henhaobio
Sign in

bioRxivBioinformaticsPreprint

Automating Morphological Profiling with Generic Deep Convolutional Networks

Nick PawlowskiiD, Juan C. CaicedoiD, Shantanu SinghiD, Anne Elizabeth CarpenteriD, Amos StorkeyiD

PDFDOICode
In-depth full-text digest available · Jump to digest ↓

Abstract

Morphological profiling aims to create signatures of genes, chemicals and diseases from microscopy images. Current approaches use classical computer vision-based segmentation and feature extraction. Deep learning models achieve state-of-the-art performance in many computer vision tasks such as classification and segmentation. We propose to transfer activation features of generic deep convolutional networks to extract features for morphological profiling. Our approach surpasses currently used methods in terms of accuracy and processing speed. Furthermore, it enables fully automated processing of microscopy images without need for single cell identification.

In-depth summary

Claude Opus 5.5 · xhigh effort · 8 Nov 2026 · from the full text of version 1 (2 Nov 2016) · CC BY 4.0

ImageNet-pretrained Inception-v3, without fine-tuning or segmentation, reaches 91% mechanism-of-action accuracy on illumination-corrected BBBC021 versus 90% for classical mean profiles, running in 22 minutes versus over 10 hours for single-core CellProfiler.

Image-based profiling of chemical and genetic perturbations commonly relies on CellProfiler, with segmentation and feature extraction hand-tuned for each assay. Here, ImageNet-trained networks serve as fixed feature extractors with no segmentation step, and they are scored with the same leave-one-compound-out protocol as the classical baselines; the authors contrast this with earlier deep learning on BBBC021 (Kraus et al.), which used a small subset and a single train/test split.

Key results

  1. ShownOff-the-shelf ImageNet-pretrained networks, with no fine-tuning and no cell segmentation, yield features that classify compound mechanism of action (MOA) well above chance; Inception-v3 is the strongest of the four networks tested.

    Accuracies range from 55% to 91% across VGG16, ResNet-101, ResNet-152 and Inception-v3 configurations (Table 1), against 8% for random chance. Scoring uses 1-nearest-neighbour MOA classification with leave-one-compound-out cross-validation. · 3 Experimental Setup; 4 Results; Tables

  2. SuggestedInception-v3 profiles slightly beat classical CellProfiler mean profiles on the same benchmark, with and without illumination correction.

    86% vs 84% on uncorrected images and 91% vs 90% on illumination-corrected images. The classical confusion matrix in Figure 1 was taken from Singh et al. (whose reported corrected-image accuracy is 90%). The authors call the accuracy gain marginal. · 2 Related work; 4 Results; 5 Conclusion; Figure legends

  3. ShownFeeding each stain to the network separately as a greyscale image and concatenating the features always beat an arbitrary RGB mapping of the three stains. Illumination correction improved almost every network and input configuration.

    Reported across all four networks in Table 1. The RGB mapping was DNA to R, tubulin to G and actin to B. The text does not give the individual values for these comparisons. · 3 Experimental Setup; 4 Results; Tables

  4. ShownExtracting deep features was much faster than running the classical CellProfiler pipeline on the same images.

    Inception-v3 took 22 minutes to transform the image set. CellProfiler took more than 10 hours for the same images on a single core. Both ran on a system with two 8-core Xeon CPUs, 128GB memory and no GPU. · 4 Results

  5. SuggestedThe errors of the deep and classical profiles seem to be complementary, which the authors take to suggest that the network captures different features than the classical method.

    Based on a visual comparison of confusion matrices: Inception-v3 with per-stain greyscale input vs classical results from Singh et al. (Figure 1). No measure of complementarity and no combined model are reported. · 4 Results; Figure legends

Methods, data and reproducibility

Downsampled BBBC021 images were passed, without fine-tuning or cell segmentation, through ImageNet-pretrained Inception-v3, VGG16, ResNet-101 and ResNet-152. Penultimate-layer activations (stains mapped to RGB, or each stain fed as greyscale and concatenated) were averaged per treatment. The profiles were scored by 1-nearest-neighbour MOA classification that leaves out every concentration of the query compound.

  • BBBC021 image set (Broad Bioimage Benchmark Collection): MCF7 breast cancer cells treated with compounds at several concentrations, stained for DNA, actin and tubulin; 1280px × 1024px images · human (MCF7 cell line) · BBBC021
  • Classical CellProfiler mean-profiling baseline (Singh et al.; Figure 1 confusion matrix taken from that paper) · human (MCF7 cell line)
Methods & tools
Inception-v3, VGG16, ResNet-101, ResNet-152, ImageNet, CellProfiler, TensorFlow, 1-nearest-neighbour classifier
Assay conditions & compute
Inception-v3 feature extraction took 22 minutes, vs more than 10 hours for CellProfiler on a single core; both ran on two 8-core Xeon CPUs with 128GB memory and no GPU.
Reproducibility gaps
No URL is given for BBBC021 (it is named only as part of the Broad Bioimage Benchmark Collection). No software versions or environment files are stated. The classical confusion matrix is taken from Singh et al.; the text does not say whether the classical accuracies were recomputed.
Limitations (6)
  • The authors describe the accuracy gain over classical mean profiling as marginal; the main advantage they claim is that no image analysis expert is needed. · stated by the authors
  • The networks were not fine-tuned. The authors suggest that fine-tuning, and deep learning in general, may need larger microscopy image sets. · stated by the authors
  • Full images were downsampled to 299px or 224px to fit the networks. Full-resolution crops and other hidden layers as features are left to future work. · stated by the authors
  • The three fluorescence stains do not correspond to natural-image RGB channels, so any RGB mapping is arbitrary. · stated by the authors
  • There is one benchmark and one cell line (BBBC021, MCF7), with no variance estimates for small gaps such as 86% vs 84% and 91% vs 90%. The baseline is mean profiling, while Ljosa et al. report 94% for factor-analysis profiles on the same set. · noted in this summary
  • The timing comparison may not be like-for-like: CellProfiler is timed on a single core, but the paper does not say how many cores Inception-v3 used on the two 8-core Xeon CPUs. · noted in this summary

Read this if you run image-based compound screens and want a segmentation-free baseline that uses off-the-shelf ImageNet networks, evaluated on the BBBC021 mechanism-of-action benchmark.

Citations
83

Reproducibility ledger

Nobody has reported trying to reproduce this paper yet. Did you replicate an experiment or run the code? Your report saves the next lab a week.

Sign in to file a reproducibility report.

Related papers

Similarity Map
SelectedSame fieldCross-field
THISKensert '19Murthy '25Godinez '18Combe '23Yao '19Xun '24
69% semantic matchBioinformatics
Alexander Kensert, Philip John Harrison, Ola Spjuth·SLAS DISCOVERY·Jan 14, 2019· 165 citations