A machine learning pipeline revealing heterogeneous responses to drug perturbations on vascular smooth muscle cell spheroid morphology and formation

Kalyanaraman Vaidyanathan, Chuangqi Wang, Amanda Krajnik, Yudong Yu, Moses Choi, Bolun Lin, Junbong Jang, Su‐Jin Heo, John Kolega, Kwonmoo Lee, Yongho BaeScientific 2021 · DOI

TL;DR

This paper introduces HETEROID, a machine learning pipeline that uses deep learning for image segmentation and clustering to analyze the complex and heterogeneous responses of vascular smooth muscle cells (VSMCs) to drug treatments. By modeling neointima formation in 3D spheroids, HETEROID reveals distinct morphological subpopulations of VSMCs under drug perturbation, offering a more nuanced understanding of drug efficacy for vascular diseases.

What Problem Does This Paper Address?

Vascular injury and diseases like atherosclerosis are characterized by the abnormal accumulation and proliferation of VSMCs, leading to neointima formation. While targeting VSMC behavior is crucial for treatment, a significant challenge is the inherent heterogeneity in how VSMCs respond to drugs, even within the same diseased condition. This variability makes it difficult to predict drug efficacy and develop targeted therapies. Traditional methods struggle to capture these subtle, heterogeneous responses at a single-cell or single-spheroid level. The authors aim to address this by developing a computational framework that can quantitatively characterize these diverse responses.

Methods

The core of this work is the HETEROID pipeline, which combines an in vitro VSMC spheroid model with a machine learning approach.

  1. VSMC Spheroid Model: They established a 3D spheroid culture system using VSMCs. This model mimics the in vivo neointima-like formation and arterial structure, providing a more physiologically relevant system than traditional 2D cultures.
  2. Deep Learning-Based Segmentation: To analyze the complex morphology of these spheroids, especially under drug treatment, they employed a deep learning model. Specifically, they combined a U-Net architecture with a VGG16 encoder pre-trained on ImageNet. This VGG16-U-Net model was trained to accurately segment spheroid boundaries from phase-contrast microscopy images, achieving high accuracy (96% at epoch 30) and outperforming standard U-Net.
  3. Unsupervised Morphological Clustering: After segmentation, morphological features (e.g., circularity, solidity, eccentricity) were extracted from the spheroid images. These features were then fed into a two-step unsupervised clustering analysis:
  4. Initial Clustering: Focused on roundness-related features to differentiate between generally circular and non-circular/disrupted spheroids.
  5. Non-circular Spheroid Clustering: For the disrupted spheroids, a wider range of morphological features was used to identify finer-grained subpopulations. This clustering approach, using a community detection algorithm and UMAP for visualization, allowed them to identify distinct morphological "types" or subpopulations of VSMC spheroids resulting from different drug treatments.

Datasets

The study utilized datasets generated from their in vitro experiments.

  • VSMC Spheroid Images: Images of human VSMC spheroids treated with DMSO (control) and various inhibitors (FAK, Rac, Rho, Cdc42) at different concentrations were collected. A total of 856 images were used for unsupervised morphology learning (Table 1).
  • Training Data for Segmentation: A subset of 23 diverse VSMC spheroid images was manually annotated with spheroid boundaries to train the VGG16-U-Net segmentation model.

No pre-existing public datasets were used for training or evaluation; the study generated its own experimental data.

Key Insights and Ideas

  • Heterogeneity is Key: The central idea is that VSMC responses to drugs are not uniform. Even with the same drug and condition, spheroids exhibit distinct morphological profiles, suggesting different underlying cellular states or pathways are being affected.
  • 3D Spheroids as a Better Model: Using 3D spheroids provides a more relevant model for neointima formation than 2D cultures, allowing for the observation of complex cell-cell and cell-ECM interactions that influence morphology.
  • Deep Learning for Nuanced Analysis: Standard image analysis struggles with the complex and varied morphologies induced by drug treatments. Deep learning segmentation and unsupervised clustering are essential for objectively identifying and quantifying these subtle differences.
  • Differential Pathway Involvement: The study highlights that different signaling pathways (FAK, Rac, Rho, Cdc42) not only affect spheroid formation but do so through distinct mechanisms, leading to unique morphological outcomes. For instance, Cdc42 inhibition resulted in multiple small spheroids, distinct from the dispersed morphologies seen with FAK, Rac, or Rho inhibition.

