NEW YORK – Leading companies in the cancer genomics field are increasingly turning their attention to artificial intelligence technologies for digital pathology and spatial biology to enhance their existing clinical test offerings, to solve sample processing challenges, and, potentially, to uncover a new generation of predictive biomarkers.
For example, earlier this year, Guardant Health adopted South Korean AI firm Lunit's digital pathology technology for a new business arm, marketed as Guardant Galaxy, which it described as a planned suite of new technology applications to enhance the performance and utility of its portfolio of cancer tests.
The firm's co-CEO Helmy Eltoukhy said in an email that Guardant closely follows the evolution of a "variety of information-generating technologies … and the launch of Guardant Galaxy as part of our portfolio is essentially a formal recognition of the growing power of AI and data science in precision tissue diagnostics."
Guardant's initial implementation features Lunit's existing CE-marked PD-L1 scoring assay. Because of the way anti-PD-L1 drugs were initially developed and the complexity of their utility in different tumor types, the field has evolved to feature more than a dozen separate indications, with four immunohistochemistry testing antibodies, three different scoring systems and two separate slide-staining platforms. In this context, tools that can streamline and standardize what is normally a manual and subjective process have become increasingly attractive.
Lunit has reported that its Lunit SCOPE PD-L1 assay has shown a more than 20 percent improvement in the detection of PD-L1 positivity in challenging samples compared to manual pathologist interpretation.
According to Eltoukhy, Guardant is considering building on this initial IHC application by adding HER2 scoring, initially for biopharma partners, sometime later this year. The move would be a relatively simple one, as Lunit has already developed a HER2-scoring tool akin to its PD-L1 test.
The company has already collected evidence that it's digital approach can identify responders to anti-HER2 drugs that traditional IHC missed.
With the recent US Food and Drug Administration approval of the antibody-drug conjugate Enhertu (trastuzumab deruxtecan) from AstraZeneca and Daiichi Sankyo for patients with HER2-low disease, pathologists and oncologists also face a new conundrum.
In the trial that supported the FDA's nod, investigators defined HER2-low patients as those IHC 1+ or 2+ with negative ISH results. Manual IHC doesn’t currently have standards for scoring patients between 1+ and 0, but data suggests there could be a population of patients being missed in that range who still express some HER2 and would be sensitive to treatment. As a result, digital pathology firms are hoping to step in with a more standardized option.
Eltoukhy said that while PD-L1 and HER2 make sense as lead indications, the ability of AI-based analyses to augment human biomarker assessment is something the company is planning to "apply broadly across both established and novel targets."
The firm's next effort is to validate, with Lunit, an inflammatory score assay to predict patients' likelihood of responding to immune checkpoint inhibitors. The two companies are refining and validating a tool trained to infer ICI responsiveness from the patterning of immune cells among tumor tissue in a single H&E-stained slide.
In a non-small cell lung cancer study last year in the Journal of Clinical Oncology, Lunit researchers wrote that a higher inflammatory score correlated with higher response rates and prolonged progression-free survival compared to cases with what they called immune-excluded or immune-desert phenotypes.
Conference data has shown the same association in a pan-cancer setting, Lunit said. And more recent publications and presentations have reiterated this for specific tumor types including liver cancer and nasopharyngeal cancer.
"We are currently testing the technology internally and plan to roll it out for our biopharma collaborators shortly," Eltoukhy said.
Although unwilling to comment on the record, other cancer sequencing firms have said that they are also looking to AI and digital pathology to enhance their existing genomic profiling.
While Guardant decided to bring in an already-developed complementary technology, other firms have embraced broader in-house digital pathology programs, building out large-scale systems for new biomarker discovery and other applications.
Caris Life Sciences Executive Medical Director Matthew Oberley said in a corporate presentation last November that his firm has invested heavily in digital pathology over the last three years.
As was the case for Guardant, one major driver for Caris has been the opportunity to improve operational efficiencies by using AI algorithms to boost the speed and the reproducibility of IHC interpretation.
"Pathologists are very good at determining whether IHC is positive or negative, above 50 percent or below 50 percent, but they get a little bit shakier when it comes to determinations like, is this 0 percent. Is it 1 percent or is it above 1 percent? We think if the machine can do the job better, it would be in the best interest of the patients to have this reliability and reproducibility increase," Oberley said.
On the research and development side, he said Caris is in the middle of analyzing all the images the company has seen in its clinical testing of over 200,000 patients, "all of whom have matched exome and transcriptome data and, through a variety of sources, clinical outcomes data," something the firm believes will be a powerful dataset to develop novel algorithms.
The company uses three different types of scanners in a triage system, so that slides that can't be imaged using one platform are passed on to a more powerful system, with the goal being 100 percent digitization. These are managed by a system called Halo, which allows for parallel clinical and research databases, such that clinical cohorts can be de-identified and distributed to Caris' partners for research.
Tempus, another cancer sequencing staple, has also had an internal AI/digital pathology biomarker generation program in place for some time, developing a portfolio of algorithms using single whole-slide H&E images to predict biomarkers or guide pathologist review. These include detection of DNA alterations, such as FGFR, RNA expression of MET, broad genomic signatures like homologous recombination deficiency, or microsatellite instability.
