In vivo tumor immune microenvironment phenotypes correlate with inflammation and vasculature to predict immunotherapy response

Aditi Sahu, Kıvanç Köse, Lukas Kraehenbuehl, Candice Byers, Aliya Holland, Teguru Tembo, Anthony Santella, Anabel Alfonso, Madison Li, Miguel Córdova, Melissa Gill, Christi Alessi‐Fox, Salvador González, Piyush Kumar, Amber Weiching Wang, Nicholas R. Kurtansky, Pratik Chandrani, Shen Yin, Paras Mehta, Cristián Navarrete‐Dechent, Gary Peterson, Kimeil King, Stephen W. Dusza, Ning Yang, Shuaitong Liu, William T. Phillips, Pascale Guitera, Anthony Rossi, Allan C. Halpern, Liang Deng, Melissa Pulitzer, Ashfaq A. Marghoob, Chih-Shan Jason Chen, Taha Merghoub, Milind RajadhyakshaNature Communications 2022 · DOI

TL;DR

This paper introduces a novel, non-invasive in vivo imaging technique to dynamically phenotype the tumor microenvironment (TiME) by combining inflammation and vascular features. This approach reveals distinct TiME phenotypes that accurately predict patient response to immunotherapy, offering a more robust alternative to static, biopsy-based methods.

What Problem Does This Paper Address?

Predicting a patient's response to immunotherapy remains a significant challenge. Current methods, like classifying tumors as "hot" (inflamed) or "cold" (non-inflamed) based on single-time point biopsies, are often insufficient. This static view misses the dynamic interplay of immune cells and the tumor's vasculature, which are crucial for anti-tumor immunity and treatment efficacy. Tumors can evade immune responses through complex mechanisms involving both inflammation and vascular changes, leading to unpredictable treatment outcomes. This paper aims to bridge this gap by developing a dynamic, in vivo phenotyping method that captures these critical, interconnected features of the TiME.

Methods

The core of this research lies in Reflectance Confocal Microscopy (RCM), a non-invasive, label-free imaging technique. RCM allows for real-time visualization of dynamic cellular processes within the tumor microenvironment, including leukocyte trafficking and vascular activity, at a depth of up to ~0.25 mm.

The authors developed a multi-stage approach:

  1. In Vivo RCM Imaging: Capturing large mosaics of skin lesions to comprehensively sample tumor, peritumoral, and normal areas.
  2. Feature Extraction: Identifying and quantifying key inflammation (e.g., intratumoral inflammation) and vasculature (e.g., vessel density, diameter) features from the RCM images.
  3. Machine Learning/Automated Quantification: Employing machine learning algorithms (specifically a U-Net model for immune cell segmentation) to automate the quantification of these features, improving reproducibility and efficiency.
  4. Unsupervised Clustering: Using Principal Component Analysis (PCA) and Hierarchical Clustering (HCPC) to group lesions into distinct TiME phenotypes based on the extracted features.
  5. Validation: Correlating RCM-derived phenotypes with histopathology, multiplexed immunofluorescence (IF), immunohistochemistry (IHC), and bulk RNA sequencing to understand the underlying molecular and cellular signatures.
  6. Treatment Response Prediction: Evaluating the predictive power of these TiME phenotypes and features in a prospective pilot study of patients receiving topical toll-like receptor agonist (TLRA) immunotherapy.

Datasets

The study utilized datasets from patients with basal cell carcinoma (BCC) and melanoma.

  • BCC Cohort: 118 lesions were imaged with RCM. A subset of 27 BCC lesions was used for unsupervised clustering and phenotyping. Further molecular analysis (RNA-seq, multiplexed IF, IHC) was performed on smaller subsets of these BCC samples.
  • Melanoma Cohort: 13 lesions were used for phenotyping.
  • TLRA Treatment Cohort: A subset of 13 BCC lesions from patients undergoing topical TLRA treatment were imaged before and after treatment to assess response.

No new benchmark datasets were introduced; rather, existing patient data was leveraged for analysis and validation.

Key Insights and Ideas

  • Dynamic TiME Phenotyping is Crucial: The study strongly advocates for moving beyond static, single-point biopsies to dynamic, in vivo imaging for a more accurate understanding of the TiME and its impact on treatment response.
  • Inflammation and Vasculature are Intertwined Predictors: The interplay between inflammation and vasculature, not just inflammation alone, is key to predicting immunotherapy response.
  • The "InflamHIGHVascLOW" Phenotype is Optimal for Response: This phenotype, characterized by high inflammation but low vasculature, showed the best response to TLRA immunotherapy, suggesting a delicate balance is needed for effective anti-tumor immunity.
  • Endothelial Anergy is a Key Evasion Mechanism: High vasculature, particularly with signs of endothelial anergy, appears to be associated with immune evasion and poorer response.
  • Automated Quantification is Feasible: The development of machine learning models for immune cell segmentation and feature quantification demonstrates the potential for robust, objective assessment of TiME features.

