What does it mean to match trials by biomarker instead of by cancer type?

Traditional trial search matches a patient to trials that name their cancer type — 'pancreatic cancer' finds trials with 'pancreatic cancer' in the eligibility. Matching by biomarker instead looks for trials that enroll by a molecular target the tumor carries, regardless of where the cancer started. Many modern trials are written this way — 'advanced solid tumors expressing HER2,' 'B7-H3-positive tumors' — because the drug targets the molecule, not the organ. Matching by biomarker means finding those cross-cancer 'basket' trials for a patient whose specific cancer may never be named in the trial's criteria.

Why do cancer-name searches miss basket trials?

Because a basket trial's eligibility often doesn't contain the patient's cancer name. A trial written as 'advanced solid tumors expressing Claudin 18.2' may be open to a gastric, pancreatic, or biliary cancer patient, but a text search for those cancer names finds no match — the words don't line up, even though the biology does. Basket trials have become one of the fastest-growing parts of the oncology pipeline, so a search that only matches cancer names can silently skip a large and growing share of relevant trials.

How does ClinTrialFinder match basket trials across cancer types?

This runs alongside — not instead of — standard diagnosis-based retrieval, which finds the trials written for the patient's specific cancer. On top of that, three things happen in sequence for the cross-cancer layer. First, a tumor-agnostic basket trial is identified as one at the corpus level, so it isn't lost among disease-specific trials. Second, its mechanism is distilled to a clean target-and-modality label — what the drug goes after (for example, a surface antigen) and how (for example, an antibody-drug conjugate or CAR-T cell therapy). Third, for a given patient, a language model decides which of those baskets actually fit that patient's profile — reasoning across cancer types the way an expert would, including whether the patient's own cancer is one that expresses the target, rather than only string-matching the cancer's name. ClinTrialFinder does not publish the internal details of this pipeline, but that is the shape of it.

What is the hardest part of cross-cancer trial matching?

The tension between recall and precision. If the matching is too permissive, patients are flooded with early-phase basket trials that mention their marker but are not real options. If it is too strict, it silently drops trials that were genuinely open — the worst failure in this domain, because the patient never learns what they did not see. Getting the right cross-cancer baskets in front of the right patient, while being honest about the rest, is harder than simply finding basket trials, and it is an ongoing area of work rather than a solved problem.

How Basket Trial Matching Works: Finding Trials by Biomarker, Not Just Diagnosis

The best clinical trial for a patient increasingly doesn't have their cancer's name on it. Here's the problem that creates — and how we approach it.

July 18, 2026

TL;DR: Modern targeted therapies are tested in basket trials that enroll by a molecular target across many cancer types, not by cancer name. That makes them nearly invisible to a search that only matches on diagnosis. ClinTrialFinder still matches by diagnosis like any tool — and also matches by the molecule: it identifies the tumor-agnostic basket, distills its mechanism to a target-and-modality label, then uses a language model to decide which baskets fit a patient's disease-and-biomarker profile. The genuinely hard part isn't finding basket trials — it's the tension between recall and precision, and being honest about it.

One Drug, Many Cancers

For most of the history of oncology, a trial enrolled one cancer type. "Metastatic pancreatic cancer." "Recurrent ovarian cancer." If your diagnosis matched, you might qualify. If it didn't, you moved on.

Modern targeted therapies broke that model. A drug like an antibody-drug conjugate or a CAR-T cell therapy doesn't attack a cancer by where it started — it attacks a specific molecule on the tumor's surface. And the same molecule shows up in cancers that started in completely different organs. A protein like HER2 appears in some breast, stomach, lung, and bladder cancers. So the trial gets written around the molecule: "advanced solid tumors expressing B7-H3." "Claudin 18.2–positive tumors." "HER2-expressing malignancies." One drug, many cancers, held together in a "basket."

            Why this matters: for patients with uncommon cancers, or who have exhausted the standard options for their specific diagnosis, a basket trial can be one of the most relevant options available — as long as their tumor carries the target.
        

Why a Cancer-Name Search Misses Them

Here's the trap. A patient's pathology report says "HER2-positive gallbladder cancer." The trial says "advanced solid tumors expressing HER2." A search by cancer name walks straight past a trial they'd actually qualify for — because the words don't line up, even though the biology does.

