Non-Invasive Quantification of Viability in Spheroids Using Deep Learning

Daniel Dubinsky, Shahar Harel, Amir Bein, Abraham Nyska, Sarah Yaari, Baran Koc, Faiza Anas, Isaac Bentwich, Lior WolfbioRxiv 2025 · DOI

TL;DR

This paper introduces Neural Viability Regression (NViR), a deep learning model that non-invasively quantifies cell viability in spheroids using microscopy images. NViR enables real-time monitoring, reduces costs, and accurately predicts drug-induced liver injury, offering a significant advancement over traditional destructive viability assays.

What Problem Does This Paper Address?

Evaluating cell viability is crucial in drug discovery and development. However, current standard methods, like bioluminescence assays, are destructive. This means they can only be performed once per cell culture, limiting the ability to track changes over time and requiring many replicates, which is costly and labor-intensive. This "endpoint assay" approach prevents researchers from observing subtle, dynamic changes in cell health. The paper aims to overcome this by developing a non-invasive method for continuous viability assessment.

Methods

The core of the paper is the Neural Viability Regression (NViR) model. It's a deep learning approach that takes bright-field microscopy images of spheroids as input and outputs a predicted viability score.

Here's a breakdown of the key methodological aspects:

  • Deep Learning Architecture: NViR uses a ResNet18 backbone, pre-trained on ImageNet, followed by a Multi-Layer Perceptron (MLP) regression head. This is a standard approach for image-based regression tasks.
  • Non-Invasive Imaging: The method relies solely on bright-field microscopy images, meaning no dyes or destructive procedures are needed.
  • Data Generation: A high-throughput microphysiological system (MPS) platform was used to generate a dataset of spheroid images under various drug treatments. This platform automates spheroid handling, drug administration, and imaging.
  • ATP Normalization: A crucial preprocessing step involves normalizing the measured ATP (viability indicator) values by the median ATP of control spheroids. This helps to mitigate batch effects and align viability distributions across different studies.
  • Classical Vision Encoder (CVE) Baseline: For comparison, the authors implemented a baseline that extracts handcrafted features (e.g., contour, brightness) from segmented spheroid images and uses a machine learning regressor.
  • Pathologist Baseline: Human experts (pathologists) also evaluated spheroid images based on morphological criteria (cytoplasmic, nuclear, and contour changes), and their scores were used to train a regressor.

Datasets

The study utilized a custom dataset for training and evaluation:

  • Spheroid Image Dataset: This dataset consists of 17,704 unlabeled and 2,359 labeled spheroid bright-field images. Each labeled image is paired with its measured viability level (ATP value).
  • Domains: The dataset was organized into four domains (A, B, C, D), distinguished by laboratory location or microscope type, to test the model's generalization capabilities. Domain A was used for training, validation, and testing, while B, C, and D were used exclusively for testing.
  • Drug-Induced Liver Injury (DILI) Dataset: A subset of 108 FDA-approved drugs was used to evaluate NViR's ability to predict DILI. This dataset was derived from previous work and included drug concentrations and DILI labels.

Key Insights and Ideas

  • Visual Cues for Viability: Spheroids with lower viability exhibit distinct morphological changes (altered contours, varied nucleus darkness) that are visually discernible. NViR leverages these subtle visual cues, which are often missed by traditional methods.
  • Generalization Across Domains: NViR demonstrates strong generalization capabilities, performing well even on images captured by microscopes or in labs it wasn't trained on. This is a significant advantage for real-world applicability.
  • Latent Space Interpretability: The study explored NViR's latent space, finding correlations between its internal representations and established fluorescence-based viability biomarkers. This suggests that NViR is learning biologically relevant features.
  • Continuous Monitoring Advantage: The non-invasive nature allows for frequent, real-time monitoring, revealing dynamic viability trends that would be hidden in endpoint assays. This is crucial for understanding drug effects over time.

Key Results

  • Viability Prediction Accuracy: NViR significantly outperformed both the Classical Vision Encoder (CVE) baseline and the pathologist-based baseline across all four domains.
  • NViR achieved an R² score of 0.83 ± 0.03 in domain A, compared to the best baseline (Pathology + CVE) at 0.56 ± 0.00.
  • Mean Absolute Error (MAE) for NViR was 0.113 ± 0.012 in domain A, significantly lower than the best baseline (0.194 ± 0.000).
  • Drug-Induced Liver Injury (DILI) Prediction: NViR accurately predicted DILI, achieving an AUC of 0.7335, comparable to a previous method (AUC 0.7338) that relied on destructive ATP assays after 14 days. NViR achieved this performance using data from just 7 days.
  • Cost and Time Savings: The non-invasive nature and high-throughput capability of NViR promise to reduce both the costs and time associated with drug discovery and development.

Limitations and Future Work

  • Latent Space Disentanglement: While NViR's latent space shows correlations with biomarkers, it's not explicitly optimized for disentanglement. Future work could focus on creating more interpretable latent spaces.
  • Adaptability to Other Spheroid Types: While NViR is trained on liver spheroids, the authors suggest it can be adapted for other spheroid types and 3D cell cultures by acquiring similar data.
  • Integration of Other Modalities: The paper mentions that other data modalities (e.g., mRNA, membrane integrity) can be used to create AI-enhanced biomarkers. Future work could explore multi-modal approaches with NViR.
  • Broader Applications: The authors envision NViR being integrated earlier in the drug development pipeline due to its cost-effectiveness and non-invasive nature, potentially reducing late-stage failures.

Summary updated

Abstract

1 Abstract In vitro viability assays are widely used in drug discovery, development, and pharmacovigilance. Traditional methods for evaluating cell viability often involve destructive processes, rendering the culture non-viable. As such, these methods are suitable only as endpoint solutions, providing a single measurement per cell culture and precluding further analyses of the cells. In this study, we introduce Neural Viability Regression (NViR), a real-time, deep learning-based method that enables non-invasive quantification of culture viability using microscopy images. The non-intrusive nature of NViR allows for frequent viability evaluations throughout experiments, capturing subtle changes while maintaining the structural integrity of the culture and significantly reducing both culture and labor costs. We demonstrate NViR’s applicability by using it to predict Drug-Induced Liver Injury (DILI) in known drugs. By exposing human liver spheroids to 108 FDA-approved drugs and capturing microscopy images over time, NViR’s viability assessments accurately predict whether a drug induces DILI in humans, playing a critical role in enhancing liver safety protocols. The cost-effectiveness and non-invasive characteristics of NViR enable high-frequency, high-throughput viability assessments. Consequently, NViR is poised to reduce both the costs and incidences of failures in drug discovery and development.
BibTeX
@misc{dubinsky2025non,
  author = {Daniel Dubinsky and Shahar Harel and Amir Bein and Abraham Nyska and Sarah Yaari and Baran Koc and Faiza Anas and Isaac Bentwich and Lior Wolf},
  title = {{Non-Invasive Quantification of Viability in Spheroids Using Deep Learning}},
  year = {2025},
  doi = {10.1101/2025.03.09.642246},
  howpublished = {bioRxiv},
}

Loading related content...

Figure