Precision oncology clinical trial matching — molecular profile to NCI-MATCH-style basket/umbrella trial arm, computationally ranked
Before any trial can be matched, a patient's tumor biology and clinical status must be converted into a structured, machine-queryable profile. Modern precision oncology draws on next-generation sequencing (NGS) panels covering hundreds of genes, immunohistochemistry (IHC) and fluorescence in situ hybridization (FISH) for protein/gene amplification, and standard clinical assessments (ECOG performance status, prior treatment history, organ function labs). Trial-matching engines like the NCI's own Clinical Trial Search tool ingest all of this simultaneously.
A comprehensive genomic profiling (CGP) report typically includes several distinct data categories, each queried differently by a trial-matching engine:
• Short variants (SNVs/indels): point mutations and small insertions/deletions — e.g., EGFR L858R, KRAS G12C, BRAF V600E — reported with variant allele frequency (VAF) • Copy number alterations: gene amplifications (HER2/ERBB2, MET) or deletions (CDKN2A) — reported as fold-amplification • Gene fusions/rearrangements: e.g., ALK, ROS1, NTRK fusions — detected by RNA-seq or FISH, often the single most "actionable" finding class • Biomarker scores: tumor mutational burden (TMB, mutations/Mb), microsatellite instability (MSI) status, PD-L1 tumor proportion score (TPS) by IHC • Germline vs. somatic flags: distinguishing inherited (BRCA1/2 germline) from tumor-acquired mutations, relevant for PARP inhibitor trials and family risk counseling
Each of these fields maps to structured eligibility criteria that trials publish in machine-readable form (e.g., ClinicalTrials.gov's eligibility module, or richer OMOP/FHIR-based representations used by institutional matching systems).
Molecular findings alone never determine trial eligibility. Every trial protocol layers on clinical criteria that must be satisfied simultaneously:
• ECOG performance status: most phase I trials require ECOG 0–1 (fully or mostly active); many phase II/III trials permit ECOG 0–2. A single point of ECOG can eliminate dozens of otherwise-matched trials • Prior lines of therapy: many trials specify a minimum ("progressed after ≥1 prior line") or maximum ("treatment-naive only," or "≤2 prior lines") — reflecting whether the drug is being tested as first-line, later-line, or salvage therapy • Organ function labs: creatinine clearance (renal), total bilirubin and AST/ALT (hepatic), absolute neutrophil count and platelets (hematologic), and for cardiotoxic agents, left ventricular ejection fraction (LVEF) — each trial sets numeric cutoffs • Washout periods: minimum time since last chemotherapy, radiation, or surgery (commonly 2–4 weeks) before trial therapy can begin • Comorbidity exclusions: active brain metastases, autoimmune disease, prior severe immune-related adverse events, and concurrent malignancies are common exclusion triggers
A matching engine represents each of these as a discrete queryable field so that eligibility logic — often boolean combinations of dozens of criteria — can be evaluated programmatically against every open trial in the registry within seconds.
In practice, a single patient profile touches 15–30 discrete structured fields. Trial protocols encode eligibility as boolean logic trees over these same fields — which is precisely what allows automated matching engines to evaluate thousands of trials in the time it takes a clinician to read one protocol PDF.
ClinicalTrials.gov lists tens of thousands of interventional oncology studies at any given time, of which several thousand are actively recruiting with molecular eligibility criteria. The first and highest-yield filtering step is biomarker matching: does the patient carry the specific mutation, amplification, fusion, or expression level that a given trial requires? This step alone eliminates the overwhelming majority of trials in the registry.
ClinicalTrials.gov, maintained by the U.S. National Library of Medicine, is the largest clinical trial registry in the world. Every trial record includes a structured "Eligibility Module" with free-text and (increasingly) coded inclusion/exclusion criteria, alongside study design fields: phase, intervention, sponsor, and recruiting status.
