AWS for Industries

Executive Conversations with Nosis Bio and Vevo: Using Generative AI on AWS to Shift the Pharmaceutical R&D Paradigm

Today, organizations are looking to harness the power of generative AI, including foundation models (FMs) and large language models (LLMs) to innovate faster to reinvent patient outcomes. While this has created a lot of hype, the adoption of these technologies in real-world drug development is still a challenge for most organizations.

Amrita Sarkar, Principal Healthcare and Life Sciences Business Development Manager at AWS speaks to two startups of AWS Generative AI Accelerator’s first cohort, Vevo and Nosis Bio. These startups share how they are successfully applying generative AI to accelerate their research initiatives, and candidly talk about their learnings as they try to separate the hype from the hope.

This Executive Conversation is part of a series of discussions held with leaders who are pushing the frontiers of the healthcare and life sciences industry with cloud technology. We were joined by Nosis cofounders Jim Martineau and Seth Myers, and Vevo cofounders Nima Alidoust and Johnny Yu.


Amrita Sarkar: To get us started, can you give a brief overview of your companies’ missions and the problems you are trying to solve?

Nosis: We are taking a revolutionary approach to targeted drug delivery. Despite the rapid advancement of new methods for treating disease, many potential life saving treatments remain undeveloped, either because they cannot appropriately reach their target tissue, or because they go too broadly in the body, causing toxicity and other side-effects. At Nosis, we use state-of-the-art deep learning, generative AI, and biochemistry to design a new class of molecules, called Targeted Delivery Vehicles (TDVs), that localize treatments to the source of disease and minimize unwanted interactions in healthy tissue. Our TDVs have been rigorously validated to deliver medicines to multiple, traditionally undruggable, tissues at the source of disease.

Vevo: At Vevo, we are combining data from organoid and in vivo disease models with AI to find better drugs and targets. Gene function is context dependent. The same gene can have vastly different functions based on how other genes are expressed, which in turn is  dependent on the broader context (patient, organ, or disease model). The failures in today’s drug discovery stem from the fact that it tries to change the disease-relevant function of a gene by finding a drug that binds to a single protein target, without having access to this context. The outcome is drugs that bind to the target but fail to treat the disease. We are solving this fundamental problem. Our Mosaic platform can measure how a drug candidate alters a gene’s function in living organisms, across 10s-100s of patients, at single-cell resolution, all in one single experiment—with unprecedented throughput and precision. We are using this throughput and precision to build the world’s first atlas of how genes function in disease-relevant contexts, to discover new drugs and targets. In parallel, we are using this data to train FMs that can capture the context-dependency of gene function, helping us find better ways to alter it in order to treat disease.

AS: Fascinating. Could you describe how you are using generative AI to advance your research?

Nosis: We are using deep learning and generative AI to build a navigation system for reaching the right tissues in the body. To deliver a medicine, our TDVs must interact strongly with the target tissue, avoid interacting with any other tissue, resist the various enzymes in the body trying to degrade them, all the while not triggering an immune system response. We have a number of deep learning models that predict these various properties for each TDV, and our built-from-scratch generative AI algorithms design TDVs that maximize these predictions.

Vevo: With a large enough amount of context-aware data, which we can generate over the next year, the FMs we are training will be able to better represent genes and the context dependency of their function i.e., how cell type, disease type, patient’s background, and perturbation by a drug affect their function. We believe such a model will have as much, if not a bigger, impact in biology than it did in natural language. It can be applied to learn new, hidden layers in how genes interact with each other, to discover new disease targets, and also to guide generative design of new drug candidates, aided by a better representation of biology.

AS: Why was generative AI the right tool for you?

Nosis: TDVs are designed to have multiple orthogonal properties that govern their delivery, efficacy, and safety. In the same way that generative AI is needed to arrange many pixels into coherent images or many words into natural language text, so too is generative AI needed to design molecular interactions into a coherent and tangible TDV. Other methods, including traditional machine learning methods, are simply incapable of this. We have been working with this technology for several years and are seeing the beginning of a massive paradigm shift in how we bring safe and effective drugs to patients.

