PD-L1 Testing in 2027: Assays, Cut-Offs and AI-Assisted Scoring

PD-L1 Testing in 2027: Assays, Cut-Offs and AI-Assisted Scoring

PD-L1 testing is still the most widely used predictive test in immuno-oncology, and still one of the most argued about. One protein is measured with four main immunohistochemistry (IHC) assays, scored in at least four different ways, and read against cut-offs that change with tumour type, line of therapy and drug. Since we first published this guide, the US label has tightened in upper gastrointestinal cancers, a new PD-L1-selected indication has arrived in ovarian cancer, and AI-assisted scoring has moved from research into some European routine practice.

This refreshed version covers the assays and scoring systems, the cut-offs on current FDA labels, the evidence on whether assays are interchangeable, why PD-L1 remains an imperfect predictor, and where MSI-H/dMMR fits alongside it. The history of how PD-L1 tests became companion or complementary diagnostics is in our separate guide to companion vs complementary diagnostics, so we do not repeat it here.

Key takeaways

• Four PD-L1 IHC assays carry FDA companion diagnostic claims today: 22C3 and 28-8 (Agilent/Dako) and SP142 and SP263 (Roche/Ventana). The 73-10 assay used in avelumab trials has no FDA claim.
• The score matters as much as the assay: TPS counts tumour cells, CPS adds immune cells, SP142 scores tumour cells (TC) and immune cells (IC) separately, and SP263 also supports a tumour area positivity (TAP) score.
• After a September 2024 FDA advisory committee vote, the FDA limited pembrolizumab and nivolumab in HER2-negative gastric, gastro-oesophageal junction (GEJ) and oesophageal cancers to PD-L1 positive tumours (letters dated 22–23 May 2025).
• The Blueprint project showed 22C3, 28-8 and SP263 stain tumour cells similarly in lung cancer, while SP142 stains fewer. That analytical similarity does not transfer a drug’s companion diagnostic claim.
• AI-assisted PD-L1 scoring is CE-IVD marked in Europe for some products, but we found no AI PD-L1 scoring algorithm on the FDA’s companion diagnostic list.
• MSI-H/dMMR is the strongest complementary IO biomarker and is tested independently of PD-L1.

What changed since our earlier version

• Assay count corrected. Our earlier version listed five FDA-approved assays including 73-10. Four assays hold FDA claims; 73-10 was a trial assay.
• New cut-off table. Added a table of PD-L1 cut-offs by tumour and drug, checked against current US prescribing information and the FDA list of authorised companion diagnostics.
• Upper GI label changes. Added the 2024 ODAC vote and the 2025 label restrictions for pembrolizumab and nivolumab, plus tislelizumab’s PD-L1 selected approvals.
• Ovarian cancer. Added the February 2026 pembrolizumab approval in platinum-resistant ovarian cancer at CPS ≥1, with 22C3 as the companion diagnostic.
• Interchangeability evidence. Added the Blueprint phase 1 and 2 findings.
• AI scoring. Replaced general claims with named products and their actual regulatory status.
• MSI-H/dMMR. Added a section on MSI-H/dMMR as the complementary IO biomarker.
• Removed unverified claims, including on exosomal PD-L1 and on the share of approvals that use PD-L1.

What PD-L1 is and why it is measured

PD-L1 (programmed death-ligand 1) is expressed on tumour cells and on immune cells such as macrophages and lymphocytes. When it binds PD-1 on activated T cells, it dampens their activity, and tumours use this pathway to escape immune attack. Anti-PD-1 and anti-PD-L1 antibodies block that interaction.

The reasoning behind PD-L1 testing is simple: tumours that express the ligand are more likely to depend on the pathway, so they are more likely to respond when it is blocked. In practice, higher PD-L1 expression is linked to greater benefit in several tumour types, but some PD-L1 negative tumours respond and many PD-L1 high tumours do not. That is why the cut-off, the scoring method and the clinical setting all matter.

The four assays and their scoring systems

Each assay was developed with a specific drug programme, on a specific staining platform, with its own scoring algorithm.

TPS vs CPS

Tumour Proportion Score (TPS) is the percentage of viable tumour cells with partial or complete membrane staining. It ignores immune cells and is the standard in non-small cell lung cancer (NSCLC).

