Skip to main content
Premium Trial:

Request an Annual Quote

Valar Labs Using $4M in Seed Financing to Advance AI Tool for Personalizing Cancer Treatment

Premium

NEW YORK – Valar Labs, a Palo Alto-based startup founded last year, is using $4 million in seed financing to jumpstart its mission to build a clinical-grade, deep learning system that analyzes microscope images from histology slides to inform therapeutic decisions for cancer patients. 

Ideally, every medical oncologist would like to take their best shot at helping patients live cancer free with the first treatment. Although the goal of precision medicine is to identify the best treatments for patients from when they are first diagnosed with cancer, molecularly informed therapies are often not offered to patients until they've developed advanced disease and undergone rounds of other drugs. Newly diagnosed patients also may not harbor therapeutically actionable biomarkers. This means that patients often end up receiving personalized treatments when they're too sick to benefit from them. 

With the seed financing it received from lead investor Andreessen Horowitz (a16z) and a lengthy roster of notable genomics tools and diagnostics executives including Dave King, former CEO of Laboratory Corporation of America, and Peer Schatz, former CEO of Qiagen, Valar Labs wants to develop a tool that can identify personalized treatment approaches for the majority of patients who don't harbor targetable molecular biomarkers. 

The firm was cofounded by Anirudh Joshi, Damir Vrabac, Viswesh Krishna, and Pranav Rajpurkar, who connected as Stanford University graduate students working on research linking AI and medicine. There, they realized that machine learning had yet to achieve its full potential in the field of medicine, particularly pathology. 

Until recently, digitized pathology slide availability has not been widespread enough to support development of artificial intelligence-based diagnostic platforms. However, the COVID-19 pandemic prompted a surge in digitization of pathology slides in order to accommodate remote examinations. Valar Labs' founders saw an opportunity in that shift to develop AI tools based on histology data that is already being collected for every cancer patient. 

"Anecdotally, at Stanford we have seen a 4X scale-up in the number of slides scanned since the pandemic," said Joshi, Valar Labs' CEO. "Academic medical centers near universally have pathology scanners at this point, and the main distinction is whether it's used for scanning research slides or diagnostic cases. Community [cancer] centers, based on recent conversations, have also purchased scanners." 

Other companies, such as PathAI, Paige, and Tempus, are also taking advantage of this trend by developing machine-learning applications in pathology. For example, Paige's suite of tools uses data from millions of digitized slides connected to pathology reports to help pathologists recognize patterns in tissue samples. However, Valar is taking a step beyond simply assisting or expediting the work of pathologists. 

"What I liked about what Valar is doing is they're saying, 'We're not trying to make the pathologist a little bit faster. What we're trying to do is just create a diagnostic tool that is totally separate from what the pathologist could do,'" said Vineeta Agarwala, a general partner at a16z and former director of product management at real-world data firm Flatiron Health, which is owned by Roche. Agarwala and Jay Rughani, another partner at a16z and former Flatiron Health employee, are advising Valar Labs. 

"We spent about six or seven months talking to hundreds of physicians, pathologists, oncologists, radiation oncologists, and surgical oncologists to understand what is top of mind for them and what was of highest need," said Joshi. 

What those physicians said they wanted was a tool to inform treatment decisions when guidelines, such as those issued by the National Comprehensive Cancer Network, are too ambiguous to make a definitive therapy choice for a patient. Those guidelines are based on the outcomes patients experienced in clinical trials. However, those trials involve patients who match detailed enrollment criteria and treatment settings. Cancer care in the real world is not so neatly segmented, making it harder for oncologists to determine the best course of care for patients who don't match the specific criteria used in these trials. 

"After analyzing a lot of data ourselves, we realized that a lot of the patients that were put under one treatment would actually have benefited from another treatment," said Vrabac, chief operating officer at Valar​​​​​​​. "But there was no tool for the oncologists today to actually select the right treatment unless they had an actionable genomic mutation." 

Valar's machine-learning algorithms don't rely on genomic testing, which isn't fully capturing tumor biology critical for predicting drug response, in Joshi's view. "The answer to predicting drug response has to be in the tumor biology," said Joshi. "That's the reason why different patients are responding differently to the drugs." 

Using AI algorithms that correlate information from pathology slide images and patients' outcomes, the firm has created a histology-based biomarker that can predict which treatment has the best chance of benefiting cancer patients. The test would be conducted by sending digitized images of the patient's histology slides to Valar for analysis by its algorithm. 

