Aro: a machine learning approach to identifying single molecules and estimating classification error in fluorescence microscopy images
© Wu and Rifkin; licensee BioMed Central. 2015
Received: 19 December 2014
Accepted: 12 March 2015
Published: 27 March 2015
Recent techniques for tagging and visualizing single molecules in fixed or living organisms and cell lines have been revolutionizing our understanding of the spatial and temporal dynamics of fundamental biological processes. However, fluorescence microscopy images are often noisy, and it can be difficult to distinguish a fluorescently labeled single molecule from background speckle.
We present a computational pipeline to distinguish the true signal of fluorescently labeled molecules from background fluorescence and noise. We test our technique using the challenging case of wide-field, epifluorescence microscope image stacks from single molecule fluorescence in situ experiments on nematode embryos where there can be substantial out-of-focus light and structured noise. The software recognizes and classifies individual mRNA spots by measuring several features of local intensity maxima and classifying them with a supervised random forest classifier. A key innovation of this software is that, by estimating the probability that each local maximum is a true spot in a statistically principled way, it makes it possible to estimate the error introduced by image classification. This can be used to assess the quality of the data and to estimate a confidence interval for the molecule count estimate, all of which are important for quantitative interpretations of the results of single-molecule experiments.
The software classifies spots in these images well, with >95% AUROC on realistic artificial data and outperforms other commonly used techniques on challenging real data. Its interval estimates provide a unique measure of the quality of an image and confidence in the classification.
In the last decade, a host of new technologies for tagging and visualizing individual molecules have yielded unprecedented quantitative insight into the spatial and temporal dynamics of fundamental biological processes as varied as ligand-receptor interactions at the cell surface , protein localization to synaptic junctions , and incomplete penetrance . For example, the ability to visualize mRNA transcripts at the single molecule level without transgenic methods has led single-molecule fluorescence in situ hybridization (smFISH) to be widely used in studying gene expression in various organisms [3-14]. Recently this technique has been pushed to image up to 32 genes simultaneously with the promise of increasing this number still more . These microscopy-based techniques rely primarily on fluorescent proteins or dyes that are bound to the molecule of interest and appear as a bright, roughly Gaussian spot. Background fluorescence can be considerable for some of these techniques, including smFISH [4,8], which makes distinguishing signal from noise an image processing challenge. However, a statistically principled, automated, and robust method for analyzing the images and classifying local intensity maxima as signal or noise, and estimating the accuracy and variability of these classifications has not been developed. This problem is acute since highly sensitive microscopy methods like smFISH are ideally suited for quantitatively studying stochastic variation in gene expression and other molecular processes within a population.
We have extended a machine-learning pipeline for identifying, localizing, and counting biologically meaningful intensity maxima in 3D image stacks  both by improving the initial spot classification and, crucially, by providing a way to both estimate the quality of the data and generate an interval estimate for the number of molecules in it. We have tested it extensively on the challenging case of wide-field epifluorescence smFISH image stacks of nematode embryos where there can be substantial background fluorescence, and it also works on other samples like yeast and mammalian cell culture where the signal to noise ratio is more favorable. Unlike other commonly used methods [3,16,17], this software does not rely on arbitrary or user-defined parameters and cutoffs, but instead recognizes and classifies individual mRNA spots by measuring several features of local intensity maxima and classifying them with a supervised random forest classifier [18,19], It is a spot-centric approach as compared with approaches that involve thresholding an entire image [3,16,17].
Our training GUI allows a user to view spots from a subset of image stacks in the dataset, generate a manually curated training set by classifying them as true signal spots or noise, and build a forest of decision trees based on the features calculated from each spot (Figure 1, Additional file 1). We find that a training set consisting of a few hundred positive and negative examples is sufficient for stable classification. For each tree, the algorithm selects a bootstrapped sample from this training data. Each split in the tree is based on a randomly chosen subset of the statistics, and the tree is grown according to pre-specified stopping criteria. The leaves can be, but are not necessarily, comprised of a single class. At the end of training, the user has a bagged ensemble of decision trees.
To classify a new local maximum, the program runs the statistics for the putative spot through each tree to a terminal leaf. The proportion of training spots in this leaf that are manually classified as good can be used to estimate the probability that the new local maximum is a true signal spot. Although such probabilities are known to be inaccurate for single decision trees , using an ensemble of bagged trees improves the probability estimate, and so we average these proportions for a single candidate spot across all the trees in the forest to estimate a preliminary probability that it is a true spot [30-33]. However, these preliminary probabilities do not necessarily reflect the long-run frequency of a spot with particular features being classified as signal or noise [34-36].
