Skip to main content
Premium Trial:

Request an Annual Quote

Ataraxis AI Aims to Address Gaps in Molecular Testing with AI Digital Pathology Tests

Premium

NEW YORK – Ataraxis AI is developing artificial intelligence-driven digital pathology tests that it believes may provide faster, more accurate, and more cost-effective answers to clinical questions about cancer patients' outcomes and recurrences than traditional molecular tests on the market.

Last month, Ataraxis introduced its first clinical test, Ataraxis Breast, which it developed using the Kestrel pan-cancer foundation model. The company is engaging with health regulators and hopes to make the test available for clinical use in early 2025.

Kestrel is a vision transformer model trained on approximately 400 million digital pathology slide image patches sourced from a set of whole-slide images representing nearly 4,700 cancer patients across 10 cohorts in six countries. The firm validated the model using data from more than 3,500 patients across five cohorts, which were not used to train the model, including patients from the Cancer Genome Atlas; Providence Health, which serves patients in Washington and several other US states; University of Chicago Medicine; Karmanos Cancer Institute in Michigan; and Cancer Center Baselland in Switzerland. Many types of cancer, including every subtype of breast cancer, are represented in the training and evaluation datasets, which came from about 8,000 patients from seven countries.

Ataraxis has published a validation study of Ataraxis Breast as a preprint in ArXiv and is currently submitting the paper to a peer-reviewed journal.

Ataraxis CEO Jan Witowski and Chief Scientist Krzysztof Geras founded the New York-based company in 2023 as a spinout from New York University after raising $4 million in a seed financing round co-led by Giant Ventures and Obvious Ventures. Both scientists have roots at NYU Langone Health, where Witowski led AI and medical imaging research and Geras is an assistant professor in the radiology department.

The model underlying Ataraxis Breast integrates pathological features extracted by Kestrel with routine clinical data, such as cancer staging; patient age; estrogen receptor, progesterone receptor, and HER2 status; and the presence of ductal or lobular histology. Based on this, the test produces a cancer recurrence risk score between 0 and 1.

Kestrel was developed based on foundational research and methods developed by Yann LeCun, a professor of computer engineering at New York University and chief AI scientist at Meta. LeCun is a strategic adviser at Ataraxis AI and has been involved in developing its models, but Meta is neither an investor in nor a technology development partner to Ataraxis.

Kestrel is a foundation model, meaning that it is trained on multiple large, diverse datasets so it can be applied to a range of different use cases. In addition to the newly unveiled breast test, Ataraxis is eyeing development of additional tests in solid tumors using the same model, rather than training and validating separate models for each cancer application.

Ataraxis' hypothesis is that if the model is trained not just on breast cancer slides but also slides from all types of cancer, it will learn not just about breast cancer biology but more broadly about tumor biology and histopathology. "The features it extracts are no longer just tied to breast cancer, but any cancer," Witowski said, adding that once a foundation model is trained, new tests can be built relatively quickly.

The other innovative technology incorporated into Kestrel is a vision transformer algorithm that was trained using a self-supervised learning method, DINOv2, borrowed from Meta. A transformer is a type of deep-learning algorithm originally developed by Google that has been widely used in large language models, computer vision, and generative pre-trained transformers (GPTs). Kestrel uses the vision transformer algorithm to break the cancer tissue images into patches and then teaches itself to spot the significant features in the image, essentially providing its own labels. This is in contrast to a supervised deep-learning algorithm, which has to be labeled by a human or an unsupervised deep-learning algorithm like ChatGPT that doesn't use labels at all.

"It's almost impossible for humans to actually label these millions of [images]," said Tinglong Dai, a professor at the Johns Hopkins Carey Business School and co-chair of the Johns Hopkins Workgroup on AI and Healthcare, who was not involved in the development of Ataraxis Breast. He pointed out that it would also be prohibitively expensive for Ataraxis to have human pathologists label all 400 million image patches in their training sets. 

"For humans, if you don't see the big picture, it's very hard to make sense of [small pieces]," he said. "Imagine you buy a puzzle from the store, and open the box, and just look at each individual piece. As humans, we cannot make sense of that."

With its AI-driven digital pathology tests, Ataraxis aims to overcome the limitations of molecular diagnostic tests, which have become part of the standard of care for guiding cancer treatment. According to Witowski, those tests consume tissue samples and often leave unanswered questions about a patient's likelihood of responding to a recommended therapy.