Joshi said that Valar's biomarkers will help guide therapies that are already standard of care, but don't currently have associated genomic markers to identify who will respond best. "For example, in some cancers there are three first-line, standard-of-care therapies, with no way for an oncologist to select the best one upfront," said Joshi. "We are providing the platform to select the best one." Currently, oncologists in that situation would have to guess. 

Valar's tool will be generalizable across different types of cancers, with pancreatic cancer as an initial area of focus. The tool will also be designed to address all stages of disease, but the company wants to demonstrate its impact particularly in the earliest stages of therapeutic planning. Because it's early days in the development of this tool, Joshi couldn't provide specific cost-effectiveness estimates, but he expects the test to be attractive to pathologists and insurers by helping avoid wasteful spending from ineffective treatments. 

A successful proof of concept 

The idea of using automated analysis to identify prognostically relevant features of tumor cells overlooked by human pathologists is not new. In 2011, a group from Harvard created an "automated pathologist" by using image processing software and programming it to identify features in breast cancer tissue that could predict outcomes. Then, in 2019, German researchers used deep learning to predict microsatellite instability from histology in gastrointestinal cancers. 

Building on this prior research, the Valar team published a proof-of-concept study in Scientific Data that illustrates the use of geometric features from a diffuse large B-cell lymphoma pathology image set to predict survival outcomes for patients. 

The NCCN international prognostic index classifies DLBCL using patients' age, lactate dehydrogenase level, extra-nodal sites of involvement, Ann Arbor stage, and ECOG performance status. That model assigns patients to good, intermediate, and poor risk categories, but it doesn't help physicians make therapeutic decisions. The standard treatment for DLBCL is rituximab, cyclophosphamide, doxorubicin, vincristine, and prednisone (R-CHOP). 

That treatment regimen is effective for a majority of DLBCL patients but not all. One recent analysis found that 78 percent of DLBCL patients treated with R-CHOP were free of cancer progression at two years. 

Some existing methods for predicting which patients will benefit from R-CHOP include gene expression profiling to determine cell of origin (COO subtyping) and testing for MYC and BCL2 gene translocations associated with more aggressive disease and poor outcomes. Targeted therapies are not yet available for patients with those genetic variations. The success of COO subtyping suggests that morphologic properties of the tumor itself could be used to predict outcomes, but DLBCL cells vary enough to make this a very challenging task. 

Researchers led by Vrabac tackled this problem by developing an automated method for analyzing the DLBCL-Morph image set based on features of the tumor cells. For their analysis, Vrabac and colleagues used a publicly available set of 42 digital, high-resolution images derived from 209 patients with DLBCL at Stanford Hospital. The resulting model predicted survival superior to random chance based on geometric features of the cell alone, and its performance was even better when patients' clinical features were incorporated. The proof-of-concept algorithm was not used to make treatment predictions, which is the company's focus now. 

Machine learning-based tools have a reputation for being black boxes in that doctors and patients have little insight into the data used to yield the reported result. But the company hopes to go beyond identifying the treatment opportunities for patients. It aims to create a system that doesn't simply output a patient prediction, but also provides information on the features used to make those predictions, which can in turn be used to generate new hypotheses about drug targets and therapeutic response. While Valar's AI algorithms are still being trained on images of tumor cells, the company believes the tool can be developed to work with a variety of imaging data such as CT scans. 

Toby Cornish, vice chair of pathology informatics at the University of Colorado School of Medicine, said Valar's concept is "entirely feasible." He said that outside of the research setting, a very small number of laboratories are signing off clinical cases based purely on digital images, but Valar's approach does not necessarily require upfront digitization of slides. Instead, Cornish said, the slides could be digitized on demand if the laboratory has that capability, or sent to an outside reference laboratory for digitization. That would fit in with existing pathology workflows, in which it is already common for slides and tissue blocks to be sent out for additional testing. 

In Cornish's view, Valar's ability to successfully commercialize this tool will likely hinge on its ability to form partnerships with hospitals, health systems, and laboratories that can provide the clinically annotated specimens it needs to train and validate its machine- learning algorithm. "If they have good partnerships, I think that this is quite doable," Cornish said. 

According to Joshi, now that Valar has secured its seed financing, its next task will be running large-scale validation studies with multiple cancer centers to ensure accuracy of its tool across diverse data sets. "The biggest thing that everyone wants to make sure of is that markers are broadly generalizable, not just in a couple of hospitals in California, but validated across the country, and across various demographics," he said.