PD-L1 as an immunotherapy biomarker: PD-L1 (programmed death-ligand 1) expression, measured by immunohistochemistry (IHC) staining of tumor tissue, is the primary biomarker used to select non-small cell lung cancer (NSCLC) patients for checkpoint immunotherapy with drugs like pembrolizumab. PD-L1 Tumor Proportion Score (TPS) determines treatment eligibility.
The problem of inconsistency: PD-L1 scoring is performed by pathologists visually examining stained slides. Studies show significant inter- and intra-observer variability - the same slide can receive different scores from different pathologists or even from the same pathologist on different days. This inconsistency can directly affect treatment decisions.
The need for AI standardization: Artificial intelligence applied to digital whole-slide images (WSIs) can provide consistent, reproducible PD-L1 scoring. This study developed and validated MG-MIL - a multi-granularity multiple instance learning model - for automated PD-L1 evaluation.
Scale of validation: The study is notable for its large-scale validation: 706 patients, 1,212 whole-slide images across three independent cohorts, including surgical specimens, biopsy samples, and metastatic tissue.
What is Multiple Instance Learning (MIL): MIL is a type of machine learning where the model learns from 'bags' of instances rather than individual data points. For whole-slide images, each slide is a bag containing thousands of small image patches. The model learns which patches contain PD-L1-positive tumor cells without needing every patch manually annotated.
PLIP foundation model: MG-MIL uses PLIP (Pathology Language and Image Pre-training) as its image encoder - a powerful vision-language model pretrained on large pathology image datasets. PLIP extracts rich visual features from image patches that capture subtle staining patterns.
Multi-granularity design: The 'multi-granularity' innovation means the model operates at two scales simultaneously: coarse-grained analysis of large tissue regions (to understand overall PD-L1 distribution) and fine-grained analysis of individual tumor cells (to precisely count PD-L1-positive cells). This combination mirrors how expert pathologists assess slides.
Clinical threshold classification: Rather than just outputting a continuous PD-L1 score, MG-MIL classifies slides into clinically meaningful categories: TPS less than 1% (PD-L1 negative), 1-49% (low positive), and 50% or above (high positive, qualifying for pembrolizumab monotherapy).
Performance in surgical specimens: On the primary validation cohort of surgical resection specimens, MG-MIL achieved AUC of 0.940 for classifying TPS threshold at 1% and AUC of 0.915 for the 50% threshold. These results are comparable to expert pathologist performance.
Generalization to biopsies: A key concern with AI pathology models is whether they generalize to small biopsy samples, which have less tissue and more artifacts. MG-MIL maintained strong performance on biopsy specimens with AUC of 0.944 for TPS 50% classification.
Metastatic site validation: The model was also tested on metastatic tissue specimens - an important clinical scenario because PD-L1 can differ between primary and metastatic tumors. Performance remained consistent, suggesting the model captures genuine PD-L1 biology rather than tissue-specific artifacts.
Comparison to existing AI methods: MG-MIL outperformed conventional MIL approaches and other published AI PD-L1 scoring tools, demonstrating that the multi-granularity design adds meaningful accuracy over simpler approaches.
Reducing treatment selection errors: By providing consistent PD-L1 scores, MG-MIL reduces the risk of misclassifying patients as PD-L1 positive or negative due to pathologist variability. This directly affects whether patients receive appropriate immunotherapy.
Efficiency in high-volume labs: AI-assisted PD-L1 scoring can process thousands of slides faster than manual assessment, reducing report turnaround times and enabling higher throughput - particularly important in cancer centers with large NSCLC patient volumes.
Resolving discordant cases: When two pathologists disagree on a PD-L1 score - a common challenge that can delay treatment - AI scoring provides an objective third opinion. This is particularly valuable at the critical 1% and 50% TPS cutoffs.
Enabling remote pathology: AI standardization enables reliable PD-L1 scoring in community hospitals or low-resource settings that lack access to specialized pathologists, extending precision oncology to underserved populations.
Importance of scale in analysis: One key technical finding was that operating at both coarse and fine-grained scales significantly outperformed either scale alone. PD-L1 expression is spatially heterogeneous - some regions of a tumor are positive while others are negative - and multi-scale analysis captures this complexity.
Foundation models for pathology: The use of PLIP, a pathology-specific vision-language model, was superior to general image encoders like ResNet or ViT. Domain-specific pretraining on pathology images provides feature representations that better capture staining patterns relevant to pathological scoring.
Data augmentation strategies: To train robust models, the team used extensive augmentation including color jitter (to handle staining variability between labs), random crops, and rotations. This reduced overfitting to specific lab staining protocols.
Interpretability: Attention maps generated by the MIL model show which regions of each slide were most influential in the scoring decision - providing pathologists with visual explanations that build trust and enable error checking.
Different antibody clones: Multiple PD-L1 antibody clones (22C3, 28-8, SP142, SP263) are used clinically and give different results. MG-MIL was developed with specific clones, and its performance with all clones needs systematic evaluation.
Tumor heterogeneity challenges: Some NSCLC tumors show extreme intratumoral heterogeneity in PD-L1 expression. Even a perfect scoring algorithm cannot resolve the biological challenge that different tumor regions from the same patient may give different scores.
Prospective clinical integration: While validation results are impressive, the next critical step is a prospective study embedding MG-MIL into clinical workflows and assessing whether AI-standardized scoring leads to better patient outcomes compared to routine pathology.
Expanding to other biomarkers: The MG-MIL framework could potentially be extended to score other immunotherapy biomarkers such as tumor-infiltrating lymphocyte (TIL) density and tumor mutational burden estimation from histology - creating a comprehensive AI pathology suite.