The firm's senior VP of pathology, Nike Beaubier, said in an email that Tempus began its digital pathology efforts in early 2022, launching a proprietary platform that allows research into "AI models intended to identify specimens with potentially actionable biomarkers and/or prognostic spatial markers" using single H&E-stained slides.
"We realized there was an unmet need … so we leveraged our multimodal data library to develop AI models that aim to help pathologists and physicians identify patients who would benefit from additional testing and may qualify for targeted therapies … earlier in their cancer care journey," Beaubier said. "We are also working with our biopharma collaborators, such as Janssen, to jointly create pathology algorithms intended to further patient pre-screening efforts for specific cancer indications, including biomarker-selected clinical trial cohorts," she added.
A significant draw of artificial intelligence in the context of cancer biology is that it might be able to explore certain aspects of how cancerous tissue diverges from normal that DNA sequencing alone cannot access. According to Beaubier, Tempus is also active in this area.
In a poster presented at last year's US and Canadian Academy of Pathology meeting, Tempus investigators shared data from a study testing an algorithm to predict MSI status from H&E whole-slide images in prostate, gastric, and esophageal cancer.
High MSI is known to correspond to response to immune checkpoint inhibitors across tumor types, but because of a low prevalence in non-colorectal cancers, testing isn't routinely performed. Researchers wanted to explore whether digital slide images could be used to identify a subpopulation of patients more likely to test positive, offering an enrichment method for guiding the use of MSI testing in non-colorectal cancers.
The team first trained a neural network using matched tissue images and MSI molecular testing results from both prostate cancers and other tumor types in which MSI-high cases are more prevalent. When they tested the resulting predictor in a fully independent holdout set of prostate cancers, it caught all the true positives, though it also predicted high MSI in a majority of the negative cases.
Applied to two other cohorts with different tumor types — gastric and esophageal cancers — the predictor caught about 80 percent of MSI-high cases, despite not having been trained on samples of these tumors.
Biomarker discovery and diagnostics development aside, Tempus has also explored AI-enabled pathology to improve the less flashy but important realm of sample preparation. In a recent preprint abstract, the company described the development of an AI-augmented pathology review system, called SmartPath, that used digitized H&E-stained slides to inform subsequent microdissection.
Following model development, an internal validation trial tested the system on 501 clinical colorectal cancer slides. Half of the slides received SmartPath-augmented reviews while the other half received traditional pathologist reviews, with the SmartPath cohort ending up with 25 percent more DNA yields within a desired target range of 100-2,000 ng.
Authors reported that SmartPath also recommended fewer slides to scrape for large tissue sections, saving tissue in these cases, while diverting more slides to scrape for samples with scant tissue sections, helping prevent the need for re-extraction due to insufficient yield.
Oberley said that Caris has also been developing AI digital analyses targeted at improving sample processing.
"One of the workflows that we've had to customize because we're a molecular profiling lab is the initial tissue evaluation," he said.
When the company receives a specimen, it is standard practice to cut several unstained slides and then stain only the first and last. When a pathologist examines these bookends, seeing enough tumor at both sides implies confidence that the same is true in the middle.
"You can kind of predict that there is going to be [sufficient tumor] in all of those unstained slides, even though you can't directly look at it," Oberley said. But once stained, the tissue can't be used for molecular analysis.
According to Oberley, Caris receives a little over 100,000 specimens per year, among which between 15 and 20 percent have limited tissue, making it difficult or impossible to cut the necessary number of unstained slides.
Pathologists have adopted the practice of staining the first slide and visually examining the rest of the slides, unstained, to try to glean whether tumor tissue persists all the way through.
"It's difficult because you can deplete the cancer tissue at the same rate as surrounding benign tissue, or you can deplete the cancer tissue earlier depending on the orientation of the cancer in the block," Oberley said. "It really matters between these two cases because in the first case you can order NGS because you have enough tissue, and in the second case there's not enough tissue remaining to do any kind of sequencing. If we knew that this was the case, we would recommend reflex IHC only."
What Caris has discovered in its digital pathology exploration is that it can obtain very high-resolution scans of unstained slides, which offer an opportunity to try to solve what Oberley called pathology's "version of the observer effect."
"If you stain a slide, you can see what's on it, but you can't do molecular testing, and if you leave it unstained, you can't see what's on it, but you could do molecular testing."
To solve this, the company explored whether it could use digital images of unstained slide to predict what a stained slide would look like — essentially creating a virtual stain — and has been using a neural net to try to do so, presenting stained and unstained images of the same slides and challenging the AI to find a way to infer the latter based on the former.
According to Oberley, Caris has had some good success. "It works in about 70 or 80 percent of cases right now, generally the cases that are well-differentiated cancers," he said.
To try to close that gap, the company is now experimenting with alternate light sources and has found that using higher wavelength light can produce autofluorescence in the tissue that provides enough extra information for the neural net to close those holes, getting the company closer to being able to virtually stain 100 percent of samples.
The key moving forward, Oberley said, is now finding a way to scale this up. "We're getting so many slides every day [that] we need somebody to help us build a scanner that can accommodate [this], and so we're currently working with some external companies on that."