At Dewpoint, we are working to unlock the full potential of condensate science to discover and develop life-changing treatments for devastating diseases, which until now have been considered untreatable. To achieve our goal, we are navigating the astounding complexity of cellular systems, combining traditional wet-lab experiments with ground-breaking artificial intelligence (AI) and machine learning (ML) models to identify unique drug targets and accelerate our journey towards the clinic. 

To hear more about Dewpoint’s current AI approach and future aspirations, we spoke with Dr. Francis Carpenter, Dewpoint’s Head of Data Science and Engineering, who is leading the development and application of innovative, data-led approaches to enhance and accelerate drug discovery and clinical translation efforts. 

Before joining Dewpoint, Francis was a consultant for both McKinsey & Co and QuantumBlack, where he led cross-functional teams building tools to help biopharma clients leverage state-of-the-art AI approaches to enhance drug discovery and to optimize clinical trials. Prior to this, Francis was a Principal Product Manager at Genomics England and led the team that built the secure cloud-based research environment and bioinformatics workflows used by thousands of researchers to analyze the whole genome sequence data of >150k cancer and rare disease participants. Francis earned his PhD at University College London, in the lab of Professors John O’Keefe, Neil Burgess, and Caswell Barry, where he combined in vivo electrophysiology, chemogenetics and computational modelling to study the mechanisms underlying the brain’s representation of space and how this supports long-term memory. 

Dewpoint is working to develop therapies based on the science of biomolecular condensates – a completely new approach. How does the novelty of the science influence your drug discovery approach? 

Focusing on biomolecular condensates as potential drug targets is indeed a totally new approach and it is opening up many new opportunities for us to discover and develop drugs to treat diseases, which have until now eluded scientists. 

In a traditional drug discovery process, you identify a target protein linked to a disease and then look for or engineer a molecule that can bind to it, to turn it on or off or change its function. While this structure-based process is straight-forward in theory, in practice there are many targets, which, due to their complexity and high levels of disorder, it has not been possible to find drugs for. This approach is also insufficient in cases when the disease is not solely caused by the incorrect functioning of just one protein. 

Condensates are however quite different from traditional protein targets. Condensates are membrane-less organelles, that form and dissolve dynamically within a cell. They concentrate communities of biomolecules, helping either to stimulate, block or otherwise regulate normal cellular functions. However, in many different diseases, from cancers to neurodegenerative and metabolic diseases, we now know that condensates are going awry, driving disease by forming or dissolving inappropriately, disrupting the biochemical processes and signaling integrated within them. We call an aberrant condensate that drives disease a ‘condensatopathy’. To repair condensatopathies, Dewpoint is working to develop condensate modifying drugs (c-mods) that target the elements which influence the formation, dissolution, or physicochemical properties of condensates and consequently modify the disease. However, because condensates are composed of highly dynamic proteins rich in intrinsic disorder, a structure-based approach is often not feasible. Historically, many such proteins, despite their strong, validated ties to the disease, were deemed ‘undruggable’. Modifying the community within the condensate provides a different route to modifying the function of these previously ‘undruggable’ proteins that don’t have a known or potential drug binding site. 

We start by identifying specific target condensates that are causally linked to the disease we’re interested in and use high content imaging and fluorescent markers to characterize the condensate phenotype in both healthy and diseased states. We use the phenotype as a readout to search for c-mods that interact with the target condensate to shift the diseased state back to the healthy state. As the whole experimental process is automated and roboticized we can quickly compare hundreds of thousands of compounds to find the most promising hits. Once we have a potential candidate, we work with our chemistry colleagues to optimize its properties and make it more drug-like, ready for testing in preclinical animal models and eventually patients in the clinic. 

This condensate-focused phenotypic drug discovery approach has really flipped what a lot of people are doing on its head, and what’s exciting is that by using specific disease-causing condensate targets as a readout we are finding novel potential therapies that we couldn’t have found with a structure-based approach. Today at Dewpoint we have several programs that have produced promising preclinical candidates for diseases such as ALS and cancer, which scientists have struggled for decades to find therapies for with traditional approaches. 

You mentioned using the phenotypes of condensate targets to read-out disease states and to find c-mods. How do you go about finding these condensate targets? 

In some of our research programs we’re focused on targets that are well known to have a strong disease link and condensate phenotype, such as TDP-43, MYC or beta catenin. Many of these targets are traditionally considered ‘undruggable’. Our programs differ from other approaches taken in the industry in two fundamental ways. First, we treat the condensate where the ‘undruggable’ protein resides as the target. Second, we use a condensate phenotype-based approach to screen for compounds that modify the function of the ‘undruggable’ target.  

However, many diseases don’t yet have well characterized and validated condensate targets. At Dewpoint, we’ve also developed a scalable computational platform for identifying novel condensate targets. The approach combines human omics evidence of gene-disease associations with knowledge graphs and protein language modelling to identify potential targets that are closely associated with the dysregulated gene network underlying the disease state, and that are likely to phase separate into condensates. From this analysis we generate a shortlist of candidate targets for which we experimentally validate both the disease link and the condensate. We’ve now successfully used this platform in multiple disease areas, and we’re excited by its potential to help us get a foothold in understanding and discovering drugs for diseases that haven’t yet been well studied through a condensate lens.  

AI tools are now routinely used for drug discovery, but with the novelty of your phenotypic approach how are you incorporating AI? 

Many companies are using AI effectively to model the structure of known targets or molecules and to speed up the identification of drugs that can bind to them. But due to the compositional complexity and dynamic properties of condensates, these structure-based AI and ML tools alone aren’t sufficient for us. We’ve therefore had to look at different approaches and technologies and build new tools that appropriately take into account and leverage condensate biology. 