Because the raw registry is largely free-text, most production-grade trial-matching tools layer a natural-language-processing (NLP) or structured-annotation pipeline on top:
• NCI's Clinical Trial Search API: exposes structured biomarker and disease-site fields that institutional matching tools query directly • Institutional matching platforms (e.g., academic cancer center "precision medicine tumor boards"): maintain curated internal databases mapping each open trial's molecular criteria to a controlled vocabulary (HGVS variant nomenclature, HGNC gene symbols) • Commercial/nonprofit matching services: extract biomarker criteria from trial protocols using rule-based and ML-based NLP, then expose a queryable API that intake systems can call with a patient's variant list
The query itself is a set-intersection problem: for a patient with N reported alterations, find every open trial whose inclusion criteria are satisfied by at least one alteration (basket-trial style) or by the specific combination required (biomarker-stratified single-arm trial).
Of ~2,800 actively recruiting oncology trials, a typical patient profile — even one with an actionable alteration — matches only on the order of dozens after biomarker filtering, and often far fewer for rare mutations. This low yield reflects deliberate trial design, not failure:
• Common actionable mutations (EGFR exon 19del/L858R in lung cancer, BRAF V600E in melanoma) already have FDA-approved targeted therapies, so fewer investigational trials target them — most current trials focus on rarer or resistance-associated alterations • Rare/novel actionable mutations (e.g., NTRK fusions, RET fusions, HER2 exon 20 insertions) have fewer competing approved options, so a smaller absolute number of trials exist, but each is a stronger scientific match for the small number of eligible patients • Basket trials (single mutation, many tumor types) and umbrella trials (single tumor type, many mutation-defined arms) partially offset this scarcity by aggregating rare-mutation patients across disease sites into one master protocol — this is precisely the NCI-MATCH design discussed in Stage 5
Biomarker specificity therefore cuts both ways: a very rare, highly actionable mutation yields fewer raw candidate trials, but the trials that do match tend to be a much stronger, more targeted fit — reflected in the higher composite match scores rare mutations receive in Stage 4.
Landmark screening studies (e.g., the original NCI-MATCH trial, Flaherty et al. 2020) found that of ~6,000+ patients whose tumors were sequenced, roughly 2–10% ultimately enrolled in a molecularly matched treatment arm — most were excluded either because no actionable alteration was found, or because a matching trial existed but the patient failed a clinical eligibility criterion (Stage 3).
Passing the biomarker filter is necessary but far from sufficient. Every candidate trial carries a dense list of clinical inclusion and exclusion criteria — often 20–40 discrete conditions — covering performance status, prior treatment history, organ function, comorbidities, and washout timing. This is the stage where the largest fraction of otherwise molecularly matched patients are screened out, and it is also the stage most prone to manual review error when done without computational support.
A representative eligibility screen for a phase I/II biomarker-matched trial cross-checks the patient profile against criteria such as:
Performance and demographics: • ECOG performance status 0–1 (occasionally 0–2 for later-phase studies) • Age ≥18 years (pediatric trials have separate, much narrower registries) • Life expectancy ≥12 weeks per investigator assessment
Disease and treatment history: • Histologically confirmed diagnosis matching the trial's tumor-type requirement (or "any solid tumor" for tissue-agnostic basket trials) • Measurable disease by RECIST 1.1 criteria • Specific number of prior systemic therapy lines (minimum and/or maximum) • Adequate washout since last chemotherapy, radiotherapy, or major surgery (typically 2–4 weeks)
Organ function laboratory thresholds: • Absolute neutrophil count ≥1,500/μL, platelets ≥100,000/μL (hematologic reserve) • Creatinine clearance ≥50–60 mL/min (renal, Cockcroft-Gault formula) • Total bilirubin ≤1.5× ULN, AST/ALT ≤2.5–5× ULN depending on liver involvement (hepatic) • LVEF ≥50% by echocardiogram for cardiotoxic study drugs
Exclusions: • Untreated or symptomatic CNS metastases (many trials exclude; some now specifically include stable brain mets) • Active autoimmune disease or prior grade ≥3 immune-related adverse event (relevant for immunotherapy combination trials) • Concurrent investigational agent or competing trial enrollment
