Tumor mutational burden, MSI status, and PD-L1 CPS as complementary biomarkers gating tissue-agnostic immune checkpoint inhibitor therapy
Tumor mutational burden begins with a straightforward but technically demanding count: how many somatic mutations does a tumor genome carry? Next-generation sequencing (NGS) panels — ranging from ~300-gene hybrid-capture panels to whole-exome sequencing (WES) — read tumor DNA (and ideally matched normal DNA) to base-level resolution, then a bioinformatics pipeline separates true tumor-acquired mutations from inherited germline variants and sequencing artifacts.
Turning a tumor biopsy into a mutation count is a multi-stage pipeline, each step designed to strip away noise that would otherwise inflate or deflate the final tally:
• DNA extraction & library prep: tumor tissue (fresh-frozen or FFPE) and, ideally, a matched normal sample (blood or adjacent normal tissue) are extracted and fragmented into sequencing libraries • Hybrid-capture or amplicon enrichment: probes pull down the genomic regions of interest — a defined gene panel or the full exome — before sequencing, concentrating reads on informative territory • Alignment: reads are mapped to the reference genome (GRCh38) with tools like BWA-MEM; duplicate reads from PCR amplification are marked and excluded • Variant calling: somatic callers (Mutect2, Strelka2, VarDict) compare tumor reads against matched normal (or against a population database when no matched normal exists — "tumor-only" mode) to flag positions where the tumor allele frequency significantly exceeds background/germline • Germline subtraction: any variant present in the normal sample, or common in population databases (gnomAD allele frequency >0.1–1%), is excluded — TMB counts only tumor-acquired, non-inherited mutations • Filtering for biological relevance: by convention, TMB counts non-synonymous coding mutations (missense, nonsense, indels) in the panel footprint; synonymous variants are usually excluded since they don't require the same detection sensitivity threshold and add noise
Tumor-only sequencing (no matched normal) is common in clinical practice but introduces a systematic risk: rare inherited variants not yet catalogued in population databases can be misclassified as somatic, artificially inflating the reported TMB — a major source of panel-to-panel discordance.
Not all "mutation counts" are measured the same way, and the sequencing footprint used to generate them matters enormously:
• Targeted hybrid-capture panels (FoundationOne CDx, MSK-IMPACT, TruSight Oncology 500): cover 300–500 cancer-relevant genes across ~1.1–1.5 Mb. These are the FDA-cleared, clinically deployed assays for TMB reporting, chosen because they are fast, affordable (~$3,000–6,000), and turnaround in 1–2 weeks • Whole-exome sequencing (WES): covers all ~20,000 protein-coding genes (~35–50 Mb). WES was the original substrate for TCGA-era TMB research and remains the reference standard against which panels are calibrated, but it is slower and costlier for routine clinical use • Whole-genome sequencing (WGS): covers non-coding regions too, but coding-region TMB from WGS correlates closely with WES TMB and is not the clinical standard
Because a 1.1 Mb panel samples only a small, gene-biased fraction of the exome, the raw mutation count from a panel must be extrapolated (mutations detected ÷ panel Mb) to estimate genome/exome-wide mutation density — an approximation, not a direct measurement, that becomes noisier as panel size shrinks.
Tumor mutational burden is defined simply: the total number of qualifying somatic mutations divided by the number of megabases of genome interrogated. TMB = mutation count ÷ panel size (Mb). This normalization is what makes TMB comparable across tumor types and, in principle, across sequencing platforms — but the simplicity of the formula conceals substantial methodological variability that has taken the field a decade to partially standardize.
TMB (mut/Mb) = (number of somatic, non-synonymous coding mutations counted) ÷ (megabases of genomic territory sequenced with adequate coverage)
On its face this is trivial division. But two tumors with an identical absolute number of mutations can report wildly different TMB values depending on the denominator:
• A tumor with 100 mutations detected on a 1 Mb panel → TMB = 100 mut/Mb • The same 100 mutations "diluted" across a 40 Mb whole-exome footprint → TMB = 2.5 mut/Mb
In reality the numerator also changes with panel size — a small panel captures a biased, gene-enriched subset of the mutation landscape (cancer genes are more frequently mutated than average genomic territory), so panel-based TMB estimates are not a simple linear extrapolation of exome TMB. Larger panels and WES sample a more representative cross-section of the genome and are considered more statistically robust, especially at low mutation counts where a handful of extra or missed calls swings the ratio substantially.
