Engineering the Future of ADCs: Innovation, Efficacy, and Clinical Integration

How tumour biology, chemistry and AI are converging to shape the next generation of antibody–drug conjugates

In a rapidly maturing antibody–drug conjugate (ADC) landscape, the central challenge is no longer simply building potent conjugates—it’s learning how to match the right design to the right tumour biology with enough confidence to reduce clinical attrition. That was the theme running through this Thought Leadership session, Engineering the Future of ADCs: Innovation, Efficacy and Clinical Integration, featuring Dr Victor Jeffrey Leyton (University of Ottawa) alongside panellists Professor Sophia Karagiannis (King’s College London), Dr Sharsti Sandall (Pfizer) and Dr Stuart Barnscher (Zymeworks).

Across the discussion, a clear message emerged: chemistry enabled the first wave of ADC successes, but biology—and increasingly data—will define the next wave.

From antibody imaging to ADC delivery: the pivot to intracellular biology

Dr Leyton’s entry into ADCs began not in therapy, but in molecular imaging. During his PhD, he developed radiolabeled antibody conjugates for PET imaging and explored reduced-Fc formats designed to clear faster and support same-day tumour visualization. That work delivered a key lesson that still guides his approach: imaging optimizes contrast, therapy demands absolute uptake. A tumour can look “bright” on a scan with a favourable tumour-to-background ratio, yet still receive insufficient drug payload for meaningful cytotoxicity.

That distinction, combined with the first solid-tumour ADC approvals around 2013, pulled Dr Leyton toward a different question: why do antibody conjugates often fail to accumulate efficiently inside target tumour cells? His work in routing antibody conjugates toward nuclear localization and understanding intracellular transport highlighted a persistent bottleneck in ADC performance: intracellular delivery is frequently the limiting step, not simply binding or circulation.

Why ADC development is still too empirical

Although the ADC field is highly interdisciplinary, Dr Leyton argued that many development decisions remain “mix-and-match” in practice—pairing different antibodies with a relatively narrow, sometimes redundant set of linker–payload systems because those combinations are established and manufacturable. That approach has delivered real progress, but it also contributes to a costly development cycle where early designs can be advanced before tumour-specific biological risks are fully understood.

Dr Leyton framed this as an unsustainable economic model across the ecosystem: from academic labs with limited infrastructure, to biotechs managing investor timelines, to large pharma investing billions into a single asset. Reducing empirical iteration early—without sacrificing rigour—could improve long-term clinical value.

Antigen selection: expression is necessary, but not sufficient

Target choice remains one of the most consequential decisions in ADC development—and also one of the hardest to standardize.

Professor Karagiannis highlighted the practical problem: most tumour targets are self-antigens, meaning the therapeutic window is constrained by normal tissue expression. Even within tumours, heterogeneity is substantial. The early assumption that “high expression across most tumour cells” is required has been challenged by clinical and preclinical evidence—especially where bystander effect can compensate for lower or uneven antigen density.

HER2 illustrates the point. It has become a reference target for “high expression,” yet its expression levels can be orders of magnitude above many other “overexpressed” antigens. Panellists agreed that this makes HER2 both instructive and misleading: it demonstrates what is possible when biology aligns favourably, but it may not generalize.

Dr Sandall added an important biological nuance: expression alone misses key variables—what is the target’s normal biology, and can it be exploited? She gave the example of an ADC targeting PD-L1, which is not always highly expressed, but is present on tumour microenvironment (TME) cells as well as tumour cells. That raises a strategic possibility: TME targeting may contribute to efficacy, not just toxicity risk—depending on biology and payload.

Tumour heterogeneity is multi-layered: intra-patient, inter-patient and longitudinal

Dr Barnscher emphasized that heterogeneity operates on multiple scales. Beyond variation between patients, different regions of a single tumour can show dramatically different target expression, and a biopsy represents only a small fraction of the full disease. He referenced analyses where single-cell expression readouts varied by tumour region, underscoring why a “single target, single threshold” mindset often breaks down.

Professor Karagiannis also pointed to an increasingly important clinical reality: metastatic and treatment-resistant disease may not retain the same target biology as primary tumours. Longitudinal biopsies can reveal targets that are lost—or selectively retained—under therapeutic pressure, which affects whether an ADC strategy remains viable later in the patient journey.

One response to heterogeneity discussed by Zymeworks is bispecific targeting: designing ADCs capable of binding either of two antigens so delivery remains possible even when one is absent or low. That approach introduces new questions around affinity, avidity, internalization and the risk of trans-binding between cells—areas where format engineering and empirical testing still dominate.

Payload and linker selection: chemistry must serve biology

If target selection determines where an ADC goes, payload and linker design determine what happens next—and these decisions are still difficult to predict.

Dr Sandall described the “humbling” challenge of making payloads work in an ADC context: it’s not enough to have a potent drug; the chemistry must allow stable conjugation, suitable drug-to-antibody ratio, controlled release, and minimal off-target uptake. Linkers, in particular, shape whether payload release occurs primarily in the endosome or lysosome—and whether drug remains inside the cell or is susceptible to non-specific release.

Importantly, she noted that some cleavable linkers can be activated in endosomal compartments, meaning effective payload release doesn’t always require full lysosomal trafficking. But the field still lacks systematic “rules of the game” that link target internalization biology to optimal linker choice.

The discussion also highlighted the future direction of payloads: rather than relying mainly on broadly cytotoxic warheads, the field is moving toward vulnerability-matched payloads—using tumour biology to select inhibitors that exploit specific dependencies. Professor Karagiannis described work using CDK2 pathway biology to guide payload selection, demonstrating tumour control with lower payload exposure by delivering an inhibitor directly to cells most dependent on that pathway.

Preclinical models and toxicity: “all models are wrong—some are less wrong”

When the conversation turned to translational hurdles, panellists were candid: predicting clinical toxicity remains one of the field’s biggest gaps, especially in solid tumours.

Dr Leyton noted the downstream emergence of toxicities and real-world evidence reshaping the understanding of approved assets. Panellists discussed the need for retrospective analyses comparing clinically successful and unsuccessful ADCs to establish better predictive assays—using known positives and negatives as anchors.

Dr Barnscher captured the reality succinctly: models are imperfect, and different assays are needed for different questions (efficacy, PK, haematologic toxicity, organ-specific risks). Bone marrow toxicity is comparatively well-modelled, but even there, translation to human dose limits is challenging. The takeaway was pragmatic: use multiple complementary models, understand their flaws, and benchmark against known clinical comparators.

Where AI can help: narrowing the design space, not replacing experiments

Across the panel, AI was positioned not as a shortcut to certainty, but as a way to reduce waste—especially in early design.

Dr Barnscher described AI’s role in the “design” stage of a build–test loop: proposing mutations, improving developability, narrowing candidate lists, and prioritizing sequences likely to retain binding and stability. Dr Sandall and Professor Karagiannis highlighted the need for “reverse translation”—using clinical outcomes and real-world data to inform future designs—while acknowledging that much clinical biomarker data is not broadly available.

Dr Leyton closed by describing his lab’s approach: integrating large-scale transcriptomic and proteomic datasets across tumour and normal cell lines with decades of curated ADC structure–activity data, including models to address missing proteomics values. The aim is modest but valuable: predict whether an ADC design is likely to be active within defined thresholds, and narrow the experimental search space.