Combined Positive Score (CPS) is the number of PD-L1 staining cells (tumour cells, lymphocytes and macrophages) divided by the total number of viable tumour cells, multiplied by 100. It is capped at 100. Because it counts immune cells, CPS is used in tumours where immune-cell PD-L1 carries much of the signal, such as gastric, oesophageal, head and neck, cervical and triple-negative breast cancer.

The TPS vs CPS distinction trips up many labs: the same 22C3 slide can be TPS 0% and CPS 5 if the PD-L1 staining is all on immune cells. A request form that names the drug and indication, not just “PD-L1”, avoids reporting the wrong score.

TC/IC and TAP

SP142 scores tumour cells (TC) and tumour-infiltrating immune cells (IC, as a percentage of tumour area) separately. SP263 is scored on tumour cells in NSCLC, and in the tislelizumab gastric and oesophageal trials it was read as tumour area positivity (TAP): the share of tumour area covered by stained tumour cells and tumour-associated immune cells.

PD-L1 cut-offs by tumour and drug (current US labels)

The table shows PD-L1 selected indications on current US labels. It is not a full list of each drug’s indications, many of which do not require PD-L1 testing (for example, pembrolizumab with chemotherapy in first-line NSCLC). Always check the current prescribing information and the FDA companion diagnostic list.

Sources: US prescribing information for Keytruda (revised 07/2026), Opdivo (08/2026), Tecentriq (10/2026) and Libtayo (08/2026); the FDA list of authorised companion diagnostics; BeiGene’s approval announcements for Tevimbra.

The upper GI label changes: what happened and when

Pembrolizumab and nivolumab were first approved with chemotherapy in first-line HER2-negative gastric, GEJ and oesophageal cancers without a PD-L1 requirement. Later analyses showed that most of the overall survival benefit sat in PD-L1 positive patients.

• 7 November 2023: the FDA restricted pembrolizumab plus trastuzumab and chemotherapy in HER2-positive gastric/GEJ adenocarcinoma to tumours with CPS ≥1, based on KEYNOTE-811, where patients with CPS <1 had no PFS or OS improvement.
• 26 September 2024: the FDA’s Oncologic Drugs Advisory Committee (ODAC) voted 10 to 2, with one abstention, that the benefit-risk of first-line PD-1 inhibitors was not favourable in HER2-negative, microsatellite-stable gastric/GEJ adenocarcinoma with PD-L1 expression below 1. In oesophageal squamous cell carcinoma the vote was 11 to 1, with one abstention, at the same threshold. The discussion covered pembrolizumab, nivolumab and tislelizumab.
• December 2024 and March 2025: tislelizumab was approved in first-line HER2-negative gastric/GEJ adenocarcinoma (27 December 2024) and first-line oesophageal SCC (4 March 2025), in both cases only for PD-L1 ≥1.
• 22–23 May 2025: FDA letters for pembrolizumab and nivolumab limited their HER2-negative gastric/GEJ and oesophageal indications to PD-L1 positive tumours (CPS ≥1 for pembrolizumab; PD-L1 ≥1 for nivolumab, including nivolumab plus ipilimumab in oesophageal SCC).
• 2026: the FDA companion diagnostic list now shows 22C3 for pembrolizumab in oesophageal/GEJ carcinoma at CPS ≥1 and 28-8 for nivolumab across oesophageal SCC and gastric, GEJ and oesophageal adenocarcinoma at CPS ≥1.

For diagnostics teams, the point is that a cut-off can move after approval, and when it does, the drug changes from “no test needed” to “test required”. That shift creates demand for testing overnight. It also exposes any differences between assays, because three drugs now share a threshold of ≥1 in the same tumours but were developed with three different assays and scoring methods.

Are PD-L1 assays interchangeable?

The best-known evidence is the Blueprint PD-L1 IHC Assay Comparison Project, run by pharma, diagnostics companies and the IASLC.

• Phase 1 (Hirsch et al., J Thorac Oncol 2017) stained 39 NSCLC tumours with 22C3, 28-8, SP142 and SP263. Three assays (22C3, 28-8 and SP263) gave similar tumour cell staining; SP142 stained fewer tumour cells, and immune cell staining varied more than tumour cell staining. Even so, 14 of 38 cases (37%) would have been classified differently depending on which assay and cut-off was used.
• Phase 2 (Tsao et al., J Thorac Oncol 2018) used 81 real-world lung cancer samples and all five trial assays, including 73-10. It confirmed the pattern (22C3, 28-8 and SP263 highly comparable, SP142 less sensitive, 73-10 more sensitive on tumour cells). Agreement between pathologists was very strong for tumour cell scoring (intraclass correlation 0.86–0.93) but poor for immune cell scoring (0.18–0.19). Glass slide and digital image scores were highly concordant (Pearson correlation above 0.96).