This simulator's two sliders make the normalization pitfall directly visible: holding the mutation count fixed while shrinking the panel size mechanically inflates the reported TMB — the same biological reality can cross the TMB-H ≥10 mut/Mb decision threshold purely as an artifact of assay design, not tumor biology.
Recognizing that different commercial and academic panels produced discordant TMB values for the same tumor, the Friends of Cancer Research (FoCR) TMB Harmonization Project (2017–2020) convened >20 sequencing labs and companies to compare panel-derived TMB against a WES reference standard across shared tumor samples.
Key findings and recommendations:
• Panels of ≥1.1 Mb footprint with ≥0.8–1× WES correlation are considered adequate for reliable TMB estimation; smaller/older panels (<0.5 Mb) show poor reproducibility, especially near the clinical decision threshold • Gene content matters: panels should include a representative mix of driver and passenger-adjacent genes, not be overly enriched for known hotspot mutations (hotspot-biased panels systematically distort TMB) • Bioinformatic pipeline choices (variant caller, filtering thresholds, synonymous vs non-synonymous counting, minor allele frequency cutoffs) independently shift reported TMB by up to 2-fold between labs using the identical raw sequencing data • FDA now requires TMB assays seeking companion-diagnostic status to demonstrate concordance with the reference method used in the pivotal KEYNOTE-158 trial (FoundationOne CDx, calibrated to WES)
Despite harmonization progress, clinicians are advised to interpret a borderline TMB value (e.g., 8–12 mut/Mb) with caution, ideally alongside MSI and PD-L1 results, rather than as a single hard cutoff.
Microsatellites are short, repetitive DNA sequences (mono-, di-, tri-nucleotide repeats) scattered throughout the genome that are especially prone to replication slippage errors. In cells with a fully functional DNA mismatch repair (MMR) system, these slippage errors are corrected. When MMR is defective — deficient mismatch repair, dMMR — errors accumulate at microsatellite loci, producing a measurable shift in repeat length called microsatellite instability (MSI-H), and, because MMR failure affects the whole genome, an extremely high overall mutation burden.
The MMR system (proteins MLH1, MSH2, MSH6, PMS2) proofreads newly replicated DNA, recognizing and excising base-pair mismatches and insertion/deletion loops that DNA polymerase misses — particularly common at repetitive microsatellite sequences where the polymerase can "slip" and add or drop repeat units.
When an MMR gene is inactivated — via germline mutation (Lynch syndrome), somatic mutation, or MLH1 promoter hypermethylation (the common sporadic mechanism, often colorectal or endometrial) — microsatellite loci across the genome accumulate length errors unchecked. This produces two clinically detectable signatures:
• MSI testing (PCR-based, Bethesda panel of 5 mononucleotide/dinucleotide markers, or newer NGS-based MSIsensor algorithms): compares repeat length distributions in tumor DNA against matched normal DNA at each microsatellite locus. MSI-High = instability at ≥2 of 5 (or ≥30–40% of NGS-assessed loci); MSI-Low = instability at 1 marker; MSS (microsatellite stable) = no instability • IHC for MMR proteins (dMMR testing): loss of nuclear staining for any of the 4 MMR proteins on immunohistochemistry directly indicates the deficient protein; MLH1 loss often pairs with PMS2 loss (heterodimer), MSH2 loss with MSH6 loss
MSI-H and dMMR are highly concordant (~90-95%) and used somewhat interchangeably as eligibility criteria, though they test different molecular layers (DNA repeat length vs. protein expression) and occasionally disagree.
A dMMR/MSI-H tumor typically carries 10–100× more somatic mutations than an MSS tumor of the same type, because the loss of proofreading allows replication errors to accumulate genome-wide with every cell division, not just at microsatellites. This mutational avalanche has a direct immunological consequence: many of those mutations, especially frameshift indels at coding microsatellites, generate novel, immunogenic frameshift peptides — neoantigens the immune system has never seen and can potentially recognize as foreign.
