Proc Natl Acad Sci U S A. 2009 Feb 10;106(6):1826-31. Epub 2009 Feb 2.
Scoring diverse cellular morphologies in image-based screens with iterative feedback and machine learning.
Jones TR, Carpenter AE, Lamprecht MR, Moffat J, Silver SJ, Grenier JK, Castoreno AB, Eggert US, Root DE, Golland P, Sabatini DM.


Many biological pathways were first uncovered by identifying mutants with visible phenotypes and by scoring every sample in a screen via tedious and subjective visual inspection. Now, automated image analysis can effectively score many phenotypes. In practical application, customizing an image-analysis algorithm or finding a sufficient number of example cells to train a machine learning algorithm can be infeasible, particularly when positive control samples are not available and the phenotype of interest is rare. Here we present a supervised machine learning approach that uses iterative feedback to readily score multiple subtle and complex morphological phenotypes in high-throughput, image-based screens. First, automated cytological profiling extracts hundreds of numerical descriptors for every cell in every image. Next, the researcher generates a rule (i.e., classifier) to recognize cells with a phenotype of interest during a short, interactive training session using iterative feedback. Finally, all of the cells in the experiment are automatically classified and each sample is scored based on the presence of cells displaying the phenotype. By using this approach, we successfully scored images in RNA interference screens in 2 organisms for the prevalence of 15 diverse cellular morphologies, some of which were previously intractable.

This article describes a machine-learning algorithm for researchers to manually train software to distinguish “positive” from “negative” cells on some particular phenotype of interest.  It seems that the main advance here is that this method works even when very few positive cells can be found.  Machine learning to distinguish phenotypes is easy when you can offer the software a positive control image with 100 cells, every one of which exhibits the phenotype of interest, and a negative control image with 100 cells, every one of which does NOT exhibit the phenotype of interest.  But if you’ve got a sample where even the positive sample only has about 5% of cells actually exhibiting the phenotype, then traditional methods would fall short; this new method aims to fill that gap and prevent overfitting due to cherry-picking the best examples of a phenotype:

Finding positive cells is straightforward when positive control samples are available and most of the cells therein show the phenotype. However, when this is not the case, as in classic exploratory screens, finding a sufficient number of positive cells can be prohibitively difficult. Even when positive control samples are available, using positive example cells from only those samples can lead to inaccurate scoring because of overfitting of the machine learning algorithm.

The upshot of all this is that the new algorithm proved highly successful– 12 out of the 14 phenotypes they looked at could be accurately scored without even adding any custom image features to the set already offered by CellProfiler.  Of the other 2 phenotypes, one was abandoned for lack of enough positive cells and one was scored successfully only with the addition of a new custom module to CellProfiler which measures the angle between a nucleus’s two nearest neighbors.

The new algorithm is now on offer under the name of Classifier as a part of CellProfiler Analyst.  To clarify: CellProfiler does image segmentation and feature recognition, while CellProfiler Analyst takes the numerical outputs of CellProfiler and helps you makes sense of them.

In terms of relevance to a potential prion therapeutic screening, this seems like something good to keep in mind.  A major challenge we need to overcome in such a screening is to accurately identify neurons and astrocytes and then score meaningful features on them, and this does not directly address that issue.  But there is clearly some good AI in this software which could theoretically be adapted to help teach CellProfiler how to recognize astrocytes if it cannot already.  It’s also possible that our initial screen to find ways to induce disease state in D178N 129M cells will not find any way to induce disease state across all cells (or all cells of a cell type) in the sample but rather will induce just some cells to take on disease state.  If the percentage of cells exhibiting prion formation turns out to be low, then Classifier could have a useful role in helping us recognize that disease phenotype as well.