of the Cell was about cancer, and there was half a page on this intriguing early concept called “immunotherapy.” It said if we could truly engender an immune response against tumors, you could achieve lasting remissions. I was inspired to work on a blue-sky opportunity like immunotherapy, since it could help patients like my mom. So, I found the one professor who was truly doing immunotherapy [at UWM], and ended up studying nucleic acid vaccines and checkpoint blockade for prostate cancer—long before any of those themes were “cool.” But by the time I graduated, through absolutely no contribution of mine, immunotherapy went from being persona non grata to the MVP. The grass =, which was brown when I started, had turned completely green.
The lesson I learned from my PhD is that we tend to think of innovation as novelty. In many people's minds those are two equivalent ideas. But many times the ideas that grow towards scale, are those that have been tried, attempted unsuccessfully, or explored in an eerily similar form, decades or centuries ago. If you think about all the things that are changing how we live today--minus transformers and ChatGPT—most of them have been worked on for decades before being “cool.”
But most scientific grants seem to be given to things like “new modalities,” where you haven’t even heard of the idea before. So after graduate school, I wanted to actually go work on an idea that was viewed unfavorably for years prior—because I actually think these ideas are most likely to mature and, like immunotherapy, change how we live and treat patients.
The only other criteria I had was that I needed to understand the idea well enough to explain it to my five-year-old cousin. Recursion happened to fall right in the center of those two intersections. In 2016, most people were writing off the emergence of big data generating technologies and lab instruments, and our ability to analyze them using cutting edge machine learning, as a fad. They [critics] said: “computer aided drug design, sequencing the human genome, have not proved to be the panacea we expected. We don't believe that just getting more data about cells should matter.” One of the most important things that I realized when talking to Chris Gibson [CEO Recursion] and others like Ron Alfa who was head of product, was that Recursion was different. This was because all the other times we used computers to make calculations or to apply rules that humans thought were correct to larger datasets. Instead, Recursion was allowing algorithms to freely define their own rules and find new patterns in data. And I thought this was very exciting. I spent a day interviewing at Recursion—at the end I was like: “you guys do facial recognition for cells.” When I told my five-year-old cousin that, he agreed it was cool. So it was decided—I moved from Madison, Wisconsin to Salt Lake City and started work at Recursion.
3. Lessons learned from working with Chris Gibson and others at Recursion?
There are a few different ones about companies and about people. I learned that really smart, ambitious people like working on hard things—often they like working on seemingly impossible tasks. One of the biggest mistakes you can make as a founder is trying to sell “certainty.” This was very surprising to me at first. The second thing I realized at Recursion and then definitely honed at Enveda is that: there is no mystical secret sauce to building great products and great technologies. Most humans are just one step away from trying to reach for the stars. What you have to do is remove the things that demotivate them: help set clear goals, have clear principles on your culture, don't get bogged down in bureaucracy. I think most people shoot for the stars naturally. Lesson number three is that companies often behave differently than the individuals that comprise it--even though companies are of course made up of people and we can all agree that if you take all the people away it won’t exist. Many companies are naturally risk averse. One of my mentors once told me: if you are sitting in the boardroom of Nokia and thinking about touchscreens (which they actually did think about) and the idea worked, everybody would take credit. But if the idea failed and you are the VP that championed it, your career may be over. Success has many fathers, failure has one. This is definitely how large companies operate. There is a lot of “best practice” in pharma that is very well deserved; “earned secrets” as Marc Andreessen calls them. Many other practices are just dogma, because they happen to be the safest possible set of parameters for you to throw it over the wall to the next vice president. From Recursion, I realized that some conventions—for example the single target single molecule hypothesis—may only be true for a very tiny number of approved drugs.
4. How did the founding of Enveda come together? What is the company’s mission?
Biology is not a set of linear cause effect relationships that you can use to pick a magical protein to ameliorate a disease—or at least this may be the exception not the rule. I realized that we need ways to effectively represent biology in its true more complex, wholesome form. The data was overwhelmingly convincing for me as a scientist. Even if you have a wonderful representation of biology of a single cell in a plastic dish, it may not validate in humans. I started to frame the problem of the industry as a lack of scientific research translating to humans. As Charlie Munger says: “start with a simple idea and take it very seriously.” So, when I took this framework [lack of translation] seriously, I came up with Enveda.