Historically, eligibility screening required a clinical research coordinator to manually read each candidate trial's protocol and cross-reference it line-by-line against a patient chart — a process taking 15–30 minutes per trial and prone to missed criteria under time pressure. Automated matching engines instead represent each criterion as a structured predicate evaluated against the same structured patient profile built in Stage 1:
• Numeric threshold criteria (labs, ECOG, age) map directly to profile fields and evaluate as simple boolean comparisons • Temporal criteria (washout periods, "progressed within the last N months") require date-arithmetic against treatment history timestamps • Free-text criteria (comorbidity exclusions, specific prior-therapy-class exclusions) require NLP extraction from clinical notes or structured problem lists — historically the weakest link, though large-language-model-assisted extraction has substantially improved recall in recent institutional pilots
The output of this stage is a per-trial checklist: each criterion marked pass/fail/unknown, with "unknown" flagging fields that require a coordinator to manually confirm (e.g., a lab value not yet in the structured record). Only trials with zero hard fails advance to ranking in Stage 4.
Multiple retrospective studies comparing manual versus computational eligibility screening at academic cancer centers have found that automated cross-checking surfaces 20–30% more genuinely eligible trial matches than manual chart review alone — primarily because coordinators, working against time constraints, tend to stop screening once an early exclusion criterion appears to disqualify a trial, sometimes missing that the criterion does not actually apply.
Once a shortlist of biomarker-matched, eligibility-passing trials is assembled, the final computational step is rank-ordering. Not all matches are equal: a phase III trial testing an FDA-adjacent regimen for a common mutation is a different proposition than a phase I dose-escalation study for a rare fusion at a single distant site. Matching engines compute a composite score so the single best-fit trial surfaces at the top of the list a clinician reviews.
A representative composite match-strength formula, conceptually similar to those used by institutional precision-medicine tumor boards, combines three weighted components:
1. Biomarker specificity fit (largest weight, ~50%): how precisely does the patient's alteration match the trial's stated criterion? An exact variant match (e.g., trial requires "EGFR L858R," patient has exactly that) scores higher than a class-level match (trial requires "any EGFR activating mutation"). Co-occurring resistance mutations or variants of unknown significance reduce the score.
2. Trial phase and evidence maturity (~25%): later-phase trials (II/III) testing regimens with existing human efficacy data are generally weighted as higher-confidence options than early phase I dose-escalation studies, all else equal — though for patients who have exhausted standard options, phase I access to a novel mechanism may be scored as more clinically relevant despite the phase.
3. Geographic and logistic feasibility (~25%): distance from the patient's home to the nearest enrolling site, site capacity/slot availability, and whether the trial is actively enrolling (versus temporarily paused) all factor in — a scientifically perfect match at a site 2,000 miles away is frequently less actionable than a strong match 40 miles away.
Final score = 0.5×(biomarker fit) + 0.25×(phase/evidence weight) + 0.25×(geographic feasibility), typically normalized to a 0–100% scale and displayed alongside the ranked list so clinicians can see why a trial ranked where it did.
The output of ranking is not an automatic enrollment decision — it is a prioritized shortlist presented to a molecular tumor board or treating oncologist for clinical judgment. Best-practice matching platforms display, per ranked trial:
• Trial identifier (NCT number) and plain-language title • The specific biomarker criterion matched and the patient's corresponding finding • Trial phase, line-of-therapy positioning, and current enrollment status • Nearest enrolling site(s) and estimated travel distance • A composite match score, decomposed into its three weighted components so the rationale is auditable rather than a black box
This transparency matters clinically: a tumor board reviewing a ranked list needs to understand why trial A outranked trial B in order to trust — and occasionally override — the computational recommendation, for instance when a patient has a strong personal preference for a nearby site or when unstructured clinical nuance (not captured in the structured profile) makes a lower-ranked trial the better real-world choice.
