British omics startup, Panakeia, said it has validated its artificial intelligence-based technology, which it is hoping will one day bypass conventional wet lab tests to identify biomarkers from images of biopsy slides.
Panakeia has evaluated the performance of PANProfiler Colorectal using imaging and clinical data from colorectal cancer patients treated at the University of Leeds, Queens University Belfast, and the University of St Andrews. The results, published in npj Digital Medicine in December, showed that the platform was able to identify which patients had microsatellite instability (MSI) and mismatch repair deficiencies (dMMR) in line with molecular profiling assays 94 percent of the time.
“That’s a very, very high percentage of agreement, which gives the clinicians confidence in terms of accuracy,” said Panakeia CEO and Founder Pahini Pandya.
The technology has been made available in select National Health Services locations in a research context, which the company said will “pave the way for broader clinical adoption and commissioning in the NHS.” Panakeia is also currently participating in the UK’s Medicines and Healthcare products Regulatory Agency (MHRA) AI Airlock program through next month, which is not a route to market but is intended to help companies and the agency develop best practices for regulating AI in the healthcare industry.
Nic Orsi, a histopathology and industrial innovation professor at St James’s Teaching Hospital in Leeds, described how his team performed the blinded analysis using 3,576 images of tissue samples from 1,243 colorectal cancer patients and found that results from Panakeia’s platform were largely consistent with standard lab-based biomarker testing. The platform identified 92.5 percent of cases where the biomarker was present and 94 percent in which it was absent, accuracy that Orsi said is in line with immunohistochemistry tests and molecular assays.
“These findings highlight the potential for AI tools like this to support more timely and accessible molecular testing, which is increasingly critical for guiding treatment decisions in colorectal cancer,” Orsi said.
Not only is the technology, known as PANProfiler, accurate, but it’s also fast, Pandya said. PANProfiler can be used right after a pathologist looks at a slide and returns results within minutes. Molecular tumor profiling reports through commercial labs can take weeks to come back to physicians. The software could expand rapid diagnosis and speed up treatment decisions, Pandya said, especially in community and resource-strained settings that often don’t have the infrastructure or capabilities to do the lab tests in house. “This is where we fit in,” she said.
Panakeia tested PANProfiler and demonstrated proof of concept across over 30 cancer indications in a 2024 Communications Medicine study. Other than the colorectal cancer version, the firm is also advancing a breast cancer-specific PANProfiler. The colorectal cancer-specific product is Panakeia’s second to carry a mark indicating compliance with certain technical safety and health requirements in the UK, after garnering both the UK’s Conformity Assessed mark and the EU’s CE mark for a breast cancer-specific PANProfiler.
AI is known to hallucinate, or confidently give false or inaccurate information, when a large language or visual model can’t find an exact match to a query. Panakeia’s technology is intentionally trained to generate a confidence score, and when the score is below a certain threshold, the software will return an “inconclusive” result. That happened about 14 percent of the time in the colorectal cancer study.
Pandya said it’s important to have these guardrails that can indicate when alternative, more traditional molecular tests should be done. “AI always gives results, but when you design it as a part of a broader product, you can always put in ways whereby AI won’t return results,” she said.
Panakeia’s team is still figuring out if there are commonalities among samples with an “inconclusive” result. Pandya guessed it could be related to sample quality — if some slides have too much or too little tumor material, the model may not be able to read it with confidence. “Quite a wide variety of reasons exist, and we are diving further deeper with our partners to be able to understand what we can do to increase the percentage of patients who can actually benefit from the product,” she said.
Pandya did not comment on any US commercialization aspirations but said Panakeia is working closely with regulators in several countries to expand access to the product and continue to validate its efficacy.
AI-based pathology is a promising alternative to traditional molecular testing. Several companies are exploring ways for AI to identify biomarkers from hematoxylin and eosin slides, including French firm Owkin, which also markets an AI digital pathology tool called MSIntuit CRC to “pre-screen” colorectal cancer patients for MSI and dMMR. In a 2023 study, MSIntuit correctly classified MSI/dMMR tumors 96 percent of the time. The test is available in the EU but hasn’t yet been approved by the US Food and Drug Administration.
When Tempus acquired digital pathology firm Paige last year for $81.3 million, it folded into its portfolio Paige Predict, which launched last month and uses AI to analyze biopsy slides and identify clinically actionable biomarkers. The firm claims Paige Predict can detect biomarkers even when slides contain too little tumor material for full molecular profiling.
Performance of these types of tools can vary, according to a study published this month that included the breast cancer version of Panakeia’s PANProfiler among other AI tools and compared their performance against each other as well as against pathologists’ determinations of HER2 expression. Both AI and pathologists had trouble identifying tumors with low HER2 expression, but they agreed 65 percent of the time. The 10 models assessed also agreed with each other 65 percent of the time, whereas the three pathologists in the study agreed with each other 70 percent of the time.
“As therapeutic decision-making increasingly depends on more granular distinctions in HER2 expression,” the authors wrote in Modern Pathology, refining and standardizing the AI models’ performance and interpretation is key for “consistent and clinically meaningful application.”
Pandya agreed that there is a need for standardization, which is why Panakeia participated in the study, which was organized by non-profit think tank and advocacy group Friends of Cancer Research. “We are working with regulators as well to help establish guidelines standards around … regulating AI solutions,” she said.