One of the most obvious things going back to my roots as an Indian, growing up in a pharmacy, was why don't we just start with a search space that has a high chance of working in people. These types of molecules that gave us morphine, aspirin, warfarin, metformin, quinine—drugs that have touched billions of lives. Where I grew up, alternative medicine was not that alternative—it was just medicine. Traditional medicine, based largely on the world around us, has been around for 5000 years, whereas modern target-based drug discovery is a little over acouple of decades old. At a distribution level I wondered: will looking to nature provide us a better starting substrate for drug discovery?
Imagine every malady you know and imagine every person in your life. The combination of what could affect anyone is massive. Currently the hope for these patients comes from medicines that represent a tiny, tiny fraction of chemical space—a space that a relatively small number of human chemists have fashioned in a lab. This is not at all inspiring, and actually anxiety inducing. The most interesting chemistry is that which is spent 4 billion years evolving with life systems. If you can figure out where to search, and how pull out the needle in the haystack, it is an area that no one else is really pursuing. Today, we [as humans] have probably have explored less than 1% of the chemistry on the planet. I also felt that I was in a unique position, given my cultural upbringing, given my life experiences, my expertise and naivete combined, to give this idea real thought.
[On the specific barriers to founding Enveda]
In the early days everybody told me this idea has been tried before and it failed, so you should leave it. Little did they know, this was exactly what I was looking for [Q#2]. Most of the reasons were perfectly plausible technical ones: natural products would mess up screens, frequently give us false positives, or these compounds would be hard to purify or isolate. Even if we managed to do these things, the third layer of insult was that we probably wouldn't like what we found—molecules that aren't amenable to structure activity relationship [SAR] or were so rapidly metabolized that they would never be medicines.
But the underlying truth, it turned out, was that we didn't really know what was in nature to begin with. We didn't have a way of finding out which one of these billions of natural molecules, or even thousands in a single plant, would make a medicine. Some had undertook this Sisyphean effort of trying to make a synthetic-like library of natural products by isolating each in an individual well and testing them. A different approach was called bioactivity guided fractionation: start with a mixture of many compounds that you find has activity, and then try to isolate the active compound from this mix. Naturally, they would end up with the most active, or the most abundant, or frequently both. This led the field to a bunch of promiscuous, ugly structures that became “characteristic” of natural products.
Even learning about these failures, I felt that natural products were at once the most validated idea and the most untapped idea. Validated because over a third of all approved small molecules come from metabolites or their derivatives. Untapped because 99 plus percent of the world's metabolites haven't even been cataloged by science, let alone tested in a drug discovery experiment.
5. What was the process by which you found your first lead drug discovery candidates?
In making new medicines, you can either go after well understood or highly desired biological targets that need new chemistry or you define new biology by running phenotypic screens. We did both right from the beginning. You can have a specific target you are interested in and deliver differentiated chemistry, or identify a biological phenomenon of interest and probe our libraries for activity. Our portfolio is some mix of the two approaches, with the most advanced programs being our oldest ones. We’ve repositioned some in terms of the indication spaces that we're going after, especially as markets have changed—it is now clear to have value a company needs to have a clinical stage asset, and likely a clinical stage asset with positive data. What we have done over time is increased our ability to find more programs, which has led to a very deep preclinical pipeline.