Witowski said a molecular test can determine whether a patient is eligible for a therapy based on biomarkers they harbor, but that test can't definitively say whether that patient will ultimately respond. "We can finally start answering questions molecular diagnostics has often ended up not answering," Witowski said, pointing out that many patients who test positive for a genetic marker and qualify for a corresponding biomarker-driven therapy still do not respond to that therapy. Ataraxis believes that its technology may eventually provide a more complete picture as to whether a patient will respond to a therapy, which in turn may lead to more accurate treatment decisions.

Ataraxis doesn't intend to replace molecular diagnostics that are gauging known treatment-predictive biomarkers or that are available as companion diagnostics to drugs. However, for patients who do not qualify for those types of tests, Ataraxis' AI tests may be an alternative.

Ataraxis assessed the accuracy of its breast cancer test in five patient cohorts in the evaluation sets and found its risk scores to be prognostic for the primary endpoint of disease-free interval and for secondary endpoints including distant disease-free survival, overall survival, and recurrence-free survival. The company did not test its ability to predict whether a patient is likely to benefit from a specific targeted or immunotherapy treatment and did not compare it to next-gen sequencing-based profiling tests.

Ataraxis, however, compared risk score predictions from its breast test to the Exact Sciences' Oncotype DX Breast Recurrence Score, using hematoxylin and eosin-stained slides from the same tissue block that was previously used for Oncotype DX testing. Oncotype DX is a 21-gene RT-PCR-based assay that estimates distant cancer recurrence risk following endocrine therapy and predicts the benefit of adding chemotherapy to endocrine therapy. The researchers found that Ataraxis Breast had a 30 percent greater accuracy for predicting cancer recurrence than Oncotype DX and was able to reclassify patients rated as having intermediate risk by Oncotype DX into low- or high-risk groups.

Witowski attributed Ataraxis Breast's superior performance over Oncotype DX in this retrospective study to the fact that Oncotype DX was developed over 20 years ago with data from about 700 patients and logistical regression modeling methodology available at the time. Ataraxis Breast, in comparison, was trained on over 8,000 patients using more advanced modeling methods. "It helps a lot for the model to have seen an order of magnitude more patients than Oncotype DX, and … the self-supervised learning approach with a pan-cancer model and deep learning on top of it allow us to use new data with a strong signal in a very effective way."

According to Exact Sciences' website for Oncotype DX, the turnaround time for the test is seven to 10 days, and studies have cited list prices above $3,000. Witowski highlighted that Ataraxis Breast can return results within one business day, and he believes the firm can do better on price, too, although it hasn't set a price for its test yet. "Ultimately, pricing will reflect both the efficiency of our approach and also the expanded value we bring to patients and providers as we will provide insights into multiple clinical questions, going beyond Oncotype's capabilities," he said.

Rick Baehner, chief medical officer for precision oncology at Exact Sciences, pointed out that Ataraxis Breast has not yet been validated prospectively in a clinical trial, whereas Oncotype DX has been prospectively validated as a test that can provide prognostic information and predict whether a patient will benefit from chemotherapy in multiple independent clinical studies. In addition, the test is included in all oncology guidelines.

Furthermore, Baehner expects that NGS testing will remain a cornerstone of precision oncology. "In the US, NGS is the gold standard for identifying genomic mutations in patients' tumor samples," he said. "Widely available and foundational to clinical trial data supporting FDA-approved targeted therapies, NGS is also recommended in global oncology guidelines."

However, he acknowledged that in areas with limited NGS testing, H&E slide-based tests could help identify patients likely to harbor tumor mutations, and follow-up NGS testing could then confirm gene mutations to enable access to biomarker-driven therapies.

Dai agreed that Ataraxis' claims will require confirmation in prospective clinical trials and real-world settings. "It's fair to say [the technology] is promising," Dai said. However, "when you look at the real-world performance of these AI models, the performance difference [compared to retrospective studies] can be huge."

Dai also cautioned that models like Kestrel can be prone to overfitting. "Even when we have 8,000 patients coming from seven countries, actually, this is [still] relatively small," Dai said. "You want your AI model to learn as much as possible from this small dataset, but you always have the danger of overfitting, where you're learning too many things that do not generalize."

As it prepares to launch the breast cancer test, Ataraxis is laying the groundwork to develop further tests including diagnostic assays to predict the risk of cancer recurrence and outcomes for patients with a range of cancer types, particularly solid tumor settings where there is a wealth of biopsy or surgical tissue samples available for analysis. In addition to partnerships with NYU and the other institutions involved in developing Ataraxis Breast, Witowski said the company will consider opportunities for additional test development and commercialization partnerships. It is already developing a new foundation model, the next iteration of Kestrel, which Witowski said will involve a larger dataset and greater computational resources to power future tests.