Gilead and Nucleai probe why target-positive tumors can still resist ADCs
Nucleai, Gilead, and target-positive tumors
Antibody-drug conjugates (ADCs) have emerged as one of the busiest and most crowded niches within oncology, given their potential to use an antibody to ferry a potent cytotoxic payload toward tumor cells. The class now spans about 15 FDA-approved products, and the segment is growing steadily. Evaluate Pharma projects the ADC category will reach $57 billion by 2032.
The core promise of ADCs is to widen chemotherapy’s therapeutic window by using a tumor-targeting antibody to concentrate a potent cytotoxic payload in cancer cells while limiting exposure to healthy tissue.
“Initially, what we were told was that ADCs are basically targeted chemotherapy. I’m going to put a GPS signal on my beachhead, on my chemotherapy target, and it’s going to zone in just to the tumor cells and release its payload,” said Ken Bloom, MD, head of pathology at Nucleai. The promise essentially is, as Bloom put it, getting “all the benefits of chemotherapy without the toxicity.” But “in practice it’s far trickier than that,” he said. A tumor can look target-positive on a pathology slide while the drug’s cell-surface target is altered or inaccessible, extracellular proteases in the tumor microenvironment can cleave some ADC linkers and release their payload, and cancer cells can use drug-efflux pumps to expel certain released payloads.
That gap between what a pathology assay sees and what an ADC encounters is one of the questions Nucleai is probing with Gilead Sciences. The companies disclosed August 11 that Nucleai has analyzed H&E and IHC whole-slide images from several Gilead clinical studies across multiple oncology indications, linking tissue features with clinical outcomes. The relationship began a few years ago with largely preclinical multiplex immunofluorescence work and has expanded into retrospective analysis of clinical-trial datasets.
Gilead described the collaboration as an extension of its existing translational medicine and biomarker capabilities. Nucleai brings “specialized expertise in AI-driven spatial biology and tissue analytics,” including experience analyzing complex tumor microenvironments at scale, said Meghna Das Thakur, senior director of oncology biomarkers at Gilead. She said combining those capabilities with Gilead’s scientific and clinical expertise could generate insights more efficiently.
Bloom said a presentation and peer-reviewed publication could come later this year or early next year, although the timeline remains fluid. He said the retrospective analysis has surfaced several candidate biomarkers. “What fell out of the analysis was several candidate biomarkers that are more predictive than standard immunohistochemistry,” he said, describing them as “things that you could do with the standard H&E and immunohistochemical slide that add further insight beyond just quantifying expression levels.” Those candidates are undergoing deeper analysis.
One reason a conventional IHC expression score can miss therapeutically relevant information is that staining can establish that a target is present without showing whether an ADC can actually engage the relevant extracellular epitope. The inside-outside epitope problem is established in the literature, with HER2 being a core example. The widely used Ventana 4B5 IHC antibody for HER2 detection binds the intracellular domain, meaning it can detect full-length HER2 as well as truncated forms that retain that region. Trastuzumab, meanwhile, binds the extracellular domain. As a 2011 Cancer Research review explains, “p95HER2 fragments arise through at least 2 different mechanisms: proteolytic shedding of the extracellular domain of the full-length receptor and translation of the mRNA encoding HER2 from internal initiation codons.” The resulting membrane-bound fragments lack the extracellular region targeted by trastuzumab.
“There’s a reason that we look at the internal side as pathologists, that we try to stay inside the membrane,” Bloom said. Tissue processing tends to preserve the intracellular portions of membrane proteins better than their extracellular regions, he explained. The exposed exterior can also undergo changes such as glycosylation. “So when we make our antibodies, and this is sort of a hidden little secret in pathology, almost all the antibodies that we use are directed at the inside membrane of the protein. Yet it’s the outside that you should really care about.”
That gap between what IHC detects and what a drug can bind has been associated with clinical outcomes. In a 2015 study of trastuzumab-treated breast cancer patients, researchers who separately measured HER2’s intracellular and extracellular domains found discordant results in 15% of cases, with higher extracellular-domain expression associated with longer disease-free survival. A 2022 study co-authored by Bloom identified a related measurement problem for ADCs: conventional HER2 assays were designed to distinguish amplified from unamplified tumors and offered limited resolution within the lower range of HER2 expression that may matter for drugs such as trastuzumab deruxtecan. The practical obstacle, Bloom said, is developing pathology assays that capture therapeutically relevant features while remaining stable and reproducible across laboratories.
That challenge helps explain why Bloom draws a distinction between digital and computational pathology. For much of his 40-year career, he said, new pathology tools primarily helped pathologists perform existing tasks more consistently. Computational pathology can extract relationships that were difficult to measure directly from a slide, including the density of particular cells, their spatial distribution and the cellular neighborhoods surrounding a target.
“Now we care about the neighborhood that a cell is in,” Bloom said. “What’s the relationship between it and other cells in the vicinity? What’s the density of things?” Quantifying those relationships adds another dimension to conventional pathology and could help explain why tumors carrying the same apparent biomarker behave differently.
The implications extend beyond drug development. Bloom argued that computational pathology could raise the floor of what pathologists can deliver reliably. He pointed to studies showing that AI assistance can improve pathologist accuracy and consistency, gains he attributed in part to reducing the effects of fatigue and differences in training. “When you bring an AI overreader into the process, everything gets better,” he said. “It catches mistakes. It gets all pathologists more uniform, which was always a big problem.”
Computational pathology hinges on expert human judgment, of course, and Bloom sees computational tools as a way to reach the level of accuracy required for clinical use while keeping a trained physician in the loop. “A fool with a tool is still a fool,” he said, citing a line he has used to open lectures since the early days of digital pathology.
He also sees a communication benefit. Pathology reports can be difficult for oncologists and surgeons to interpret, Bloom said, and computational tools can paint tumor and non-tumor tissue in distinct colors, giving non-pathologists an intuitive spatial picture they could never extract from narrative text alone.
The value still depends on pairing those capabilities with domain expertise. “AI by itself isn’t going to do it, but a well-trained physician with AI is going to be light years ahead of one without,” Bloom said.
How quickly that model spreads across pharma remains harder to predict. Bloom sees several organizations pushing ahead, though the industry has yet to establish a repeatable route that others can readily follow. “I think there still isn’t a pathway that’s been forged clearly yet for somebody else to follow,” he said. “But the good news is that there are several leaders out there attempting to forge that first path. What you’re going to see, as soon as the first one hits, is a wave that follows.”