The phase 2 authors concluded that the data support the analytical interchangeability of 22C3, 28-8 and SP263 for tumour cell scoring. Two caveats limit how far this goes. Blueprint was done in lung cancer using TPS-type tumour cell scoring; it says little about CPS in gastric or head and neck cancer, where immune cells count. And analytical concordance is not the same as clinical validation. Only the assay and cut-off validated in the pivotal trial carries the label claim. Labs that use a laboratory-developed test or a different platform need their own bridging and validation data, a point covered in our guide to digital pathology biomarker validation.

Why PD-L1 is an imperfect predictor

Biology

PD-L1 expression is heterogeneous across a tumour and between primary and metastatic sites, and it changes over time and with treatment, including chemotherapy and radiotherapy. A single core biopsy samples a small area at one moment. Expression can also be driven by interferon-gamma from infiltrating T cells (an “inflamed” tumour) or by oncogenic signalling inside the tumour cell. These mean different things for response, but they look the same on an IHC slide.

Measurement

Pre-analytical factors (fixation time, tissue age, decalcification) affect staining. Scoring near low cut-offs such as CPS 1 or TC 1% is where pathologists disagree most, and immune cell scoring is less reproducible than tumour cell scoring. A 1% threshold can move a patient across a treatment boundary because of a handful of cells.

Clinical design

Many trials enrolled all patients and analysed PD-L1 subgroups afterwards, sometimes using different assays and scoring rules between competing drugs. That makes cross-trial comparison of “PD-L1 positive” populations unreliable. PD-L1 negative patients still respond in some settings, which is why many first-line combinations, such as pembrolizumab with chemotherapy in NSCLC, carry no PD-L1 requirement.

For these reasons PD-L1 is increasingly read alongside other markers: MSI-H/dMMR, tumour mutational burden (covered in our TMB guide), tumour-infiltrating lymphocytes, and spatial and gene expression signatures, none of which has yet replaced it.

AI-assisted PD-L1 scoring: where it stands

AI-assisted scoring targets the parts of PD-L1 testing that are hardest for people: counting cells close to low thresholds, and scoring immune cells consistently. What is actually available for clinical use is narrower than the research literature suggests.

We found no AI PD-L1 scoring algorithm on the FDA’s companion diagnostic list. In the US, AI PD-L1 tools are used in research, in trials and as lab-validated workflows, not as FDA-authorised companion diagnostics. For a wider view of what evidence AI biomarker tools need, see our article on AI in biomarker discovery.

The practical questions for sponsors and labs are the same as for any digital pathology tool: scanner and stain validation, locked versus adaptive algorithms, how pathologist overrides are handled, and whether the algorithm was trained on the same assay and scoring method that the label requires.

MSI-H and dMMR: the complementary IO biomarker

Microsatellite instability-high (MSI-H) and mismatch repair deficiency (dMMR) describe tumours that cannot repair DNA replication errors. They build up very high numbers of mutations and neoantigens, which makes them among the most immunotherapy-responsive tumours. Unlike PD-L1, MSI-H/dMMR is a genomic or protein-loss result that is close to binary, so it suffers less from the cut-off problems above.

• Pembrolizumab became the first tissue-agnostic cancer approval in 2017 for MSI-H/dMMR solid tumours. Its current label covers previously treated MSI-H/dMMR solid tumours and MSI-H/dMMR colorectal cancer.
• Nivolumab plus ipilimumab was approved on 8 April 2025 for unresectable or metastatic MSI-H/dMMR colorectal cancer, based on CheckMate 8HW.
• Atezolizumab’s label now includes adjuvant treatment, with chemotherapy, of stage III dMMR colon cancer (label updated October 2026).
• Companion tests include the VENTANA MMR RxDx Panel (IHC for MLH1, PMS2, MSH2 and MSH6) and NGS-based tests such as FoundationOne CDx and Caris MI Cancer Seek.

