A Robust Deep Learning Workflow for Quickly Establishing and Performing Image Analysis in Phenotypic Assays

September 10, 2017


Presented at SIBS, Zurich, Switzerland

We present an innovative workflow based on convolutional neural networks (‘Deep Learning’), providing life science research with more relevant information from today’s imaging technologies and increasing throughput (the principles of deep learning in HCS have recently been shown, e.g. by [1-2]). The workflow is tailored towards pharma-relevant assays, minimizing the need for expert time and knowledge, both during assay development and production. It aims to support complex research situations such as the phenotypic in-vitro assays that are highly relevant. These are typically performed in early drug discovery programs, and require costly expert time for many weeks per assay.

The key elements of the workflow are: 1. Visual and algorithmic tools to define known phenotypes and to reliably detect new phenotypes, 2. tools for effective training data generation and curation, 3. trained deep learning networks for production-level classification of high-content screening image-sets. In particular the easy production of training data is a prerequisite for productively using any deep learning approach. We illustrate the workflow using production assay data, showing a result quality that surpasses traditional image analysis (from basic classification performance to compound potency). We also compare the time and expertise investment between this workflow and traditional approaches and highlight the enormous timesaving, in particular for the tedious phenotype definition and curation task.

We connect the workflow principles with the underlying algorithmic approaches and discuss the pros and cons of possible strategies like auto-encoders, transfer-learning, and saliency maps in view of generalizability and robustness. Both are strong requirements to fully support scaled-up analysis and the complex phenotypes typical for assays in today’s life science and pharma research.

[1] Dürr O, Sick B, 2016. Single-Cell Phenotype Classification Using Deep Convolutional Neural Networks. JBS 21(9), 998-1003.
[2] Godinez WJ, et al., 2017. A multi-scale convolutional neural network for phenotyping high-content cellular images. Bioinformatics.

Back to list