The calibrated probability reflects uncertainty in the classification of any particular spot, and, consequently, can be used to measure the uncertainty in the count of the number of true spots in an image. A local intensity maximum with a particular preliminary probability can be thought of as a sample of size 1 from the population of all candidate spots with the same preliminary probability, of which some fraction (the calibrated probability) are true spots. We would like to estimate a confidence interval for the count of true spots in an image. The width of the confidence interval is a measure of the quality of an image because it will largely be driven by the fraction of spots for which the user himself or herself would be ambivalent, based on how he or she has classified similar spots in the training set.
By rerunning this model 1000 times, we can derive a confidence interval for the total spot number (T). This interval will be tight for high quality images and will widen as image quality degrades.
Estimating the variance of random forest and other bagged predictions is still an open problem, in part because the variance is comprised of both (a) sampling variance from training on a limited set of data and (b) Monte Carlo effects arising from a finite amount of bootstrapping [39-41]. Because we can empirically calibrate random forest probabilities in our classification task, we can take advantage of standard probability theory to construct an interval estimate.
Results and discussion
One key difficulty with evaluating image-based, molecule counting methods is that there is not an independent way to count the number of molecules in the specimen. For smFISH (as well as other techniques) it has been experimentally established [4,8] that, with the exception of transcriptional foci, spots in these images do represent single, fluorescently-labeled, diffraction-limited molecules. We can, however, use artificially generated data to investigate how well our method performs in the face of background noise.
In order to avoid making arbitrary assumptions about the structure of background noise, we used three 3D image stacks from actual specimens without any transcripts as the background. The background therefore consists of both autofluorescence and any diffuse fluorescence due to unbound probes that were not removed by washing. To generate signal, we sprinkled point sources of a specified magnitude throughout a blank image stack of the same size as the background stack, convolved them with a point spread function based on typical microscopy parameters, added the background and signal stacks together, and then blurred them with a Gaussian filter. The spots in these images look very much like actual data (Additional files 3, 4 and 5).
Another approach [43-45] has the primary goal of spot localization and starts by identifying individual candidate spots after intensity thresholding a 2D maximum projection, correcting for local background, and fitting them to 2D Gaussians. It then removes purported duplicates and thresholds a measure of the intensity of the entire spot to distinguish signal from noise. Because our algorithm also starts directly from the local maxima, it also works robustly for images with high or inhomogeneous backgrounds. However, it uses the 3D image, not a maximum projection, and is able to resolve clustered spots. Furthermore, while our local-maxima-centric approach uses a similar method for localization, its primary goal is robust classification without setting semi-arbitrary thresholds. The supervised learning process and the GUI allow the user to manually curate the classification of individual spots, and then feed these corrections back into the classification algorithm. This is particularly useful for low quality images, allowing the user to overrule the algorithm for spots on the boundary between signal and noise.
As the throughput of microscopy-based single-molecule techniques increases, robust image processing techniques will be ever more crucial. We present a machine-learning-based pipeline for identifying and classifying fluorescently labeled molecules in 3D image stacks that performs well under conditions where other algorithms fail. The software (called Aro Spot Finding Suite after Arothron hispidus) includes MATLAB  GUIs to generate the training set and review the classifications and a detailed manual with examples. The ability to infer biological meaning from a quantitative imaging experiment depends upon extracting reliable measurements from images. For single molecule imaging, our software uniquely provides a way to measure this reliability.
Availability and requirements
Project name: Aro Spot Finding Suite.
Project home page: https://gitlab.com/evodevosys/AroSpotFindingSuite.
Operating system: Platform independent.
Programming language: MATLAB.
Other requirements: MATLAB statistical toolbox; third-party MATLAB packages that are included with their own licenses with this distribution.
License: Apache 2.0.
We thank Dan Pollard and Larry Du for helpful feedback on the software. This work was supported by a Hellman fellowship to SAR and by NIH grants R01GM103782 and P50 GM085764.