In practice, MSI/MMR and PD-L1 are tested in parallel, not either-or. In gastric and oesophageal adenocarcinoma, the 2024 ODAC question was specifically about microsatellite-stable disease, because MSI-H tumours tend to respond whatever their PD-L1 score.

What this means for diagnostics and pharma teams

• Plan for cut-off changes after approval. The upper GI story shows that an “all-comers” label can become biomarker-restricted years later. Prospective, assay-locked PD-L1 data protect a programme.
• Specify the score, not just the assay. TPS, CPS, TC/IC and TAP are not interchangeable even on the same slide.
• Build bridging data early if your test may be used in place of the trial assay, and expect regulators to ask for it.
• Treat AI scoring as a regulated device where it drives a treatment decision, with the same validation burden as the assay. Our FDA companion diagnostic approval guide sets out the pathway.

Discuss PD-L1 and IO biomarker strategy in London

These questions sit at the centre of Stream 2 at ImmunoMark Summit London, our biomarker and companion diagnostics conference in London, on 3–4 March 2027. The 2nd annual UK edition targets 400+ attendees across two parallel streams: Stream 1: Next-Gen Immuno-Oncology and Stream 2: Biomarkers & Companion Diagnostics. Stream 2 sessions include “Keynote – Emerging Predictive & Prognostic Biomarkers in IO” and “Digital Pathology & Spatial Transcriptomics in Oncology”, alongside keynotes, panels, case studies, pre-arranged 1:1 meetings and a Day 1 drinks reception.

The 2026 London edition drew 300+ attendees from 100+ companies, 70% of them senior decision-makers, with companies such as GSK, Roche, Merck, AstraZeneca, J&J and BMS taking part.

Register for ImmunoMark London 2027

If your company wants to meet this audience, whether you supply IHC assays, digital pathology and AI scoring, or companion diagnostic development services, see the London 2027 sponsorship opportunities or our page on the companion diagnostics conference in London.

FAQs

What is the difference between TPS and CPS in PD-L1 testing?

Tumour Proportion Score (TPS) is the percentage of viable tumour cells with membrane PD-L1 staining and ignores immune cells. Combined Positive Score (CPS) counts PD-L1 staining tumour cells, lymphocytes and macrophages, divides by the number of viable tumour cells and multiplies by 100, capped at 100. TPS is used for pembrolizumab in NSCLC; CPS is used in gastric, oesophageal, head and neck, cervical, triple-negative breast and ovarian cancer.

Which PD-L1 assays are FDA-approved companion diagnostics?

Four: PD-L1 IHC 22C3 pharmDx and PD-L1 IHC 28-8 pharmDx from Agilent (Dako), and VENTANA PD-L1 (SP142) and VENTANA PD-L1 (SP263) from Roche. Each is linked to specific drugs, tumours and cut-offs on the FDA’s companion diagnostic list. The 73-10 assay was used in avelumab trials but has no FDA claim.

What is the PD-L1 cut-off for immunotherapy in gastric and oesophageal cancer?

On current US labels, pembrolizumab with chemotherapy in first-line HER2-negative and HER2-positive gastric/GEJ adenocarcinoma, and in oesophageal/GEJ carcinoma, requires CPS ≥1. Nivolumab in gastric, GEJ and oesophageal adenocarcinoma and in first-line oesophageal squamous cell carcinoma requires PD-L1 ≥1, and tislelizumab is approved for PD-L1 ≥1. Pembrolizumab monotherapy in previously treated oesophageal squamous cell carcinoma requires CPS ≥10.

Are 22C3, 28-8, SP263 and SP142 interchangeable?

Not for label purposes. The Blueprint project found that 22C3, 28-8 and SP263 give similar tumour cell staining in lung cancer, while SP142 stains fewer tumour cells, and that immune cell scoring is much less reproducible. Only the assay and scoring method validated for a drug carry its companion diagnostic claim, so any substitute needs local validation and bridging data.

Is AI PD-L1 scoring approved?

Some products are CE-IVD marked in Europe, including Roche’s uPath PD-L1 (SP263) image analysis algorithm for NSCLC and Aiforia’s PD-L1 lung cancer model. We found no AI PD-L1 scoring algorithm on the FDA’s companion diagnostic list, so in the US these tools are used in research, trials and laboratory-validated workflows.