This is the mechanistic link between dMMR/MSI-H and checkpoint inhibitor response: a tumor riddled with neoantigens attracts a pre-existing cytotoxic T-cell infiltrate that is held in check by PD-1/PD-L1-mediated immune exhaustion. Blocking PD-1 (pembrolizumab, nivolumab) releases that brake, allowing already-primed T-cells to attack. This is why MSI-H tumors — regardless of tissue of origin — respond to checkpoint blockade at strikingly high rates (~30-50% objective response), motivating the first-ever tissue-agnostic FDA approval (pembrolizumab, 2017, KEYNOTE-164/158) years before the TMB-H tissue-agnostic approval that followed the same logic with a broader, continuous biomarker.
MSI-H and TMB-H are biologically overlapping but not identical gates: essentially all MSI-H tumors are also TMB-H (hypermutation is the shared mechanism), but the reverse is not true — many TMB-H tumors (e.g., UV-mutagenized melanoma, smoking-associated NSCLC) are MSS, driven by exogenous mutagens rather than defective repair.
While TMB and MSI status measure the tumor genome, PD-L1 immunohistochemistry (IHC) measures a functional immune-evasion protein directly on the tissue — the ligand that engages the PD-1 receptor on T-cells to suppress their activity. PD-L1 expression is scored by a pathologist reading a stained tissue section, and different scoring systems (TPS, CPS, IC) are calibrated to different drugs and tumor types, making it the most assay- and context-dependent of the three major immunotherapy biomarkers.
PD-L1 IHC scoring is not one single number — three distinct scoring conventions are used across different drug labels and tumor types, and confusing them is a common clinical error:
• Tumor Proportion Score (TPS): the percentage of viable tumor cells showing partial or complete membrane staining, out of all viable tumor cells — used primarily in NSCLC for pembrolizumab monotherapy eligibility (TPS ≥50% for 1L monotherapy, ≥1% for combination therapy) • Combined Positive Score (CPS): the number of PD-L1-staining cells — tumor cells, lymphocytes, AND macrophages — divided by the total number of viable tumor cells, multiplied by 100. Because the numerator includes immune cells while the denominator counts only tumor cells, CPS can mathematically exceed 100 in immune-cell-rich tumors. CPS is used in cervical, gastric/GEJ, esophageal, head & neck, and triple-negative breast cancer indications • Immune Cell score (IC): the percentage of tumor area occupied by PD-L1-staining immune cells — used for atezolizumab in some indications
A pathologist manually counts staining cells across representative tumor fields under the microscope (or with digital image analysis assistance), a process with meaningful inter-observer variability (~5-15% discordance rate at borderline cutoffs between trained readers).
Unlike TMB (a relatively stable genomic property once measured) and MSI status (essentially binary and stable), PD-L1 expression is a dynamic transcriptional/translational readout that varies with:
• Intratumoral heterogeneity: PD-L1 expression can differ substantially between regions of the same tumor and between primary tumor and metastases — a single biopsy may not represent the whole disease burden • Temporal variability: PD-L1 expression can be upregulated by prior therapy (including chemotherapy and radiation, which increase interferon-gamma signaling) or downregulated over time — a biopsy from 6 months ago may not reflect current status • Assay-to-assay variability: different PD-L1 antibody clones (22C3, 28-8, SP142, SP263) used across companion diagnostics for different drugs show only moderate concordance, and each clone is only formally validated for its paired drug
Because of this imperfection, PD-L1 alone is an incomplete predictor: many PD-L1-negative tumors still respond to checkpoint blockade, and many PD-L1-high tumors do not respond — response rates rise with CPS but never approach 100% or fall to 0% at any cutoff. This is precisely why PD-L1 is used as one axis among several (alongside TMB and MSI) rather than a sole gatekeeper for most indications.