At a high level, the way the platform works is a sample from a plant is loaded on to a liquid chromatograph, which separates all the molecules based on a certain chemical property such as size or polarity. These are then fractionated out across a 1536-well plate. A small fraction of each well is then sent to our mass spec facility for structure prediction and the rest is used in a range of ~50 different high-throughput bioassays. Depending on the program, these screens could be phenotypic screens, target-based screens, biodistribution screens, or a combination. We then use AI algorithms that we developed in-house to link the activities seen in particular wells to the specific molecules within that well and their predicted chemical structures. All of this information gets integrated into our platform and is now searchable by our medicinal chemists through our GUI called Enveda Search. So now our drug hunters can go in and search for molecules with particular criteria such as active in this screen but not in that screen, available in this organ but not in that organ, etc, and Enveda Search returns a list of all the molecules we’ve unearthed with those properties. Only now, after doing this really thorough initial characterization, do we go back and do the hard work of isolating and testing molecules one at a time, which is a complete paradigm shift to how this kind of drug discovery was done in the past, which started with the isolations and one-by-one testing. The scale here is critical as I like the say, there are lots of needles in this haystack, but the haystack huge.
6. Briefly, what are the lead programs that Enveda is developing?
I learned at Recursion that if you're a drug discovery company, you need to have drugs. I am not a perfectionist by nature, and I consider that a strength. I knew that the platform was always going to look very different in four years than it did at 6 months. Despite a platform still in development, I prioritized launching actual drug discovery programs from day one—as long as they met the bar. My first hire was a lab and not a platform hire. As we grew, we scaled both teams in parallel. We began to apply some of the technologies from Pieter Dorrestein’s lab—Peter is one of the stalwarts of computational metabolomics and our scientific co-founder.
We wanted to create real value right from day one, so we quickly had to decide where to play. NLRP3, depending on who you talk to, is an interesting target. If you talk to Roche or Novo, they absolutely love it. If you talk to growth investors or Abbvie, they hate it. We just want to go after targets that our CSO likes to call “fundamental to disease biology.” If we can do that, then we can push off some of the more involved strategic decisions for later on.
With indications and targets I like to have a risk spread—in our pipeline MRGPRX, TGF beta have pretty low target risk, while NLRP3 is probably intermediate. We also have completely phenotypically discovered assets, like our inhibitors of neutrophil migration that work in both CXCR dependent and independent chemotaxis. These novel inhibitors show remarkable data across like nearly half a dozen preclinical models beating standard of care and even high potency steroids. These compounds have worked in models of lung disease, lung fibrosis, intestinal fibrosis, inflammatory bowel disease, multiple skin diseases; but as of now our plan is to take it these into atopic dermatitis.
7. Enveda recently released a pre-print detailing the development of MS2Mol, a transformer-encoder-decoder model, built for chemical drug discovery. Why are generative AI tools so useful for this type of application?
If you boil down the problem that our platform must solve, it is to take any sample from nature, and ask two questions: what are the molecules in the sample and what do they do? We are building technologies to answer both of those questions.
We realized that the current state of the art was isolating and purifying a single compound from a 10,000-compound mixture that you'll get from a humble leaf extract and doing an NMR to figure out the structure. With proteins, the comparison is to use x-ray crystallography to derive structure rather than alpha fold. If we wanted to find the structures of many natural products, we had two options: invent a cheaper, faster, better NMR or to use computational methods of interpreting data that already existed or could be easily generated. For us, the answer [and data] came in the form mass spectrometry.
All these natural samples can be loaded onto a tandem mass spectrometer, which orders them by size and charge, and then puts them in a chamber of ionized radiation where they break apart. Mass spec then captures the mass of all the pieces of a molecule after it's broken apart. Caffeine, morphine, cocaine, aspirin all have a unique molecular mass spec fingerprint. These fingerprints are extremely high precision and high-fidelity data. If you run the same sample on the same machine decades later you get the exact same data. So mass spec is high volume, high precision, high fidelity.
The only thing that mass spec was being used for previously was to fingerprint the molecules in a sample and retrieve all of the known molecules by matching fingerprints. It's like you go into a crime scene, you collect the fingerprints, you run it against an existing Interpol database. If you don't get a match, you are out of luck. In other words, chemists were using mass spec on samples to rediscover known molecules in nature. We asked: what if there was enough information in the way a novel molecule broke apart, that we can begin to tell what it is? Can algorithms interpret the signal without NMR confirmation?
The big breakthrough came when we realized that mass spectra were actually very similar to sentences in human language. So what do I mean by that?
