Putting AI at the service of the healthcare sector is a delicate task; this is even more true in oncology, where the stakes are particularly high. Biotech startup Valar Labs aims high but starts small with a tool that accurately predicts certain treatment outcomes, saving patients valuable time. It raised $22 million to expand into new cancers and therapies.
Every cancer is different, but many have established best practices perfected over years of testing. Sometimes, however, this means following a given treatment regimen for months in order to find out if it works.
Bladder cancer is one of them, Valar co-founders told TechCrunch. A common first treatment recommended by oncologists, called BCG therapy, has a chance of working a coin toss, which is actually pretty good! But wouldn’t it be nice to not have to flip that coin to begin with? This is the problem Valar is trying to solve.
CEO Anirudh Joshi said the team met at Stanford, where they were studying AI support for clinical decision-making. In other words, helping patients and doctors decide which treatment path to pursue, whether it’s two or a dozen.
“What we’ve learned is that for the majority of cancer patients today, their treatment plan is really unclear,” Joshi said. “They have options, but it’s hard to say what will work well – you just have to try things. Our idea was therefore to make an informed decision. In the treatment of bladder cancer, only one in two patients responds to standard care. If we knew which patient this was, we wouldn’t have to waste a year of therapy for something that doesn’t work. »
The first test they developed, called Vesta, focuses on this specific situation. And it’s not a theoretical software solution: the team worked with a dozen medical centers around the world to study more than 1,000 patients and find out what exactly makes them respond to certain therapies.
The process has two elements: first, a visual AI (or computer vision model) trained on thousands of histological images of cancer patients. These thin slices of affected tissue are increasingly being scanned and inspected by experts, although the process can be somewhat rough.
“This very high-resolution image tells you a lot about what is happening at the cellular level of a tumor,” explained technical director Viswesh Krishna. “We run our models on this image to extract a very large amount of features, similar to a genomic panel; we generate thousands of histological readings (i.e. important image features) and take the most important ones that pathologists can review, but can’t really quantify. They can see that they are different but cannot measure the differences between them.
Joshi was careful to add that they were not trying to replace the pathologist, but to augment them. You could think of it as a smart microscope that helps an expert take precise measurements on things like cell damage, immune response, and other structures that indicate disease progression or inhibition.
“At the end of the day, the doctor is always in charge. It’s just more data, and they like it. And offering tests like this is a fundamental external perspective, and patients really like that,” Joshi said.
The imaging component, the team noted, was trained on tons of data and is generalizable to many areas and cancers; Counting lymphocytes in breast cancer tissue is largely the same task as counting them in skin cancer tissue. But what this means, or any other quantifiable biomarker the model can identify, indicates that the likelihood of a patient responding to treatment is much more limited to specific conditions.
Accordingly, the second element of the Valar system is what actually needs to be adapted to a particular clinical situation. And to that end, the company has demonstrated that, in the specific case of bladder cancer and the standard treatment regimen, its test is a far more accurate indicator of success than any other available indicator.
Risk factors such as age, medical history, smoking, etc. are variable indicators of some treatment outcomes, but these are “very crude,” Joshi noted. Valar claims that their AI models “outperform all of these variables (in terms of predictive power) and are independent of them” – meaning they can be used in addition to the standard risk factor, not just in place of them .
They also noted that it was important that the results remained interpretable: the last thing doctors or patients need is a black box. So if it says a patient will respond well, this is supported by “because their immune system does A and their nuclei do B, etc.” »
The company, founded in 2021, devoted much of its efforts to developing the image model and its first clinical model for the aforementioned BCG treatment in bladder cancer patients. As Valar noted in a recent announcement, the test identifies individuals at triple risk of not responding to BCG, meaning (at the care team’s discretion) it’s probably best to try something else. thing. If it saves even a month of wasted effort, it could be life-changing for some.
As anyone who has experienced cancer care will tell you, not only is each day of treatment incredibly precious, but trust is hard to find. Valar may not offer certainty (almost impossible in oncology), but it could be a powerful arrow in the tremors of caregivers.
Coinciding with the impending release of its first product, Valar closed a $22 million Series A round led by DCVC and Andreessen Horowitz, with participation from Pear VC.
“The fundraiser was timed perfectly,” Joshi said. “We were able to finalize this validation, and now this funding will help fuel the commercialization of Vesta, and at the same time we begin to expand it to other types of cancer.”
The founders said they hope to expand steadily, using a commercial laboratory model, much like genomic testing has followed in recent years, COO Damir Vrabac said: “It’s very similar to these other tests that have preceded us, it does not add any friction to the health system. » We hope that this will allow them to pass the cost on to insurers and, ultimately, to completely reduce the cost of care by avoiding unnecessary and ineffective treatments.
techcrunch
