Global Training Program Achieves High Agreement in PD-L1 Scoring for Lung Cancer

…

…

The clinical significance of the tumor proportion score rests on its role as a gatekeeper for immunotherapy decisions, which is why the reproducibility data from this training program carry such weight. When pembrolizumab was first approved, the companion diagnostic test established a binary decision framework: patients whose tumors expressed PD-L1 on at least half of their tumor cells were candidates for first-line monotherapy, while those falling below that threshold faced different treatment pathways. A scoring error that crosses either the 1% or the 50% threshold therefore has direct therapeutic consequences, potentially exposing patients to a costly therapy with limited expected benefit or, conversely, withholding a treatment from which they might have derived substantial survival advantage. This is the practical backdrop against which the agreement figures reported in the study must be understood.

The asymmetry between the agreement metrics at the two cutoffs deserves particular attention. At the 1% threshold, positive percent agreement reached 97.6%, meaning that when the expert consensus panel designated a case as positive, participating pathologists overwhelmingly concurred. Negative percent agreement was lower at 85.6%, indicating that some cases scored below 1% by the reference panel were occasionally read as positive by trainees. At the 50% cutoff, the pattern inverted: negative percent agreement rose to 91.9% while positive percent agreement fell to 81.0%. In other words, pathologists were more reliable at recognizing truly negative tumors at the higher threshold but less consistent at confirming strong positivity. This asymmetry is consistent with the biology of PD-L1 expression, where heterogeneous and patchy membrane staining can make it difficult to decide whether a tumor has crossed the halfway mark. Cases sitting near a cutoff are inherently the hardest to classify, and the study deliberately enriched its slide sets around both thresholds precisely to stress-test readers at these decision boundaries rather than presenting only unambiguous specimens.

The distinction between overall percent agreement and its positive and negative components also illustrates a broader principle in diagnostic test evaluation. A high overall agreement figure can mask divergent performance in the two directions of misclassification. A reader who tends to overcall positivity will inflate negative percent agreement while depressing positive percent agreement, yet the overall figure may still appear reassuring. By decomposing agreement into these components, the study offers a more honest picture of where scoring errors concentrate. For clinical practice, the direction of the error matters: false-positive calls at the 1% cutoff could lead to pembrolizumab being offered to patients with negligible target expression, whereas false-negative calls at the 50% cutoff could steer high-expressing patients toward chemotherapy instead of the regimen associated with superior outcomes in that subgroup.

The training methodology itself reflects lessons accumulated over more than a decade of biomarker standardization work. The same organization had previously run formal pathologist training programs for HER2 scoring in breast and gastric cancer, an effort that helped establish reproducible HER2 assessment as a prerequisite for targeted therapy. That institutional experience informed the structure of the PD-L1 program: a small cadre of pathologists received direct training from the assay manufacturer, those individuals then trained additional colleagues, and the certified trainers subsequently delivered standardized sessions worldwide. This cascade model, sometimes described as a train-the-trainer approach, allows a consistent interpretive framework to propagate across geographically dispersed audiences while retaining fidelity to the manufacturer’s approved interpretation manual. The two-year enrollment window spanning 2016 and 2017 coincided with the period when PD-L1 testing was transitioning from a research concept to a routine clinical requirement, making timely standardization particularly consequential.

The composition of the training material was carefully constructed. Fifty-four NSCLC resection specimens, comprising both adenocarcinomas and squamous cell carcinomas, were stained with the 22C3 assay under controlled conditions. A panel of three expert pathologists independently evaluated each case and reached consensus on the reference tumor proportion score, providing a stable gold standard against which trainee performance could be measured. Fourteen cases served for demonstration, ten were divided into two sets of five for pretesting, and thirty formed the main test. The inclusion of a pretest allowed instructors to gauge baseline competency before formal instruction, while the main test measured post-training performance. Specimens were required to contain at least 100 viable tumor cells, a minimum cellularity criterion that guards against scoring instability in sparse samples. The scoring rules themselves are exacting: only viable tumor cells with partial or complete membranous staining at any intensity count toward the numerator, while cytoplasmic staining, stained infiltrating immune cells, normal parenchymal cells, and necrotic tissue are all excluded. Each of these exclusions addresses a known pitfall, since macrophages and other immune cells frequently show PD-L1 staining in lung tumors and can be mistaken for tumor cells by inexperienced observers.