This is not a rare edge case. Basket-style, biomarker-defined trials have become one of the fastest-growing parts of the oncology pipeline — drawing on trial records from ClinicalTrials.gov, ClinTrialFinder's corpus identifies more than 1,300 actively-recruiting trials (as of July 2026) that enroll across cancer types by a shared biomarker or antigen. For a patient searching one cancer name at a time, most of those are effectively invisible. Closing that gap is the reason this kind of matching has to exist.

Matching by the Molecule, Not Just the Diagnosis

To be clear, the diagnosis still does most of the work. Like any trial-matching tool, ClinTrialFinder retrieves disease-specific trials by matching the patient's cancer and conditions — that's the standard, essential path, and it finds the trials written for their diagnosis. The difference is what happens on top of that: a separate layer that also looks across cancer types for basket trials the patient's tumor may qualify for by biomarker. Notably, even that layer still uses the diagnosis — it reasons about whether the patient's specific cancer expresses the target — so it's diagnosis and molecule, not one instead of the other.

That cross-cancer layer is where the interesting work is. At a high level, three things happen in sequence:

1. Identify the basket

A tumor-agnostic basket trial is recognized as one at the corpus level — separated from ordinary disease-specific trials so it isn't lost among them. This is what keeps "advanced solid tumors expressing X" trials in play at all for a patient whose specific cancer they never name.

2. Distill the mechanism to a target-and-modality label

Each basket's mechanism is reduced to a clean, comparable label — what the drug goes after (a surface antigen, for instance) and how (an antibody-drug conjugate, a CAR-T cell therapy, a bispecific antibody). This turns messy trial text into something a matcher can reason about consistently.

3. Decide relevance for this patient

For a given patient, a language model decides which of those baskets actually fit that patient's profile — reasoning across cancer types the way an experienced oncologist would ("this marker is expressed in that cancer, so this cross-cancer basket may apply"), rather than only string-matching the cancer's name. Note that this step still uses the diagnosis — it has to reason about whether the patient's cancer expresses the target — so it's diagnosis and biomarker together.

We don't publish the internal details of this pipeline, but that's the shape of it — and the shape is the interesting part, because it's where the hard problem lives.

The Hard Part: Recall vs. Precision

Finding basket trials isn't the hard part. The hard part is the tension between recall (surfacing everything genuinely relevant) and precision (not burying the patient in noise) — and in this domain that tension is unusually unforgiving.

Turn the matching up, and every patient drowns in early-phase baskets that mention their marker but aren't real options for them. Turn it down, and you silently drop trials that were genuinely open — the worst possible failure here, because the patient never knows what they didn't see. A short results list feels complete whether or not it is.

⚠️ How sharp that edge is — a real case

A patient with advanced nasopharyngeal carcinoma — heavily pretreated, out of standard options — should have matched roughly a dozen B7-H3 basket trials. B7-H3 is expressed in nasopharyngeal cancer; it's in the literature. They matched zero.

When we traced it, the relevance step had effectively asked "does this cancer express B7-H3?" and — leaning on second-tier evidence a general model doesn't reliably recall — answered no. Technically defensible. Clinically wrong. And completely invisible to the patient, who saw a short results list and reasonably assumed that was all there was. Catching and correcting that kind of miss, without flooding every other patient with irrelevant baskets, is the actual engineering problem — and it's ongoing work, not a solved one.

That's the real distinction between a trial search that looks impressive and one that's genuinely useful: not "can it find basket trials," but "can it find the right ones for this specific person, and be honest about the rest."

What This Means for You

            If you're a patient or clinician: know your tumor's biomarkers — HER2, TROP2, Claudin 18.2, B7-H3, Nectin-4, and the rest. Then make sure whatever you use to search looks across cancer types by that marker, not just by your diagnosis. The trial that matters may be filed under a name that isn't yours.
        

If you build in this space: the disease-name search is the easy 80%. Cross-cancer, biomarker-defined matching — done with honest recall — is where the actual patient value is, and it's harder than it looks. It's the kind of problem ClinTrialFinder is built around.

Find trials matched to your situation

A note on honesty. ClinTrialFinder helps you surface candidate trials to discuss with your oncologist and verify on ClinicalTrials.gov. It is a starting point, not a determination of eligibility, and it will not find every trial. Carrying a biomarker is necessary but not sufficient for a basket trial — the other eligibility rules (prior treatment, expression level, performance status) still apply. Enrollment decisions belong with you and your care team.

Related reading:

How We Rank Trials: Evidence-Informed Scoring
Browse trials by mechanism / target — B7-H3, HER2 ADCs, CLDN18.2, TROP2, and more

Questions? Contact us

← Back to Blog