Q&A: Predicting response and resistance in ADCs with AI and multimodal data

Razik Yousfi: AI adds value across the entire development lifecycle, including screening for and developing novel biomarkers, de-risking existing programs, expanding labels into new indications, going deeper into mechanisms of action and resistance, and then deploying those tools in the clinic. Computational pathology, in particular, points to a clearer path to better patient selection. We already see early signals: a computational TROP2 biomarker associated with a 43%² risk reduction, and the first breakthrough designation granted to a computational pathology companion diagnostic.³ What's changed is the technology itself. Foundation models trained on very large, multimodal datasets now outperform more traditional AI approaches. Because they learn rich, general representations, we can adapt them with far less data, generalize across indications, and even interpolate modalities a patient may be missing. That work is built on Tempus' de-identified dataset of more than 500 petabytes, roughly 9 million whole-slide images, and the ability to reach about half of oncologists in the US.

Kevin Kolahi, MD, PhD: ADC efficacy comes down to three things: target expression, internalization of the ADC, and payload sensitivity. Antibody engineering has advanced so quickly that target expression is no longer the bottleneck it once was, which is why we now see responders in HER2-low and even HER2-negative disease. The problem is that IHC measures only that one factor, and it consumes scarce tissue to do it. We're trying to binarize a multifactorial question with a single-variable assay. My belief as a pathologist is that the other determinants of efficacy are encoded in the morphology of the H&E slide, the standard hematoxylin and eosin stain every patient already receives, and can be read directly with the right AI tools. The path forward is panel-based, computational profiling that interrogates all of these determinants at once. The goal is better decisions with less tissue.

Razik Yousfi: Imaging is a modality rich in information that remains vastly underutilized today. Gene expression expresses itself morphologically, and it may be difficult for a human to see those patterns through a microscope. But when you feed billions of tiles and millions of images to a complex network, it learns those patterns and can identify the signatures of those expressions. That was the hypothesis we started with, and it's the journey we're still on. In our experiments, we can emulate the output of a DNA panel from an H&E image and, more often than not, predict IHC status without running the IHC. We can also tune the model for negative or positive predictive value, so you can decide whether an IHC is even worth running for a given patient. This only works with very large paired datasets; H&E linked to DNA, enriched with IHC images and whole-transcriptome RNA. Once you have that, a digital biomarker becomes a screening tool: run it on every patient, then reflex to IHC, PCR, or comprehensive genomic testing as needed, getting more patients onto the right therapy faster.

Aliki Taylor, MD, PhD: It's an exciting time in epidemiology, because we simply haven't had this kind of data before. Real-world data is messier than a randomized controlled trial; there's missing data, scans and biopsies aren't aligned to trial schedules, follow-up varies, and patients are often older and sicker than trial populations. But having multimodal datasets like the ones Tempus has now is changing the game. It lets us combine the electronic health record with claims, pathology, and outcomes, and potentially histology and CT scans, which we haven't been able to do before, and then bring in AI. The key is a robust methodology that emulates the trial as closely as possible. Generating regulatory-grade evidence is the gold standard; most of the time, we're informing clinical decision-making.

Kevin Kolahi, MD, PhD: Ordinarily, we study resistance by comparing pre- and post-treatment biopsies, but pathology gives us a static picture at a single point in time; we don't see the dynamics of response. There's also a survivorship bias built into the study design: in patients where the drug works, you've eradicated the tumor, so by the time you take the post-treatment biopsy, there's little viable tumor left to analyze. The patients who could teach us the most are the ones we can't study. So we study the failures and use causal-inference and counterfactual methods to model what would have happened in responders, or to identify which resistance features were latently present at baseline. That's the kind of complexity where AI earns its place. It's not just pattern recognition; it's really helping us answer questions about a world we can't directly observe.

Razik Yousfi: On our side, we're building foundation models over longitudinal patient trajectories; pre-treatment through multiple lines of therapy, with DNA, RNA, ctDNA, treatments, and outcomes such as real-world overall survival and progression-free survival. The models are engineered to predict the next event from everything we know about a patient, and by decomposing those trajectories we can see which variables push a patient in one direction or another. That opens the door to counterfactual analysis, for example, asking whether a patient is likely to respond better to one class of drug than another. It isn't magic. We learn what a patient like this looks like, then compare the benefit of one treatment versus another.

Kevin Kolahi, MD, PhD: I'd separate cultural adoption from logistical adoption, meaning the infrastructure and investment. On the logistical side, you have to demonstrate the impact, and adoption follows, because people want to do better by the patient. On the cultural side, I'd reframe the conversation: pathologists have always inferred molecular information from morphology; we've been doing a form of computational inference in our heads. We've just never been able to do it in a systematic, quantitative, reproducible way. It also matters whether we describe AI as an efficiency tool or as a genuinely new diagnostic test; conflating the two creates confusion. Our goal isn't to replace pathologists; it's to give them a powerful new instrument that will enable better patient care. This is really a tool that augments the expert, and that always wins.

Aliki Taylor, MD, PhD: For regulators and payers, it comes down to transparency, early conversations, the quality of the data, and a study design that is robust and fit for purpose. Acceptance of real-world evidence has shifted enormously over the past five to ten years; it's well established for external control arms in single-arm oncology trials, and it's being used more broadly all the time. AI adds another layer of complexity, and regulators and health technology assessment bodies need time to determine how they'll validate and reproduce these models. The onus is on industry to keep doing high-quality, transparent studies, and I'd expect guidance to continue evolving relatively soon.