CPS ≥1 is a low bar deliberately — regulatory approvals at this threshold (e.g., cervical cancer, gastric cancer) prioritize sensitivity (not missing responders) over specificity, accepting that many CPS-low patients who receive therapy will not respond, in exchange for not excluding patients who would.
| Product | Indication | Trial Design | Key Result |
|---|---|---|---|
| TMB (mut/Mb) | Genomic — NGS panel/WES | Counts somatic non-synonymous mutations, normalizes by sequenced Mb | Tissue-agnostic threshold (≥10), captures mutagen + repair-driven burden |
| MSI/dMMR status | Genomic + protein — PCR / IHC | Detects mismatch repair failure via microsatellite length shift or protein loss | Near-binary, highly reproducible, strong mechanistic link to neoantigen load |
| PD-L1 CPS/TPS | Protein — immunohistochemistry | Pathologist-scored staining of PD-L1 on tumor + immune cells | Directly measures the drug target; captures dynamic immune-evasion state |
| Combined use | All three, non-redundant | Each axis catches responders the others miss | Maximizes sensitivity for identifying checkpoint-inhibitor candidates |
On June 16, 2020, the FDA granted accelerated approval to pembrolizumab for adult and pediatric patients with unresectable or metastatic solid tumors with TMB-H (≥10 mutations/megabase), as determined by an FDA-approved test, that have progressed following prior treatment and have no satisfactory alternative treatment options — the second-ever tissue-agnostic oncology approval, and the first based on a purely quantitative genomic biomarker rather than a specific gene alteration.
KEYNOTE-158 was a multi-cohort, open-label phase II basket trial testing pembrolizumab monotherapy across ten advanced, previously-treated solid tumor types that individually lacked strong checkpoint-inhibitor data (including anal, biliary, cervical, endometrial, mesothelioma, neuroendocrine, salivary, small-cell lung, thyroid, and vulvar cancers).
Within this trial, TMB was retrospectively assessed using the FoundationOne CDx assay on 790 patients with sufficient tissue, and patients were stratified into TMB-H (≥10 mut/Mb, n=102) versus non-TMB-H (n=688) groups:
• TMB-H group: objective response rate (ORR) 29% (30/102), including 4 complete responses • Non-TMB-H group: ORR 6% (43/688) • Median duration of response: not reached in the TMB-H group at data cutoff, with the majority of responses ongoing beyond 12 months • The TMB-H effect held up across most represented tumor types within the cohort, though sample sizes per tumor type were often small
This ~5-fold difference in response rate between TMB-H and non-TMB-H patients, replicated in a purely biomarker-defined (not tissue-defined) population, was the evidentiary basis for the FDA's accelerated approval — explicitly divorced from organ of origin, mirroring the logic that had already succeeded for MSI-H/dMMR in 2017.
The accelerated approval carried a post-marketing requirement: subsequent real-world and confirmatory data (including analyses questioning whether TMB-H benefit is as uniform across all tumor types as initially estimated) have kept TMB-H interpretation an active area of clinical and regulatory scrutiny, illustrating that tissue-agnostic biomarkers still interact with tumor biology in tissue-specific ways.
In clinical practice, a patient's NGS report typically returns all three biomarkers simultaneously, and oncologists reason through them as a converging, not mutually exclusive, decision tree:
1. Is the tumor MSI-H/dMMR? If yes, pembrolizumab is FDA-approved tissue-agnostically (2017 approval) regardless of TMB or PD-L1 — this is usually the most mechanistically robust single gate, especially relevant in colorectal and endometrial cancer where MSI testing is standard of care
2. If MSS, is TMB ≥10 mut/Mb? If yes, the 2020 tissue-agnostic pembrolizumab approval applies for previously-treated patients with no satisfactory alternative — this gate captures hypermutated-but-repair-proficient tumors (UV-driven melanoma, smoking-driven NSCLC) that MSI testing alone would miss
3. If TMB-low and MSS, is PD-L1 CPS/TPS above the relevant tumor-type-specific cutoff? Many first-line indications (NSCLC, gastric, cervical, head & neck) grant eligibility for specific drug regimens based on PD-L1 alone, independent of TMB/MSI status
4. Combination reasoning: a small but clinically important subset of patients are positive on multiple axes (e.g., TMB-H AND PD-L1-high), which some data suggest predicts an even higher response probability than either alone — though formal multi-biomarker composite scores are not yet standard of care
Patients meeting none of the three gates may still be considered for checkpoint inhibitors in specific approved combination regimens or clinical trials, but lack a biomarker-driven monotherapy rationale — for these patients, chemotherapy, targeted therapy against an identified driver mutation, or trial enrollment remains the standard pathway.