Key Results

  • FAK Inhibition Disrupts Spheroid Formation: Inhibition of FAK phosphorylation significantly disrupted VSMC spheroid formation, leading to a wide range of morphologies, indicating heterogeneity in response.
  • Rac, Rho, and Cdc42 Inhibitors Show Differential Effects:
    • High doses of Rac and Rho inhibitors caused significant disruption of spheroid formation.
  • Cdc42 inhibition led to the formation of multiple small cell aggregates, a distinct phenotype.
  • HETEROID Identifies Morphological Subpopulations: The machine learning pipeline successfully identified four distinct morphological clusters (Types 1-4) of VSMC spheroids.
  • Type 1: Predominantly circular, unperturbed spheroids (common in DMSO controls and low-dose inhibitor treatments).
  • Type 2 & 4: Non-circular, dispersed, or disrupted morphologies.
  • Type 3: Small circular shapes, often containing multiple small spheroids (enriched with Cdc42 inhibition).
  • Drug-Specific Subpopulation Distributions: Different drugs and concentrations led to distinct proportions of spheroids within these clusters. For example, FAK inhibition significantly shifted spheroids towards Type 2 and Type 4, while Cdc42 inhibition strongly enriched Type 3.
  • Further Clustering Reveals Finer Details: A secondary clustering analysis on disrupted spheroids (Types 2 & 4) identified four new subpopulations (N1-N4), revealing more specific patterns of disruption linked to FAK, Rac, and Rho inhibition. For instance, Type N3 morphology was strongly associated with Rho and Rac disruption.

Limitations and Future Work

  • Limitations:
    • The study primarily used human VSMCs, with some validation in mouse VSMCs.
  • The analysis focused on morphological changes; further investigation into the underlying molecular mechanisms (transcriptome, proteome) of these distinct spheroid types is needed.
  • The current clustering is based on morphology; integrating functional assays (viability, proliferation, apoptosis) within these morphological clusters would provide deeper biological insights.
  • Future Work:
    • Investigate the biological and functional roles of FAK, Rac, Rho, and Cdc42 in VSMC spheroid formation using 'Omics' technologies to understand the molecular basis of the identified morphological clusters.
  • Determine why only a subpopulation of VSMCs participates in neointima formation.
  • Leverage the HETEROID pipeline for personalized medicine by testing drug effects on patient-specific spheroids before prescribing treatments.
  • Expand the application of this machine learning approach to study other diseases involving VSMC dysfunction, such as hypertension and restenosis.

Summary updated

Abstract

Machine learning approaches have shown great promise in biology and medicine discovering hidden information to further understand complex biological and pathological processes. In this study, we developed a deep learning-based machine learning algorithm to meaningfully process image data and facilitate studies in vascular biology and pathology. Vascular injury and atherosclerosis are characterized by neointima formation caused by the aberrant accumulation and proliferation of vascular smooth muscle cells (VSMCs) within the vessel wall. Understanding how to control VSMC behaviors would promote the development of therapeutic targets to treat vascular diseases. However, the response to drug treatments among VSMCs with the same diseased vascular condition is often heterogeneous. Here, to identify the heterogeneous responses of drug treatments, we created an in vitro experimental model system using VSMC spheroids and developed a machine learning-based computational method called HETEROID (heterogeneous spheroid). First, we established a VSMC spheroid model that mimics neointima-like formation and the structure of arteries. Then, to identify the morphological subpopulations of drug-treated VSMC spheroids, we used a machine learning framework that combines deep learning-based spheroid segmentation and morphological clustering analysis. Our machine learning approach successfully showed that FAK, Rac, Rho, and Cdc42 inhibitors differentially affect spheroid morphology, suggesting that multiple drug responses of VSMC spheroid formation exist. Overall, our HETEROID pipeline enables detailed quantitative drug characterization of morphological changes in neointima formation, that occurs in vivo, by single-spheroid analysis.
BibTeX
@inproceedings{vaidyanathan2021machine,
  author = {Kalyanaraman Vaidyanathan and Chuangqi Wang and Amanda Krajnik and Yudong Yu and Moses Choi and Bolun Lin and Junbong Jang and Su‐Jin Heo and John Kolega and Kwonmoo Lee and Yongho Bae},
  title = {{A machine learning pipeline revealing heterogeneous responses to drug perturbations on vascular smooth muscle cell spheroid morphology and formation}},
  booktitle = {Scientific},
  year = {2021},
  doi = {10.1038/s41598-021-02683-4},
}
Figure