- Sako Y, Minoghchi S, Yanagida T. Single-molecule imaging of EGFR signalling on the surface of living cells. Nat Cell Biol. 2000;2:168–72.View ArticlePubMedGoogle Scholar
- Dani A, Huang B, Bergan J, Dulac C, Zhuang X. Superresolution imaging of chemical synapses in the brain. Neuron. 2010;68:843–56.View ArticlePubMedPubMed CentralGoogle Scholar
- Raj A, Rifkin SA, Andersen E, van Oudenaarden A. Variability in gene expression underlies incomplete penetrance. Nature. 2010;463:913–8.View ArticlePubMedPubMed CentralGoogle Scholar
- Raj A, van den Bogaard P, Rifkin SA, van Oudenaarden A, Tyagi S. Imaging individual mRNA molecules using multiple singly labeled probes. Nat Methods. 2008;5:877–9.View ArticlePubMedPubMed CentralGoogle Scholar
- Neuert G, Munsky B, Tan RZ, Teytelman L, Khammash M, van Oudenaarden A. Systematic identification of signal-activated stochastic gene regulation. Science. 2013;339:584–7.View ArticlePubMedPubMed CentralGoogle Scholar
- Bumgarner SL, Neuert G, Voight BF, Symbor-Nagrabska A, Grisafi P, van Oudenaarden A, et al. Single-cell analysis reveals that Noncoding RNAs contribute to clonal heterogeneity by modulating transcription factor recruitment. Mol Cell. 2012;45:470–82.View ArticlePubMedPubMed CentralGoogle Scholar
- Darzacq X, Yao J, Larson DR, Causse SZ, Bosanac L, de Turris V, et al. Imaging transcription in living cells. Annu Rev Biophys. 2009;38:173–96.View ArticlePubMedPubMed CentralGoogle Scholar
- Femino AM, Fay FS, Fogarty K, Singer RH. Visualization of single RNA transcripts in situ. Science. 1998;280:585–90.View ArticlePubMedGoogle Scholar
- Ji N, Middelkoop TC, Mentink RA, Betist MC, Tonegawa S, Mooijman D, et al. Feedback control of gene expression variability in the Caenorhabditis elegans Wnt pathway. Cell. 2013;155:869–80.View ArticlePubMedGoogle Scholar
- Barkoulas M, van Zon JS, Milloz J, van Oudenaarden A, Félix M-A. Robustness and Epistasis in the C. elegans vulval signaling network revealed by pathway dosage modulation. Dev Cell. 2013;24:64–75.View ArticlePubMedGoogle Scholar
- Lubeck E, Cai L. Single-cell systems biology by super-resolution imaging and combinatorial labeling. Nat Methods. 2012;9:743–8.View ArticlePubMedPubMed CentralGoogle Scholar
- Mohn F, Sienski G, Handler D, Brennecke J. The rhino-deadlock-cutoff complex licenses Noncanonical transcription of dual-strand piRNA clusters in Drosophila. Cell. 2014;157:1364–79.View ArticlePubMedGoogle Scholar
- Oliveira JTD, Matos AJD, Barros R, Ribeiro C, Chen A, Hespanhol V, et al. Differential expression of Galectin-1 and Galectin-3 in canine non-malignant and malignant mammary tissues and in progression to metastases in mammary tumors. Anticancer Res. 2014;34:2211–21.Google Scholar
- Whitehead CL, Walker SP, Ye L, Mendis S, Kaitu’u-Lino TJ, Lappas M, et al. Placental specific mrna in the maternal circulation are globally dysregulated in pregnancies complicated by fetal growth restriction. J Clin Endocrinol Metab. 2013;98:E429–36.View ArticlePubMedGoogle Scholar
- Rifkin SA. Identifying fluorescently labeled single molecules in image stacks using machine learning. Methods Mol Biol. 2011;772:329–48.View ArticlePubMedGoogle Scholar
- Mueller F, Senecal A, Tantale K, Marie-Nelly H, Ly N, Collin O, et al. FISH-quant: automatic counting of transcripts in 3D FISH images. Nat Methods. 2013;10:277–8.View ArticlePubMedGoogle Scholar
- McIsaac RS, Silverman SJ, Parsons L, Xu P, Briehof R, McClean MN, et al. Visualization and analysis of mRNA molecules using fluorescence in situ hybridization in Saccharomyces cerevisiae. J Vis Exp. 2013;76:e50382.Google Scholar
- MATLAB. Version 7.10.0 (R2010a). Natick, Massachusetts: The MathWorks Inc.; 2010.Google Scholar
- Breiman L. Random forests. Mach Learn. 2001;45:5–32.View ArticleGoogle Scholar
- Hastie T, Tibshirani R, Friedman J. The elements of statistical learning: data mining, inference, and prediction. 2nd ed. New York, NY: Springer; 2011.Google Scholar
- Flach P. Machine learning: the art and science of algorithms that make sense of data. Cambridge. New York: Cambridge University Press; 2012.View ArticleGoogle Scholar
- Caruana R, Karampatziakis N, Yessenalina A. An empirical evaluation of supervised learning in high dimensions. In: Proceedings of the 25th International Conference on Machine Learning. New York: ACM; 2008. p. 96–103.Google Scholar
