The biotech improving immunotherapies with AI

DDW Editor Reece Armstrong speaks to Ron Alfa, CEO & Co-founder at NOETIK, about the company’s mission to utilise artificial intelligence (AI) and spatial biology to develop better immunotherapies for cancer patients. Reece A: What are some of the human limitations to current drug discovery research for immunotherapies? Ron A: One of the bigger challenges is translatability from preclinical work to human studies. We have discovered promising molecules and mechanisms, including novel checkpoint inhibitors for example, but it has been difficult to understand the human tumour context where these new mechanisms have the potential to be most efficacious. There are a few fundamental challenges: First, we don’t have a complete understanding of tumour immune biology to define therapeutically relevant tumour subsets that are mechanistically relevant. While we have made a lot of progress in targeted therapeutics by identifying molecular tumour drivers that define patient subsets, genomics has not quite produced the same level of understanding for tumour-immune biology. Secondarily, preclinical work for cancer immunotherapies involves a panel of a limited set of mouse models that are used to advance most programmes. While these systems were developed to recapitulate some aspects of tumour immune biology, they do not translate well to specific human tumour biology. While they may enable us to identify and validate interesting biological findings, when it comes to bringing a programme to the clinic, they do not provide a lot of insight into patient selection strategies. As a result, we find ourselves trying to use clinical data from early human trials to do patient stratification, but these studies are underpowered to solve this problem without additional data. What we need is a more complete understanding of tumour immune biology that integrates molecular, cell and tissue spatial biology, and patient information to build a more complete atlas for precision immunotherapies. It is a challenging undertaking to integrate so much data, but contemporary advances in machine learning provide us the right tools to take on this challenge. From these data, we can discover new medicines with a higher probability for success. Reece A: In pharmaceutical research there is an abundance of data, much of which is untapped. Is AI / machine learning (ML) the best way to utilise this data at scale? Ron A: It is a misconception that most companies are sitting on treasure troves of data that can easily be tapped by AI methods in my opinion. We’ve learned over the years that to successfully deploy AI methods in biology, one needs to build datasets fit for purpose for machine learning. What that means is often making certain design choices in the data generation process that might result in very different approaches than datasets collected by more traditional methods. At NOETIK we are generating a massive atlas of spatial and molecular data from human tumours designed up front to train foundation models with the potential to unlock the complexity of tumour immune biology. In designing the approach, we’ve been very thoughtful about things such as batch to batch variation, generating many representations of each tumour sample, among many other considerations that are specifically designed to enable our machine learning methods to learn fundamental human tumour biology. Once models have been trained and validated rigorously on the right data, we do believe that there are opportunities to deploy these models for out of distribution inference on other datasets or bring multiple datasets together, but in the beginning, it is valuable to maintain a lot of control over the data generation process. Reece A: AI has typically been used to fast-track certain stages of drug discovery and development but NOETIK is using it to tackle fundamental issues in cancer. Can you tell us about this approach? Ron A: One of the most important problems to solve in drug discovery is improving the probability of clinical success. In oncology and other areas of medicine, significant progress has been made by advances in genomics to better define patient biology to enable targeted precision therapeutics. If you consider the impact of genomics epistemologically, it gave us a new way to see the biological world in order to discover new definitions of biology and disease. Where cancers were previously understood by their tissue of origin and cell morphology, we can now define them based on their genotypes and molecular drivers. This radically changed how we think about both cancer biology and also the treatment landscape because we could develop medicines targeting these specific genotypes and pathways, giving rise to precision therapeutics. In the context of cancer immunotherapies, we’ve arrived at some drugs such as immune checkpoint inhibitors that have proven really transformational in some patient contexts, but we don’t have a good mapping between tumour immune biology and response to these medicines. This is a complex problem because it involves both tumour biology, genomics, and immune biology. Unpacking a complex problem that involves multiple data types is the perfect setup for modern machine learning. We’re building a very significant multimodal dataset of human tumour biology to form the basis for an atlas of tumour immune profiles, or immunotypes. These immunotypes will be the basis for internal work at Noetik to discover and develop new precision cancer immunotherapies. Reece A: Using this approach how would you like to improve upon current cancer treatments? Ron A: Our mission is to discover and develop new therapeutics to patients with diverse cancers. We think the future of cancer immunotherapy is going to look a lot more like the current state of targeted therapeutics where we understand different types of cancers to be driven by specific immune signatures. In this future, physicians are able to profile tumours using state of the art immune profiling methods to deploy a diversity of therapeutics that can be matched to the right patients. This is an ambitious vision but I think a lot of people in the field recognise it is the right direction, and science and technology are advancing at such a pace to bring it within reach