Vevo: FMs are the right tools for us because of the type of data we are generating, and its scale. The sequence-based data we generate is able to capture function in a diverse set of contexts. We’ve seen FMs are good at learning context-dependency for other kinds of data, like understanding meaning and function of words in spoken language. So, in-spite of the differences between language and biology, we are convinced it is the right tools for the biological systems we are studying. However, FMs need a lot of data. We are working to create that. In each batch of experiment, we can generate 1M single-cell data-points. Within 2 years, we will have more data than the entirety of publicly available single-cell data-points (~30M).

AS: How does being a member of the AWS Generative AI Accelerator help you innovate faster?

Nosis: The AWS Generative AI Accelerator has been hugely enabling—being surrounded by other entrepreneurs thinking about generative AI has been a fantastic experience. In terms of technology, AWS provided us with significant compute resources and a range of AI/ML tools, alongside cost-effective cloud infrastructure that allowed us to scale. In addition, AWS is connecting us with its network of life sciences and pharma companies, so we are very excited about the commercial progress happening there.

Vevo: Over time and as our models get bigger, the cost of training will also get higher. One way to keep that in check is incremental improvements on the algorithms and implementation side. We are working with a solid team at AWS—not just cloud experts, but bioinformatics experts—to optimize our infrastructure to solve our operational and cloud infrastructure-related bottlenecks. This is critical for startups like ours. It helps us make our models more efficient from the get-go, iterate faster on the models, and train even bigger models in the next iterations. That’s the part we are most excited about!

AS: What are the biggest differences in the utility and applicability of generative AI in pharma R&D when compared to the consumer space, or even the direct-to-consumer space in healthcare?

Nosis: The pace of experimentation with generative AI in the consumer space is rapid, and that results in some failures. But similar failures in pharma R&D can have massive adverse impacts on patients. Hence the skepticism around its applicability. At Nosis, we work extremely hard to define rigorous validation standards to demonstrate the safety and efficacy of our approach at levels as high as, if not higher, than traditional approaches to drug discovery. We see Generative AI playing an incredible role in unlocking biology, accelerating drug discovery and development, but we believe that the same rigorous standards we use for clinical trials today are necessary to protect patients.

Vevo: A lot of generative AI applications in the consumer space are creating new use cases and applications by mimicking what we, as a collective, already know and understand, like, human language. However, in biology, many of the causative elements of diseases are unknown. Over time, we hope that these models can actually teach us something completely new that can potentially save people’s lives.

AS: This space is moving extremely fast. What in your oinion are the biggest areas of promise for generative AI in pharma R&D? And what will be the drivers that make it possible?

Nosis: Generative AI will permeate every single activity of pharma R&D to make it more efficient over time. From indication planning, target selection, target validation, to drug development—Generative AI is a true process accelerator. The tight integration among disciplines will be a critical driver. Data scientists need to sit with biologists and chemists, and work together on individual experimental plans. The results need to then be validated in a biological model quickly. Today, the data science team in pharma companies and scientists in the wet lab are still somewhat siloed. So the handoff between in silico to the physical world is time-consuming and error prone. This organization structure needs to change to keep up with emerging technologies.

Vevo: We believe that the biggest impact will be in building in silico models capable of better representing the biological and chemical space and using them to guide drug discovery and design. To give you an example: there have been a lot of advances in generative chemical design over the last 5-10 years. The models have gotten better at generating new chemicals that are drug-like, synthesizable, and valid; where they fail is that the drug structures they generate are still not guided by the right biology. They output drugs that are optimized to bind to a protein target, but not necessarily treat the disease. We need more of the latter; diverse sets of precise data from more sophisticated disease models (in vivo or organoid models) coupled with advanced in LLMs will bridge that gap. It helps that sequencing costs and advances in single-cell technologies are also going in the right direction.

AS: With all the noise around generative AI today, how do you separate the hype from the hope?

Nosis: We validate everything in biological systems to ensure the translatability of our platform. It’s this strong core of rigor in our work that gives us a very clear indication of whether something is working or not.

Vevo: It’s important to choose your models wisely. For certain applications, fine-tuning existing models may work. But, for highly specialized applications like small molecule drug discovery, they will not, since they have not been trained on the right kind of data. You’ll have to build your own models there. We are hopeful that once we start having scalable multimodal data from the right disease models, the adoption of generative AI in life sciences will become the standard.

AS: What do you wish you knew when you started? And your advice to other companies getting started on this wave?