In language the context of the word matters. If we read the sentence: “the animal didn't cross the street, because it was tired,” you have no idea what the word it means, unless you read the word tired. If I change that last word from “tired” to “crowded,” now “it” refers to the street. So the complexity is that every word can change the meaning of the other words around it, as well as the meaning of the sentence as a whole. This means that one needs algorithms that understand the larger context of information, and know which information to weight differently. Transformers have helped solve this problem for natural language. There is a parallel between language and mass spectra [shown in the below image].
The top is essentially is the “sentence” [mass spectra] and the bottom is the meaning [structure]. As more “words” or fragments are added to the spectrum, you see not just the size of the molecule change, but actually the nature of the molecule. This means that addition of one of these fragments is changing the meaning of the fragments that were there before, just like in language. When we realized this, it made us very excited, because we could extend transformers to mass spectra and the study of chemistry. An advantage of such transformer models is that they can be trained on large unlabeled data sets. You can have a transformer, read billions of English sentences, and predict missing words. We realized that there are hundreds of millions of unlabeled mass spectra that we could utilize for our model.
We asked: what if we just built transformer models that encode a representation of the mass spectrum and decode into a representation of SMILES strings ( Simplified Molecular-Input Line-Entry System, a way of writing a three-dimensional chemical structure in a linear manner)]? This is essentially what ChatGPT does when you ask it to come up with a block of text and then translate it to another language. So as illustrated below, I asked ChatGPT to come up with a poem about mass spectrometry and translate it to Hindi, which it does beautifully. This is effectively what we're doing for the mass spec and smile strings.
8. How did you benchmark MS2Mol’s performance against other tools? How useful is meaningfully similar matches for drug discovery?
The ultimate goal is to get exact matches on everything. The advantage to pursuing this is to create a high-fidelity chemical sequence of the world, like we are doing in genetically sequencing different organisms. If you have a model that gets very good at exact predictions, then what you actually have in nature, are millions of SAR campaigns pre-built for you. There is redundancy in biosynthetic pathways. So, for a scaffold of one of our leads, there are 550 known analogs across three plant genera. If you can identify those analogs right off the bat, you can see how activity varies in a natural 500 molecule SAR series that has been made for you over billions of years. This is what the second piece of the platform, which we haven’t discussed, is focused on: how we can confidently annotate what related molecules do in a particular bio assay without needing to isolate them individually.
But as of today, using our MS2Mol algorithm, we can predict structures that are 60%, meaningfully similar. This is still a massive improvement over going in blind or only looking for known compounds that have been previously catalogued. We are trying to now develop a minimum viable product, which I think is still revolutionary, because it gives us the ability to de-prioritize bad molecules. If you have the good, the bad and the ugly, we really want to pick up the good and forget about the bad and the ugly. We don't need to be exact matches or even close matches, to be able to call out the bad and the ugly. This will allow us to only spend our time focusing on characterizing the good molecules further.
9. What is another company you look to for inspiration?
I really appreciate the virtues of the Nimbus model. I think Nimbus has demonstrated how lean can be mean, and how you can create medicines that will hopefully also create a lot of value. They have gone about this in a very intentional manner. If you listen to Jeb Keipfer speaking about the origins of the company, it's very clear that they wanted to build small, high-powered unit enabled by computational physics. They have now done this twice. I think they've shown that they'renot just a one hit wonder, and have been thoughtful about pioneering corporate models to enable their M&A. I find Jeb very impressive as an individual and Nimbus to be inspirational as a company.
10. What is something you are currently reading or watching?
I'm reading a book by Arthur Clarke called Profiles of the Future, which was written in 1962. He talks about how scientists and experts are particularly bad at predicting arcs of technology. One of the sentences he writes is that: “they said, it couldn't be done,” is almost the universal story behind every great innovation. He talks about how the light bulb, alternating current, bicycles, and even cars—at one time we thought that driving over 30 miles an hour would lead to asphyxiation of the passenger – were written off. There were leading engineering professors that wrote off commercial air flight, space travel, and rockets. The interesting piece is that they did this even after people had actually shown that it was possible. There are a lot of great lessons for drug discovery embedded in this book.