Key Results

  • High Agreement with Histopathology: RCM-derived TiME features showed substantial to almost perfect agreement with histopathological evaluations (Cohen's kappa: 0.62-1.0).
  • Three Distinct TiME Phenotypes in BCC: Unsupervised clustering identified three main phenotypes in BCC: InflamLowVascHIGH, InflamHIGHVascLOW, and InflamHIGHVascHIGH.
  • Two Distinct TiME Phenotypes in Melanoma: Two main phenotypes were identified: InflamLowVascHIGH and InflamHIGHVascHIGH.
  • Molecular Signatures Correlate with Phenotypes: Gene expression analysis revealed distinct molecular signatures associated with inflammation (M2 module) and vasculature (M5 module), with InflamHIGH phenotypes showing stronger immunostimulatory signatures.
  • InflamHIGHVascLOW Predicts Best Response: In the TLRA treatment cohort, the InflamHIGHVascLOW phenotype was significantly associated with the best treatment response.
  • Combined Features Improve Prediction: Linear regression models combining inflammation and vascular features achieved significantly better response prediction accuracy (71% sensitivity, 83% specificity, 76% accuracy) compared to inflammation alone.
  • Automated Quantification Accuracy: The U-Net model for immune cell segmentation achieved a Dice coefficient of 0.72, and vascular feature quantification showed Dice coefficients ranging from 0.29-0.78.

Limitations and Future Work

  • Limited Sample Size for Melanoma and Treatment Cohorts: The study acknowledges that the melanoma and treatment response cohorts were relatively small, necessitating larger studies for further validation.
  • Depth Limitation of RCM: RCM's imaging depth is limited to approximately 0.25 mm, potentially missing deeper TiME features or tertiary lymphoid structures (TLS) that might influence outcomes.
  • Grayscale-Limited Specificity: While label-free, RCM's grayscale output limits specificity for functional phenotyping.
  • Need for Single-Cell Resolution: Future work should incorporate single-cell RNA sequencing to comprehensively understand cellular heterogeneity and gene expression differences across phenotypes, particularly in stromal cells.
  • Further Validation of Automated Pipelines: Continued refinement and validation of automated quantification pipelines for vascular features and individual trafficking events are planned.
  • Longitudinal Monitoring: The potential for longitudinal monitoring of TiME changes during treatment to assess response and uncover resistance mechanisms is a promising future direction.
  • Complementary Imaging Modalities: Integrating RCM with other imaging techniques like OCT, photoacoustic microscopy, or fluorescence lifetime imaging could enhance TiME visualization and functional phenotyping.
  • Broader Application: While demonstrated in skin cancers, the approach holds potential for application in other cancer types.

Summary updated

Abstract

Response to immunotherapies can be variable and unpredictable. Pathology-based phenotyping of tumors into 'hot' and 'cold' is static, relying solely on T-cell infiltration in single-time single-site biopsies, resulting in suboptimal treatment response prediction. Dynamic vascular events (tumor angiogenesis, leukocyte trafficking) within tumor immune microenvironment (TiME) also influence anti-tumor immunity and treatment response. Here, we report dynamic cellular-level TiME phenotyping in vivo that combines inflammation profiles with vascular features through non-invasive reflectance confocal microscopic imaging. In skin cancer patients, we demonstrate three main TiME phenotypes that correlate with gene and protein expression, and response to toll-like receptor agonist immune-therapy. Notably, phenotypes with high inflammation associate with immunostimulatory signatures and those with high vasculature with angiogenic and endothelial anergy signatures. Moreover, phenotypes with high inflammation and low vasculature demonstrate the best treatment response. This non-invasive in vivo phenotyping approach integrating dynamic vasculature with inflammation serves as a reliable predictor of response to topical immune-therapy in patients.
BibTeX
@inproceedings{sahu2022vivo,
  author = {Aditi Sahu and Kıvanç Köse and Lukas Kraehenbuehl and Candice Byers and Aliya Holland and Teguru Tembo and Anthony Santella and Anabel Alfonso and Madison Li and Miguel Córdova and Melissa Gill and Christi Alessi‐Fox and Salvador González and Piyush Kumar and Amber Weiching Wang and Nicholas R. Kurtansky and Pratik Chandrani and Shen Yin and Paras Mehta and Cristián Navarrete‐Dechent and Gary Peterson and Kimeil King and Stephen W. Dusza and Ning Yang and Shuaitong Liu and William T. Phillips and Pascale Guitera and Anthony Rossi and Allan C. Halpern and Liang Deng and Melissa Pulitzer and Ashfaq A. Marghoob and Chih-Shan Jason Chen and Taha Merghoub and Milind Rajadhyaksha},
  title = {{In vivo tumor immune microenvironment phenotypes correlate with inflammation and vasculature to predict immunotherapy response}},
  booktitle = {Nature Communications},
  year = {2022},
  doi = {10.1038/s41467-022-32738-7},
}
Figure