The dual delivery format, combining traditional glass-slide microscopy with whole-slide digital imaging, merits consideration in light of how pathology practice has evolved. Digital slides were scanned at forty-fold magnification using commercial scanners and viewed with standard image analysis software, while hematoxylin and eosin and negative control sections were scanned at twenty-fold magnification. Whole-slide imaging had by this period matured into a validated medium for primary diagnosis in several settings, and its use in training offered obvious logistical advantages: a single set of expertly stained and consensus-scored cases could be distributed to audiences on multiple continents without the degradation, loss, or scheduling constraints associated with shipping physical slides. The authors note that the digital approach also opens the door to computer-assisted scoring systems, foreshadowing the image analysis tools that have since become increasingly prominent in biomarker pathology. Remote delivery additionally reduced the carbon and financial footprint of international training and enabled participation by pathologists in regions without easy access to in-person expert-led sessions.

The scale and diversity of participation strengthen the generalizability of the findings. Seven hundred fifty-one pathologists from sixty-three countries represent an unusually broad sampling of professional backgrounds, training traditions, and levels of prior experience with PD-L1 interpretation. Inter-reader variability in biomarker scoring has historically been one of the most persistent obstacles to companion diagnostic implementation, with early HER2 and hormone receptor programs documenting substantial disagreement rates before structured training and standardized assays were introduced. Against that backdrop, overall inter-reader agreement of 95.6% at the 1% cutoff and 87.3% at the 50% cutoff, achieved across such a heterogeneous cohort, represents a favorable outcome. Intra-reader agreement, which measures each participant’s consistency with their own earlier readings, was similarly high at 95.9% and 91.4% respectively, suggesting that the training produced not just momentary alignment with the reference standard but a stable internalized scoring practice.

It is worth situating these results within the broader ecosystem of PD-L1 assays and their clinical use. The 22C3 pharmDx assay was the first PD-L1 immunohistochemistry test to receive regulatory approval as a companion diagnostic for anti-PD-1 therapy in NSCLC, and the tumor proportion score it defines applies specifically to tumor-cell membrane staining. Other PD-L1 antibodies and scoring systems exist, each with its own cell types, scoring compartments, and cutoff values, and results are generally not interchangeable between platforms. The training program described here therefore addresses one specific but widely used testing paradigm, and the interpretive discipline it instills, careful exclusion of non-tumor elements, attention to partial membrane staining, and rigorous application of percentage thresholds, exemplifies the kind of structured approach that any biomarker scoring system requires.

The finding that agreement was harder to achieve at the 50% cutoff than at the 1% cutoff has practical implications for laboratory quality assurance. Laboratories might reasonably direct additional proficiency attention toward cases in the intermediate-to-high expression range, where heterogeneity of staining distribution is greatest, and toward borderline cases immediately adjacent to either threshold. Periodic revalidation, reference-case review, and participation in external quality assessment schemes can help sustain the performance levels documented immediately after training. The study’s demonstration that structured, expert-led instruction can bring a large international cohort to high concordance within one or two days of training provides an encouraging template for future biomarker rollouts, particularly as the repertoire of predictive immunohistochemical markers in oncology continues to expand and as digital pathology platforms make standardized remote education increasingly feasible.

Subject of Research: …

Article Title: Training on PD-L1 scoring in non-small cell lung cancer with high intra- and inter-reader agreement: results of a worldwide microscopic/digital image-based training of 751 pathologists

Article References: Bänfer, G., Diezko, R., Oberthür, R., Kumar, G. L., Rüschoff, J., Schildhaus, H.-U., & Jasani, B. (2026). Training on PD-L1 scoring in non-small cell lung cancer with high intra- and inter-reader agreement: results of a worldwide microscopic/digital image-based training of 751 pathologists. British Journal of Biomedical Science, 83, Article 16477. https://doi.org/10.3389/bjbs.2026.16477

Image Credits: AI Generated

DOI: 10.3389/bjbs.2026.16477

Keywords: Training, PD-L1, scoring, non-small, cell, lung, cancer, high, intra-, inter-reader, agreement, results