The archetypal destination for a molecularly matched patient is a master-protocol trial: a single overarching study with dozens of parallel, biomarker-defined treatment arms sharing one infrastructure, one IRB approval, and one screening pathway. The NCI-MATCH trial (EAY131) is the definitive example — and understanding the distinction between "basket" and "umbrella" design clarifies exactly what happens once a patient is matched.
Both designs solve the same fundamental problem — that any single actionable mutation is too rare within any single tumor type to power a traditional randomized trial — but they cut across the disease-vs-mutation matrix in opposite directions:
Basket trials (one mutation, many tumor types): • A single targeted therapy is tested across multiple different cancer types, all unified by sharing one specific molecular alteration • Example: NCI-MATCH itself is fundamentally basket-structured at the arm level — each arm tests one targeted drug in patients with the matching mutation regardless of whether the primary tumor is lung, colon, breast, or a rare sarcoma • Rationale: if a mutation drives the same oncogenic pathway regardless of tissue of origin (tissue-agnostic biology), the drug should work similarly regardless of primary site — validated dramatically by tissue-agnostic FDA approvals like larotrectinib for NTRK fusions and pembrolizumab for MSI-high tumors
Umbrella trials (one tumor type, many mutation-defined arms): • A single cancer type (e.g., non-small-cell lung cancer) is the constant, and patients are sub-stratified into multiple parallel arms by their specific driver mutation, each arm testing a different targeted agent • Example: Lung-MAP (S1400) — a National Cancer Institute umbrella trial for previously treated advanced NSCLC with multiple biomarker-defined sub-studies running under one master protocol • Rationale: efficient shared infrastructure (screening, statistical design, regulatory oversight) for a single high-incidence tumor type with many known molecular subtypes
NCI-MATCH is often described as blending both: it screens broadly across tumor types (basket-like intake) and then routes each patient into one of >40 mutation-defined single-arm sub-studies (umbrella-like arm structure) — making it the template for tissue-agnostic precision oncology master protocols worldwide.
Once the ranking engine surfaces a basket/umbrella master protocol as the top match, enrollment proceeds through a defined pathway distinct from enrolling in a conventional single-arm trial:
1. Central IRB and master screening registration: the patient is registered under the master protocol's central institutional review board approval — a single consent and screening process covers eligibility for all arms simultaneously, rather than requiring separate consent per candidate arm
2. Molecular confirmation: many master protocols require confirmation of the driving alteration through a central, protocol-designated sequencing assay (rather than accepting outside lab results directly) to ensure uniform variant-calling quality across all enrolling sites
3. Arm assignment: the confirmed alteration is cross-referenced against the current roster of open arms — arms open and close dynamically over the life of a master protocol as some reach statistical futility and new arms are added for newly druggable targets, so the "menu" of available arms in year 5 looks very different from year 1
4. Slot availability check: each arm has a target enrollment (often ~35 patients per arm for NCI-MATCH-style single-arm phase II designs) — a molecularly eligible patient can still be turned away if their specific arm has already reached its enrollment cap and is statistically closed
5. Site-level activation: not every enrolling site has every arm open (drug supply and staffing constraints), so geographic feasibility (scored in Stage 4) determines whether the patient can actually access the assigned arm without excessive travel burden
Across the full NCI-MATCH program, roughly 6,000+ patients had their tumors centrally sequenced, but only about 2–10% of all screened patients ultimately enrolled in and received treatment on a molecularly assigned arm — a gap driven by the combination of rare mutation prevalence, arms reaching enrollment caps before a matched patient could be assigned, and clinical eligibility failures at Stage 3. This modest yield is the central quantitative lesson of a decade of basket-trial experience: broad molecular screening is necessary but the funnel from "sequenced" to "treated on a matched trial" remains narrow, motivating continued investment in larger, more dynamic master protocols with more arms and broader eligibility.