Skip to content

Non-Invasive Quantification of Viability in Spheroids Using Deep Learning

2026
Paper Title: Non-Invasive Quantification of Viability in Spheroids Using Deep Learning
Published in: Frontiers in Bioengineering and Biotechnology
Link: Read the paper (DOI: 10.3389/fbioe.2026.1797474)

Summary

In drug discovery, assessing cell viability is critical but often relies on destructive "endpoint" assays (like measuring ATP), which kill the culture and prevent further analysis. This project introduces Neural Viability Regression (NViR), a deep learning-based method that quantifies culture viability non-invasively using standard microscopy images.

By preserving the culture, NViR enables tracking of the same sample over time, capturing temporal changes and significantly reducing costs.

Key Innovation: Neural Viability Regression (NViR)

NViR replaces the destructive chemical assay with a computer vision proxy. The model takes a bright-field microscopy image of a spheroid and regresses the viability score (normalized ATP level).

Technical Architecture

  • Backbone: ResNet18 pre-trained on ImageNet.
  • Head: Multi-Layer Perceptron (MLP) regression head.
  • Input: Bright-field microscopy images of liver spheroids.
  • Output: viability score (normalized ATP).

The model was trained on a dataset of human liver spheroids exposed to various compounds.

Application: Drug-Induced Liver Injury (DILI) Prediction

To validate the method, NViR was applied to the task of predicting Drug-Induced Liver Injury (DILI). * Dataset: 108 FDA-approved drugs. * Method: Spheroids were imaged over time while exposed to drugs. * Result: The viability profiles generated by NViR accurately distinguished between toxic and non-toxic drugs, matching the performance of destructive gold-standard assays but with a fraction of the cost and labor.

Impact

  1. Non-Invasive: Enables "online" monitoring of experiments without destroying samples.
  2. High-Throughput: Automated imaging allows for massive scaling of viability screens.
  3. Cost Reduction: Eliminates expensive reagents needed for chemical assays.

Code Availability

The code for this project is available on GitHub: DanielDubinsky/atp_paper