Medtech startup brings Oracle AI to bear on cancer drug research


In recent years, human ingenuity and scientific know-how has developed advanced drugs promising relief and even recovery to millions of people worldwide diagnosed and living with previously incurable forms of cancer.

Historically, identifying the active ingredient for candidate drugs is an easy process. Drug discovery drew on old knowledge of traditional folk medicine – drugs such as aspirin began in this way, or came about by happy accident, such as in the case of penicillin, the first identified antibiotic, which was spotted by accident on a mouldy petri dish in 1928.

A century after Alexander Fleming’s chance encounter with penicillium chrysogenum, modern-day drug discovery is big business, with a large and ever-growing pharmaceutical industry pouring billions of dollars into the process, often backed by governments – as was the case with the first Covid-19 vaccines – but the process is still hard and dogged by inefficiencies.

What if there was a better way? With the technology being developed and used at Imagene AI, an Israel-based startup, we may be on the verge of big change. Imagene and its founder and CEO Dean Bitan are hoping to kickstart the process of drug discovery and find new ways to help oncologists provide the best care possible by bringing artificial intelligence (AI) capabilities courtesy of Oracle to bear on the scourge of cancer.

Bitan is a lifelong technologist who earned his degree in computer science at the tender age of 15. But it was some years later when he became interested in the field of cancer research. Cancer is a disease that will touch the lives of almost everybody on the planet in some way during their lifetimes, and in Bitan’s case, he sadly lost a close relative to it.

Sitting down with Computer Weekly on the fringes of Oracle Cloud World in Las Vegas, Bitan says: “I was interested in technology and entrepreneurship and that was my field. Initially, I had no relationship to cancer.

“But when it happened, I learned a lot about the disease, and I learned about the gaps and what might we do better…There are not always enough opportunities in terms of drugs – we see many cancer types where we don’t have enough drugs to offer patients.

“So that’s the story of Imagene, I decided to get into that field and I really wanted to assist physicians to better navigate the diagnostics and treatment decisions,” says Bitan.

Is AI the answer?

Did Bitan always have an inkling that AI might offer a path forward? He says he started considering the possibility very early on in the research that Imagene grew from.

“The assumption was that we might be able to leverage technology into that area. [But] I’m an engineer. And talking with physicians, there’s a different mindset,” says Bitan.

“What we did was to sit together and think – we know what the challenge is, how can we do it better? Some physicians had told me, before we established the company, that they have a strong sense of intuition when they look at a biopsy image.

“They say they can identify patterns that are probably related to the presence of biomarkers indicating whether a patient will respond, or not, to a specific drug. They build that up over many years of practice and observing their patients, but that is something that can be enhanced…Intuition and AI go very well together. So, we understood that we would try to see if we could discover more information out of those biopsy images.”

Bitan and his team leveraged 630,000 anonymised biopsy images taken from multiple cancers at multiple sites in the body, and from there developed a foundation model to deliver what he now describes as “oncology intelligence”.

This 1.1 billion parameter foundation model is called CanvOI. At its heart, it captures the complex features and patterns in biopsy images that a human might never see unaided to enhance a researcher’s understanding of various pathological features and derive new understanding from that.

The ultimate idea is to deliver a “robust vision data backbone” for the development of downstream applications in oncology research. This doesn’t just have to apply to identifying new drugs, it can also predict how people with various different biomarkers might respond to them, eventually enabling frontline physicians to deliver bespoke cancer care based on their patients’ unique physiologies.

OCI powers CanvOI

Imagene’s model runs on Oracle Cloud Infrastructure (OCI), taking advantage of OCI AI Infrastructure and OCI Supercluster, which can scale to tens of thousands of GPUs right now for AI inference, and will be able to tick over 130,000 in the very near future. 

Bitan says that by applying Oracle’s compute capabilities and Imagene’s new approach to digital pathology foundation models, CanvOI is already achieving industry-leading performance in its various tasks, even when using minimally labelled data.

“Oracle is in a unique place to support AI companies in those challenges. So, when we talk about computing power, the fact that they had that strategic agreement with Nvidia allows us to get more availability of GPUs. And more availability of GPUs means more computing power and it’s better for us,” he says.

“We need companies like Oracle to go with us on this long journey because the challenges are real, and we want to continue and show more milestones related to this. With ChatGPT and LLMs we saw how the technology, in less than two years, moved from elementary school level to high school level, and now they are talking about PhD expert level. We want to see similar stuff in the world of oncology, and we’ve done it with biopsy images, but as we move forward, we will add more models.”

More widely, CanvOI forms the cornerstone of the firm’s new OISuite, a platform designed to support researchers and diagnostics developers and enable them to explore answers to a wide range of questions needed to conduct their research. Bitan says this alleviates the need for AI expertise and data acquisition, enabling new breakthroughs while still adhering to the highest possible standards of data privacy and security. On which note, says Bitan, all the data used by Imagene’s systems is de-identified in advance.

“And of course,” he continues, “we work based on the highest standards of GDPR, HIPAA compliance, etcetera. That much is obvious, but besides that, we are also working on zero-trust approaches, we run vulnerability scans, we encrypt data at rest and in transit, despite the fact it is de-identified.”

Future goals

Imagene is already working with medical institutions in multiple countries, including world-renowned US research facilities such as Johns Hopkins in Baltimore, and Northwestern University in Chicago, as well as leading cancer centres in Brazil and Israel. Bitan wants to go further, to bring in not only academic medical centres, but private and reference labs as well.

“In the field of cancer research, we don’t have the privilege not to do what we can because every day is important,” he says.

Looking further ahead, Bitan says he sees opportunities to apply the technology developed at Imagene to other areas of medical research as well, such as Covid-19 or HIV/AIDS.

“We will go towards those areas as we move forward. You will see more and more models that aggregate different modalities – so not only biopsy images, we could maybe add radiology, MRIs and X-rays. Or microbiomes or maybe even genome sequencing. Then we will be able to answer much more complex questions related to different aspects of healthcare,” he concludes hopefully.



Source link