sooner than we think. Reece A: Earlier this year you raised $14 million in a funding round. Is that funding helping the company expand? Ron A: We were really privileged to be able to launch the company with $14 million in seed financing from a syndicate of incredible investors during an overall difficult fundraising environment. Biotechs tend to launch a bit more slowly from conception to proof of concept, but given the urgency and importance of this problem, we felt this was the right time to build quickly. Having spent half a decade building one of the leading AI biotech companies in the space, we knew that we could define a solid roadmap and execute quickly. Our goal for the seed financing was to get the first version of our data generation engine off ground with hundreds of tumour samples and to start training foundation models to learn tumour biology. In less than, 12 months of operations we were able to recruit a really exceptional team of veterans from some of the first-generation AI biotech companies and build the capabilities to generate and process a quite complex stack of human tumour data that includes standard pathology images, as well as multimodal spatial proteomics and transcriptomics. We’re thrilled with the team’s progress moving so quickly, but there is still a lot of work ahead of us as we start to translate these data into therapeutics. Reece A: You’ve analysed 1,000 human lung cancer samples within 12 months to understand the biological function of these tumours. How encouraged were you by these data and are you looking to grow the number of samples? Ron A: This has been an immense data generation effort and there are a few aspects that make this dataset incredibly unique. First, from the beginning, we designed the data production strategy with machine learning in mind and considered many factors that we knew were going to be important. For example, each tumour is represented many times in the dataset and we have methods in place to detect potential batch effects, and other artifacts that will confuse our models. Second, these data are really cutting edge in terms of what we’re capturing. The techniques and instruments that we’re leveraging have only been around for months at this point and we’ve had to build a lot of custom pipelines for processing this data. Third, having built AI platforms before, we knew data infrastructure was going to be incredibly important from the start. Even before we started to generate data, we had an engineering team in place building that infrastructure. We are only scratching the surface of analysis at this point but we’re already seeing some really interesting biology. For example, we’re finding tumour immunotypes that haven’t been described in the literature and different immune subtypes of major tumour populations that could inform therapeutics. There has been a decade of cancer immunotherapy development that has come about independent of these types of approaches, and now we are starting to look for these biological pathways and understand much more of the story. This dataset is a starting point for us - a proof of concept to build the platform – we are already expanding the biobank into other oncology indications and working on next generating methods to get massive scale from the platform. Reece A: Does immunotherapy represent the future of cancer care? Ron A: Absolutely. While there is still a lot of exciting science happening in targeted therapies, we are at the beginning of realising the potential of cancer immunotherapies in my opinion. We have only significantly explored the space of therapeutics opportunity for anti-PD(L)1 therapies and we’re seeing positive clinical outcomes in new populations on a weekly basis for these. Yet we still don't get great definitions of therapeutically relevant tumour immunotypes that are predictive of response. We think that mapping this third dimension of tumour immune biology will unlock the field in a big way. Reece A: With your focus on immunotherapies, are you planning to tackle challenges such as the high-cost associated with these therapies and limited patient response? Ron A: Healthcare costs is certainly an important area for discussion, and a very complex topic with many different factors at play. In general, I think that improving the success rate and shortening the development time of bringing new medicines to patients will contribute to decreasing the costs of medicines, and AI methods will certainly play a role there. However, this will happen at the scale of improving industry- wide productivity overall and will not occur overnight. At Noetik, we’re focused specifically on discovering and developing the next class of drugs, to get the right drugs to the right patients. Reece A: Noetik is building a precise-map of oncological traits. How are you hoping to understand what immunotherapies patients respond to? Will this be through more biomarkers? Ron A: Biomarkers are often framed as something that a translational team works on as a molecule enters the clinic to refine patient selection in the clinic. However, the lesson learned from targeted therapies is that the right definition of tumour biology should underpin the therapeutic approach at the beginning of discovery. This concept, reverse translation, is gaining traction in the industry with some companies creating reverse translation teams. We are building a map of tumour immunotypes as the basis for identifying therapeutically relevant biology that translates from discovery to clinical development. This map will provide insights for how we understand the activity of existing molecules and mechanisms. We've had a lot of interest in that area, but ultimately we are most excited about discovering and developing the best mechanism for a particular patient. DDW Volume 25 – Issue 1, Winter 2023/2024 Biography: Ron Alfa has dedicated his career to tackling the most challenging areas of unmet need in medicine. Prior to Noetik, Alfa was SVP, Head of Research (acting-CSO) at Recursion (RXRX) and an early founding employee where he led the company’s scientific and portfolio strategy from pre-Series A through IPO. He has led research programmes across rare disease, neuroscience, oncology, and immunology, that have advanced molecules from discovery to clinical development.