One area of our work in AI, which I think really sets us apart, is computer vision. Condensates are a network-level phenomenon where communities of biomolecules assemble. The condensate phenotypes we image are extremely information rich, encoding the biology and chemistry of their biomolecular networks. To make sense of this wealth of information, we have had to develop AI models to help us decode them. These models are based on approaches built by companies like Meta and Open AI, that are trained on pictures from the internet to differentiate between say images of cats and dogs, but we teach them to identify different types and characteristics of condensates. With the help of these models, we are analyzing the millions of condensate phenotype images we have collected for candidate drugs, to identify what changes in the condensates are linked to the therapeutic profiles we’re looking for. For example, we can identify what change in a specific condensate is typical of drugs that selectively kill cancer cells but not healthy cells. These models have already helped us to find novel compounds that show therapeutic promise but that we would have otherwise missed, because the way they affect the target condensate was unexpected. 

Another example is the models that we have developed to help our chemists ‘expand’ the chemistry around an existing hit compound, to find other compounds with similarly desirable phenotypic effects but more promising chemical properties. In one of our oncology programs, the high-throughput screen and our AI analyses together found a compound with a promising phenotype, but it wasn’t an ideal starting point for a drug because its chemical properties meant that it would be metabolized in the body into a different reactive compound. We developed ML-based methods to identify patterns in the high-throughput screen between the chemical structure of compounds and the phenotypes they induce, and successfully found compounds that retained their biological characteristics, while gaining promising drug-like attributes.  

These are just a couple of examples of the fully integrated, end-to-end data science platform we have built at Dewpoint, which we lovingly named ERSAi – after Ersa, the goddess of dew. ERSAi supports every step of c-mod discovery and development with predictions and analyses, from target discovery to c-mod identification and optimization into clinical translation. ERSAi includes multiple proprietary models, and is trained on the world’s largest condensate database, including >3 petabytes of data and >5 million condensate experiment images. The predictions and analyses from ERSAi are tightly coupled with our wet lab experiments, meaning that its models are continually improving. 

What are you working on next from an AI perspective? 

We’ve already incorporated computational methods and tools into every stage of the drug discovery pipeline, and these tools are helping us to accelerate and optimize the process. But we need to keep innovating, both to refine and improve our existing tools, and to develop or integrate new methods that help address other challenges in the drug discovery process. 

One area we are currently focusing on is a new model to help us better understand the molecular mechanisms of action and interactions of our potential drugs. The way we do phenotypic screening is condensate-target specific but mechanism agnostic. That is, we look for c-mods that have a specific condensate effect, but we don’t specify a single biomolecule target that it has to work through. This means that when we initially have a set of hit compounds, we know their effects converge on the aberrant condensate that drives the disease, but we don’t necessarily know their specific mechanism or how they’re interacting. We can of course find that out experimentally, and there’s lots of assays you can do to try and do that, but they’re usually quite time consuming, and you maybe only do them for a subset. So, we’re now working on using both phenotypic and structure-based methods to infer the type of molecular mechanisms and targets of hit compounds so that we can evaluate at scale each hit compound more efficiently. This will help us to prioritize and progress the hit compounds with the most promising molecular mechanisms and with the highest chance of success in the clinic. 

What are your goals and aspirations for AI and ML at Dewpoint? 

Despite being a biology-first company, I’m proud that we’ve already demonstrated our ability to build and apply cutting-edge computational methods. I believe that our team, under the guidance of experts such as Regina Barzilay (MIT) who sits on our scientific advisory board, can achieve remarkable results in this space.   

In the medium term, I want to expand the number of ways that computational approaches guide and inform our drug discovery process. That means continuing to refine the methods and the approaches that we already have, but also continuing to find and develop new models. I would, for example, like to find a way to better understand the structure and dynamics of the targets and the condensates earlier in the process, so that we can bring methods like structure-based drug design into the equation. We have some ideas about how this might be possible, and I’m excited to work with the team to find a structure-based drug design approach that works despite the challenges and complexities of condensates.  

Long term, I see condensates as having potential in a lot of different disease areas, and I’m really excited about the potential for AI and ML to help scale that up. AI gives us the opportunity to effectively identify and work on many different disease-driving condensates in parallel, allowing us to build a much bigger pipeline of potential treatments for a wider range of diseases. 

What impact do you think that AI will have on the drug discovery industry as a whole?  

I don’t see AI as a stand-alone silver bullet, but I do believe that combining AI with traditional wet-lab approaches will help us identify and design more effective and safer drugs and do so more efficiently. AI is already helping to accelerate the process of drug discovery and development, and I think the potential is there to help us focus on drugs with a higher probability of success, so that we’re not spending so many billions of dollars on trials that have may have been doomed to failure.  

Beyond the discovery of new drugs, I think computational approaches can help us make much more of the opportunities for repurposing drugs. There are a huge number of drugs out there with the potential to be used for other associated indications with common underlying causes, but it hasn’t been economically attractive or feasible to test them. AI gives us a way to open up these opportunities.  

I also hope that AI will allow us to tackle rare diseases more effectively. There is such a huge unmet need, but looking at rare diseases individually, it’s challenging to carry out traditional drug discovery approaches because of the cost involved relative to the economics of a small patient population. There is work being done to overcome this, such as by targeting aberrant condensates which are shared between patients of diverse backgrounds, as is the case for ALS. But to go further, I hope that combining the advantages brought about by leveraging condensate science with computational methods and AI, will really be a game changer here by allowing us to develop drugs in a sufficiently cost-effective way that will, over time, allow us to deliver treatments for these diseases too.