Could AI Unlock Zero-Cost Gene Therapy Delivery?
Dr. Eric Kelsic explains how AI could reduce the cost of gene therapy delivery to $0.
Complete the form below to unlock access to ALL audio articles.
More than three decades after the first human gene therapy clinical trial, these therapies are starting to deliver on their promise to treat complex genetic disorders. Over 40 cell and gene therapies are now approved for human use by the US Food and Drug Administration (FDA). 2023 marked a blockbuster year for the field with 7 new approvals, including novel CRISPR-based gene therapy CASGEVY®.
While this new class of medical treatment offers hope for patients lacking alternative options, there is a barrier to its accessibility: cost. A 2023 study calculated the average gene therapy cost at $43,110 per quality-adjusted life year. It's a pricepoint that poses challenges even for high-income countries and makes widespread access in low- to middle-income countries nearly unattainable.
Implementing strategies to reduce the costs associated with gene therapies is therefore paramount to support wider groups of patients. Eric Kelsic, co-founder and CEO of Dyno Therapeutics, believes that artificial intelligence (AI) can help. Kelsic says that AI could effectively reduce the cost of gene therapy delivery to $0. In this interview with Technology Networks, he explains why.
Can you explain why cost-effective gene therapy development is increasingly important?
Eric Kelsic, PhD (EK):
As research advances, gene therapies are becoming safer, more effective and applicable for even more diseases. However, today, gene therapies are still some of the most expensive medicines on the market, often priced in the range of several million dollars per treatment.
As these gene therapies become increasingly useful, we need to make them even more affordable and accessible to enable millions to billions of patients to benefit.
MC:
MC: Can you discuss the current “gold standard” approaches to delivering gene therapies, where they may fall short and why?
EK:
MC:
EK:
Optimizing AAV capsids is a design problem with huge possibility, which makes it an ideal application for AI. However, even a single amino acid change on the capsid protein sequence can completely change its therapeutic properties, and most changes break function, which makes it almost impossible for humans to approach rationally.
That’s why Dyno is applying AI to solve the in vivo delivery challenge, enabling drug developers to deliver genes to every organ, every cell and every patient. We’ve developed technology that enables us to generate large and detailed datasets, measuring hundreds of thousands of different sequences and their in vivo delivery properties. AI is exceptional for automatically identifying patterns within this data and applying these insights to generate new sequences with higher performance.
To expand on how we got here, in a 2019 Science paper we measured protein fitness landscapes comprehensively and at an unprecedented scale, tracking large libraries of AAV capsid variants using DNA barcoding and NGS as they biodistributed and delivered genetic payloads in mice. A short time later, in Nature Biotechnology we published a pioneering application of machine learning to explore deep sequence space with the goal of protein diversification, towards creating synthetic AAV capsids that can help every patient by avoiding pre-existing immunity to natural AAVs.
The exceptional depth and complexity of the data set also enabled us to show that deep neural networks can be more efficient than a random directed evolution approach, as an early demonstration of the power of AI for sequence design. Today, Dyno’s LEAPSM technology, as shared at conferences and in preprints over the past few years, combines these approaches with more recent architectural machine learning innovations and applies them to measuring and optimizing the delivery properties of capsids in non-human primates to enable more effective translation to human patients.
MC:
You believe that, ultimately, the cost of delivering gene therapies could be brought down to $0. Can you expand on this point?
EK:
The high price of gene delivery today is due to many factors, including delivery challenges, high manufacturing and research costs, and long clinical development timelines. Each of these factors compounds costs–but this is also why massive cost reductions are possible.
There is precedent for this in other industries; for example, the cost of computing has dropped exponentially over the last five decades. I believe that reducing the costs of gene delivery is similarly important, and will be similarly impactful, making gene therapy affordable and accessible for billions of patients.
I believe that the biggest near-term opportunity could come from improving the delivery efficiency per capsid. We’ve demonstrated 100x improvements for brain delivery and I believe improvements of more than 1000x are possible. In the long term, machine learning approaches can also be applied to optimize genetic sequences for improved manufacturability, and make it easier to store, distribute and administer gene therapies, all of which enable further cost savings all along the supply chain. Ultimately the cost of the raw materials (i.e., carbon, nitrogen, oxygen etc.,) that go into a gene therapy is so low that $0 is a reasonable approximation of what we can achieve with the right optimizations.
MC:
Are there any barriers to the use of AI in this context? If so, how do you envision they could be overcome?
EK:
It’s well known that AI is only as good as the data it’s trained on. Until we started at Dyno, there wasn’t much high-quality data connecting AAV sequences to their in vivo functions. To solve this, Dyno created new multiplexing technologies and a computational platform for training our machine learning models at scale.
Now, in an average month, we make billions of distinct in vivo measurements, which enables us to build better models, and design even better capsids. With a better capsid we can make even better measurements, which further improves the models, and so on.
I’ve been working on this for 10 years and I’m very excited about all the progress Dyno has made to date applying AI to protein sequence design. It feels like this is just the beginning of the transformative patient impact these technologies will bring to genetic medicine.
