An Endpoint Is Not a Biomarker, and That Difference Can Kill a Drug : Rethinking Endpoints in the Age of AI (Part 2)

In the previous chapter, we looked at a case that generated considerable noise in the community: the supposed “approval of an AI-based endpoint.”

But when we broke it down, something more interesting emerged.

It wasn’t a new type of endpoint. It wasn’t a redefinition of clinical benefit. It was something else entirely an improvement in how we measure it.

The PathAI case didn’t redefine what an endpoint is. It revealed something deeper: that even when we think we know what to measure, we remain dependent on how we measure it.

That’s where this second chapter begins.

Because understanding what an endpoint is isn’t enough. You have to understand how it’s constructed, what types exist, and above all, why some bring us closer to clinical truth — while others pull us further away without us realizing it.

There’s a curious paradox in clinical development: the longer you’ve been in the field, the easier it is to start using words imprecisely.

Not because you don’t know what they mean. But because you use them so often that the nuances stop mattering.

Endpoint. Biomarker. Outcome. Surrogate. Primary criterion.

In day-to-day conversations, these words blur together. In internal presentations, they get swapped. In scientific papers, they appear as synonyms when they aren’t.

For a long time, that didn’t seem like a serious problem.

It is. And not for semantic reasons.

This imprecision introduces a silent distortion into how we design trials, how we interpret results, and ultimately, how we make clinical decisions.

The technical definition is familiar: an endpoint is the predefined measure used to evaluate the effect of a treatment in a clinical trial, the variable on which the study’s statistical and clinical conclusions are built.

But stopping there means staying on the surface.

What that definition doesn’t say is this: an endpoint is the formalization of what a society (through its regulatory agencies and scientific community) considers a meaningful clinical benefit.

It’s not just a variable. It’s a filter. A filter that decides what counts as evidence… and what doesn’t.

The FDA expresses this through a triad worth memorizing: clinical benefit is a favorable effect on how a patient feels, functions, or survives.

Sensation, function and survival, that’s the gold standard. Everything else any variable that doesn’t directly measure those things, requires justification.

And this is where the real complexity begins.

Because in many diseases, waiting to measure how a patient feels, functions, or survives can take decades for example: Alzheimer’s, MASH and Idiopathic pulmonary fibrosis. Diseases where the definitive clinical events as death, severe dementia, transplantation may take twenty years to manifest.

Drug development cannot wait twenty years.

So we use shortcuts, well-intentioned shortcuts, sometimes well-validated, and sometimes not. And it’s in that tension between what we want to measure and what we can measure that most of the problems live.

Every endpoint is, at its core, a negotiation between what matters and what’s measurable. And like all negotiations, it involves trade-offs.

We lose time if we wait for the real clinical endpoint. We lose certainty if we use a surrogate. But the most important thing isn’t that these trade-offs exist. It’s that we rarely make them explicit.

We design trials as if we were measuring clinical truth directly, when in reality we’re measuring an approximation. Sometimes a good one. Sometimes a dangerous one.