Integration of Biodynamic Imaging and RNA-seq classifies chemotherapy response in canine diffuse large B-cell lymphoma

Jonathan Fine, Sagar M. Utturkar, Deepika Dhawan, Phillip San Miguel, Gaurav Chopra, John Turek, David D. Nolte, Michael O. Childress, Nadia A. LanmanbioRxiv 2020 · DOI

TL;DR

This study integrates two powerful techniques, BioDynamic Imaging (BDI) and RNA sequencing (RNA-seq), to predict chemotherapy response in canine diffuse large B-cell lymphoma (DLBCL). By combining phenotypic and transcriptomic data, they developed a machine learning model that accurately classifies treatment response, offering a promising framework for personalized cancer medicine.

What Problem Does This Paper Address?

Diffuse large B-cell lymphoma (DLBCL) is an aggressive cancer with significant genetic and cellular variability, making it challenging to predict how individual patients will respond to chemotherapy. Current methods, relying primarily on genomic data, have limited success. There's a critical need for a more comprehensive approach that captures both the cellular behavior (phenotype) and the underlying genetic activity (transcriptome) to predict treatment efficacy before administering therapy. This could prevent patients from undergoing ineffective treatments and experiencing unnecessary side effects.

Methods

The researchers employed a multi-modal approach:

  • BioDynamic Imaging (BDI): This technique uses Doppler spectroscopy to measure time-dependent fluctuations in back-scattered light from living tumor tissues. It captures changes in intracellular dynamics, essentially observing how cells behave under drug exposure. This generates "spectrograms" that are then converted into feature vectors representing phenotypic biomarkers.
  • RNA Sequencing (RNA-seq): This standard technique measures gene expression levels in tumor samples, providing a snapshot of the cellular machinery at work.
  • Machine Learning (Regularized Logistic Regression): To integrate these two data types, the authors used regularized logistic regression (LASSO and ridge) to build a predictive model.
  • Feature Selection: Given the high dimensionality of the data, aggressive feature selection was crucial. They used Area-Under-the-Precision-Recall Curve (AUPRC) to select the most informative BDI and RNA-seq features.
  • Network Analysis (WGCNA): Weighted Gene Co-expression Network Analysis (WGCNA) was used to identify modules of co-expressed genes and correlate them with BDI biomarkers, aiming to provide biological mechanistic interpretations for the BDI signals.

Datasets

The study utilized tumor samples from nineteen pet dogs naturally afflicted with DLBCL. This canine model is considered relevant to human DLBCL due to similar disease characteristics and treatment responses. The data collected included:

  • BDI measurements from ex vivo tumor biopsies.
  • RNA-seq data from the same tumor samples.

Key Insights and Ideas

  • Phenotype + Transcriptome = Better Prediction: The core insight is that combining phenotypic information (how cells behave) with transcriptomic information (what genes are active) significantly improves the prediction of chemotherapy response compared to using either data type alone.
  • BDI as a Window into Cellular Dynamics: BDI provides a unique, dynamic view of cellular responses to drugs, capturing changes in transcription, microtubule processes, and apoptosis, which are crucial for understanding drug sensitivity.
  • Bridging the Gap Between BDI and Biology: By correlating BDI biomarkers with gene expression modules, the study offers biological interpretations for BDI signals, making them more clinically actionable. For example, certain BDI signals were linked to processes like microtubule function and apoptosis.
  • Canine Model's Translational Value: The study reinforces the utility of the canine DLBCL model for developing personalized medicine strategies, as their responses to chemotherapy mirror those seen in humans.

Key Results

  • Accurate Classification: A model incorporating 3 BDI biomarkers and 4 RNA-seq variables achieved a kappa of 1.00 and an accuracy of 100% in predicting chemotherapy response using leave-one-out testing.
  • Identification of Key Genes and Biomarkers: The selected features included BDI biomarkers like SDIP1dox, LOFOchop, and ALLF1pred, and RNA-seq variables such as ZFP92, KIAA1217, SH2D4A, and FGFR4.
  • Biological Correlations: Network analysis revealed correlations between BDI biomarkers and gene expression modules involved in processes like microtubule function, cytoskeleton dynamics, and apoptosis, providing mechanistic links.
  • Enriched Pathways in Resistant Tumors: RNA-seq analysis identified enriched pathways in resistant tumors, including the Jak-Stat pathway, oxidative phosphorylation, and MYC targets, mirroring findings in human DLBCL.

Limitations and Future Work

  • Small Sample Size: The authors acknowledge that the small sample size (19 dogs) necessitated aggressive feature selection and that the results should be considered preliminary.
  • Subpopulation Heterogeneity: The study grouped dogs with primary refractory DLBCL and those that became rapidly resistant after initial response. Future work with larger datasets could potentially resolve these distinct subpopulations.
  • External Validation: The authors emphasize the need for external validation of the developed model on independent datasets.
  • Broader Applications: They suggest performing similar analyses with larger sample sizes and across different drug treatments and cancer types to confirm the biological interpretations and expand the model's applicability.
  • Spatially Resolved Data: Future improvements could involve incorporating spatially resolved gene expression analyses to further enhance the mechanistic understanding of BDI data.

Summary updated

Abstract

Abstract Diffuse large B-cell lymphoma (DLBCL) is a common, aggressive cancer of notorious genotypic and phenotypic heterogeneity. A major challenge is predicting response to drug treatment, which has typically been done using genomic tools alone with little success. A novel method that incorporates phenotypic profiling for predicting the effectiveness of therapy for individual patients is desperately needed. BioDynamic Imaging (BDI) is a technique for measuring time-dependent fluctuations in back-scattered light through living tumor tissues to identify critical changes in intracellular dynamics that are associated with phenotypic response to drugs. In this study, BDI and RNA sequencing (RNA-seq) data were collected on tumor samples from dogs with naturally occurring DLBCL, an animal model of increasingly recognized relevance to the human disease. BDI and RNA-seq data were combined to identify correlations between gene co-expression modules and linear combinations of biomarkers to provide biological mechanistic interpretations of BDI biomarkers. Using regularized multivariate logistic regression, we combined RNA-seq and BDI data to develop a novel model to accurately classify the clinical response of canine DLBCL to combination chemotherapy (i.e. CHOP). Our model incorporates data on the expression of 4 genes and 3 BDI-derived phenotypic biomarkers, capturing changes in transcription, microtubule related processes, and apoptosis. This pilot study suggests that the combination of multi-scale transcriptomic and phenotypic data can identify patients that respond to a given treatment a priori in a disease that has been difficult to treat. Our work provides an important framework for future development of strategies and treatments in precision cancer medicine.
BibTeX
@misc{fine2020integration,
  author = {Jonathan Fine and Sagar M. Utturkar and Deepika Dhawan and Phillip San Miguel and Gaurav Chopra and John Turek and David D. Nolte and Michael O. Childress and Nadia A. Lanman},
  title = {{Integration of Biodynamic Imaging and RNA-seq classifies chemotherapy response in canine diffuse large B-cell lymphoma}},
  year = {2020},
  doi = {10.1101/2020.09.11.290353},
  howpublished = {bioRxiv},
}
Figure