- Booth A, Gerding E, McGroarty F. Predicting equity market price impact with performance weighted ensembles of random forests. In: 2104 IEEE Conference on Computational Intelligence for Financial Engineering Economics (CIFEr). London: IEEE; 2014. p. 286–93.Google Scholar
- Tüselmann H, Sinkovics RR, Pishchulov G. Towards a consolidation of worldwide journal rankings – a classification using random forests and aggregate rating via data envelopment analysis. Omega. 2015;51:11–23.View ArticleGoogle Scholar
- Cutler DR, Edwards TC, Beard KH, Cutler A, Hess KT, Gibson J, et al. Random forests for classification in ecology. Ecology. 2007;88:2783–92.View ArticlePubMedGoogle Scholar
- Chen X, Ishwaran H. Random forests for genomic data analysis. Genomics. 2012;99:323–9.View ArticlePubMedPubMed CentralGoogle Scholar
- Verikas A, Gelzinis A, Bacauskiene M. Mining data with random forests: a survey and results of new tests. Pattern Recogn. 2011;44:330–49.View ArticleGoogle Scholar
- Fanelli G, Dantone M, Gall J, Fossati A, Gool LV. Random forests for real time 3D face analysis. Int J Comput Vis. 2012;101:437–58.View ArticleGoogle Scholar
- Gall J, Razavi N, Gool LV. An introduction to random forests for multi-class object detection. In: Dellaert F, Frahm J-M, Pollefeys M, Leal-Taixé L, Rosenhahn B, editors. Outdoor and large-scale real-world scene analysis. Heidelberg: Springer; 2012. p. 243–63.View ArticleGoogle Scholar
- Provost F, Domingos P. Tree induction for probability-based ranking. Mach Learn. 2003;52:199–215.View ArticleGoogle Scholar
- Malley JD, Kruppa J, Dasgupta A, Malley KG, Ziegler A. Probability machines: consistent probability estimation using nonparametric learning machines. Methods Inf Med. 2012;51:74–81.View ArticlePubMedGoogle Scholar
- Biau G. Analysis of a random forests model. J Mach Learn Res. 2012;13:1063–95.Google Scholar
- Kruppa J, Schwarz A, Arminger G, Ziegler A. Consumer credit risk: individual probability estimates using machine learning. Expert Syst Appl. 2013;40:5125–31.View ArticleGoogle Scholar
- Gebel M, Weihs C. Calibrating classifier scores into probabilities. In: Decker PDR, Lenz PDH-J, editors. Advances in data analysis. Heidelberg: Springer; 2007. p. 141–8.View ArticleGoogle Scholar
- Zadrozny B, Elkan C. Obtaining calibrated probability estimates from decision trees and naive Bayesian classifiers. In Proceedings of the Eighteenth International Conference on Machine Learning; 2001. p. 609–616.Google Scholar
- Niculescu-mizil A, Caruana R. Predicting good probabilities with supervised learning. In: Proceedings of the 22th International Conference on Machine Learning. New York: ACM; 2005. p. 625–32.Google Scholar
- Platt JC. Probabilistic outputs for support vector machines and comparisons to regularized likelihood methods. In: Advances in Large Margin Classifiers. Cambridge: MIT Press; 1999. p. 61–74.Google Scholar
- Feller W. An introduction to probability theory and its applications, vol. 1. Sydney: J. Wiley & sons; 1968.Google Scholar
- Duan J. Bootstrap-based variance estimators for a bagging predictor. In: PhD thesis. Chapel Hill: North Carolina State University; 2011.Google Scholar
- Sexton J, Laake P. Standard errors for bagged and random forest estimators. Comput Stat Data Anal. 2009;53:801–11.View ArticleGoogle Scholar
- Wager S, Hastie T, Efron B. Confidence intervals for random forests: the jackknife and the infinitesimal jackknife. J Mach Learn Res. 2014;15:1625–51.PubMedPubMed CentralGoogle Scholar
- Tan RZ, van Oudenaarden A. Transcript counting in single cells reveals dynamics of rDNA transcription. Mol Syst Biol. 2010; doi:10.1038/msb.2010.14.Google Scholar
- Trcek T, Chao JA, Larson DR, Park HY, Zenklusen D, Shenoy SM, et al. Single-mRNA counting using fluorescent in situ hybridization in budding yeast. Nat Protoc. 2012;7:408–19.View ArticlePubMedPubMed CentralGoogle Scholar
- Thompson RE, Larson DR, Webb WW. Precise nanometer localization analysis for individual fluorescent probes. Biophys J. 2002;82:2775–83.View ArticlePubMedPubMed CentralGoogle Scholar
- Zenklusen D, Larson DR, Singer RH. Single-RNA counting reveals alternative modes of gene expression in yeast. Nat Struct Mol Biol. 2008;15:1263–71.View ArticlePubMedPubMed CentralGoogle Scholar
This is an Open Access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/4.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly credited. The Creative Commons Public Domain Dedication waiver (http://creativecommons.org/publicdomain/zero/1.0/) applies to the data made available in this article, unless otherwise stated.