Nosis: Innovate with humility, focus on the patient, and be agile to make the right adjustments as technology changes. If there’s better technology that’s not yours that can solve the problem, adopt it. Keep a lookout for all ideas but stay focused on the problem that you’re solving. Generative AI is incredibly powerful, and the companies that stay agile and humble are the ones who will win.

Vevo: The recent wave has increased our conviction in our science. While the promise is massive, a lot depends on adapting the right models, refining them for the data you are using to train them, and most importantly, having the right data to train them.

AS: This is the most exciting time to be working in therapeutic development, and this new wave of technology will enable cures for dozens of diseases that we previously thought were incurable. At AWS, we will continue to democratize generative AI for companies of all sizes and deliver new services to make it easier and more cost effective for organizations to access and securely customize the right models while protecting their IP and data. Thank you so much for joining us today.

Amrita Sarkar

Amrita Sarkar

Amrita Sarkar, Ph.D. is a Principal on the Healthcare & Life Sciences Startups team at AWS, where she works with the founders and investors behind the world’s most innovative HCLS startups to help them build at scale, resolve technical/regulatory challenges and win more business. Prior to joining AWS four years ago, she was a Venture Capital investor based in Paris for several years, investing in the Life Sciences among other sectors. She has mentored hundreds of startups across the United States and Europe, including for Techstars. She holds a BA in Mathematics and Molecular Biology from Middlebury College, a PhD in Computational Biology from NYU/Mount Sinai School of Medicine and an MBA from the Collège des Ingénieurs in Paris. In 2018, Forbes named her one of the top tech disrupters in France. She currently lives in New York.

Jim Martineau

Jim Martineau

Jim Martineau, MS, MBA is the Cofounder / CEO of Nosis Bio. Jim co-founded Nosis after working for more than 15 years at the intersection of data science and healthcare innovation including a focus in precision medicine, genomics, and drug development. Prior to co-founding Nosis, Jim was Head of Oncology Products at Natera (NASDAQ: NTRA), a leading molecular diagnostics provider, and Director of Product & Operations at Flatiron Health (acquired by Roche in 2018 for $2 billion). Jim received his MBA in Health Care Management from The Wharton School at the University of Pennsylvania. He earned his MS in Computational & Mathematical Engineering from Stanford University and his BS in Engineering Physics from the Colorado School of Mines.

Johnny Yu

Johnny Yu

Johnny Yu, Ph.D. is the CSO and co-founder of Vevo Therapeutics. His focus is on single-cell methods, oncology targets, and high-throughput functional genomics in the field of small molecule drug discovery. He received his PhD from University of California, San Francisco in Biomedical Sciences. At UCSF, Johnny developed single-cell sequencing methods focused on application to drug discovery, in-vivo oncology models, and lineage tracing, under the supervision of Vevo's other co-founders, Hani Goodrzi and Kevan Shokat. Prior to UCSF, Johnny worked in target validation efforts at the Broad Institute, and upstream and downstream process development at Takeda (Millenium) and Biogen.

Nima Alidoust

Nima Alidoust

Nima Alidoust, Ph.D. is the co-founder and CEO of Vevo Therapeutics. Over the last decade, Nima has been leading engineering, scientific and business teams at the intersection of computational chemistry, advanced compute, and AI. Prior to founding Vevo, Nima was the CEO of the Computational Chemistry Division at 1QBit (now spun out as Good Chemistry Company) and VP of Product and Strategy at Rigetti Computing ($RGTI). He also spent 2 years at McKinsey & Co, where he primarily served clients on strategy, corporate finance, and innovation topics. Nima received his PhD from Princeton University, focusing on computational and quantum chemistry. He received his BA from Middlebury College in Physics and Classics.

Seth Myers

Seth Myers

Dr. Seth Myers is the Cofounder / Chief Technology Officer of Nosis Bio. Seth co-founded Nosis after working at the cutting edge of deep learning, graph neural networks, and natural language processing for over a decade. Prior to co-founding Nosis, Seth worked as the Vice President of Product & Head of Data Science at DemandBase, a late stage venture-backed technology firm in San Francisco, CA and was the Chief Scientist for Spiderbook, a machine learning startup focused on organizing the world’s information on how companies collaborate. Seth is a prolific researcher whose published work on machine learning has been cited over 6,000 times. Seth received his PhD in Computational Mathematics from Stanford University. Seth earned his BA from Northwestern University, where he triple majored in Math, Physics, and Integrated Science.