Neuron count from cell-based
drug screen images
Vladimir Morozov, ALS-TDI
October , 2014
Cell Profiler pipeline
1. Thresholding
2. Nuclei
identification
3. Neurite
attachment
4. Final
neurons
Making training set for detecting healthy neurons
•We consider good looking cells as
positive
•Microphages were identified as
negative examples
•Other non-neuron cells (e.g. astrocytes)
were NOT specified as negative
CellProfiler rule-based classifier, ~77% accuracy
IF (Nuclei_Intensity_IntegratedIntensity_ch1 >
6.1516400000000004, [0.36232803790685297, -
0.36232803790685297], [-0.80075679385436416,
0.80075679385436416])
IF (Neurite_AreaShape_Solidity > 0.52507099999999995, [-
0.68639735377870881, 0.68639735377870881],
[0.25029235002592609, -0.25029235002592609])
IF (Nuclei_Intensity_StdIntensity_ch2 > 0.00113378,
[0.11906406640858008, -0.11906406640858008], [-
0.65599426031103425, 0.65599426031103425])
IF (Neurite_Intensity_StdIntensity_ch1 > 0.0028765700000000002,
[0.042546288651953375, -0.042546288651953375], [-
0.88894769206543311, 0.88894769206543311])
IF (Neurite_Intensity_MassDisplacement_ch1 >
3.5285700000000002, [0.08250012593027499, -
0.08250012593027499], [-0.50046974611076089,
0.50046974611076089])
CellProfiler classifier on the picked image
Manually created training set
•We consider good looking cells as
positive
•Microphages were identified as
negative examples
•Other non-neuron cells (e.g. astrocytes)
were NOT specified as negative
Built own classifier in R
Using the training set I tried
• Yeo-Johnson feature transformation (can
handle negative values) to make them more
normally distributed
• 4-5 non-linear (rules, trees, boosting) classifier
algorithms
• The best performance ,85% accuracy with the
Multivariate Adaptive Regression Splines
(MARS,”earth”) algorithm. This model was
used for the final neuron classification
Toxic compounds
•Validity of the neuron detection pipeline is confirmed by compounds
with the largest cell toxic effect
•These compounds are known cytotoxic agents
Protective compounds
•Statically significant protective compounds were identified
•These compounds don’t show statistical enrichment for specific
pharmacological or structural classes or mechanism of action

Neuron detection and counting in high-throughput screening images

  • 1.
    Neuron count fromcell-based drug screen images Vladimir Morozov, ALS-TDI October , 2014
  • 2.
    Cell Profiler pipeline 1.Thresholding 2. Nuclei identification 3. Neurite attachment 4. Final neurons
  • 3.
    Making training setfor detecting healthy neurons •We consider good looking cells as positive •Microphages were identified as negative examples •Other non-neuron cells (e.g. astrocytes) were NOT specified as negative
  • 4.
    CellProfiler rule-based classifier,~77% accuracy IF (Nuclei_Intensity_IntegratedIntensity_ch1 > 6.1516400000000004, [0.36232803790685297, - 0.36232803790685297], [-0.80075679385436416, 0.80075679385436416]) IF (Neurite_AreaShape_Solidity > 0.52507099999999995, [- 0.68639735377870881, 0.68639735377870881], [0.25029235002592609, -0.25029235002592609]) IF (Nuclei_Intensity_StdIntensity_ch2 > 0.00113378, [0.11906406640858008, -0.11906406640858008], [- 0.65599426031103425, 0.65599426031103425]) IF (Neurite_Intensity_StdIntensity_ch1 > 0.0028765700000000002, [0.042546288651953375, -0.042546288651953375], [- 0.88894769206543311, 0.88894769206543311]) IF (Neurite_Intensity_MassDisplacement_ch1 > 3.5285700000000002, [0.08250012593027499, - 0.08250012593027499], [-0.50046974611076089, 0.50046974611076089])
  • 5.
  • 6.
    Manually created trainingset •We consider good looking cells as positive •Microphages were identified as negative examples •Other non-neuron cells (e.g. astrocytes) were NOT specified as negative
  • 7.
    Built own classifierin R Using the training set I tried • Yeo-Johnson feature transformation (can handle negative values) to make them more normally distributed • 4-5 non-linear (rules, trees, boosting) classifier algorithms • The best performance ,85% accuracy with the Multivariate Adaptive Regression Splines (MARS,”earth”) algorithm. This model was used for the final neuron classification
  • 8.
    Toxic compounds •Validity ofthe neuron detection pipeline is confirmed by compounds with the largest cell toxic effect •These compounds are known cytotoxic agents
  • 9.
    Protective compounds •Statically significantprotective compounds were identified •These compounds don’t show statistical enrichment for specific pharmacological or structural classes or mechanism of action