Most expression analyses tend to focus on a single parameter: transcript abundance. But cells are not homogeneous bags of molecules; there’s order in the chaos, and location matters.
Single-molecule fluorescence in situ hybridization (smFISH) is the one gene expression analysis method capable of reporting both transcript abundance and spatial localization. But until now, smFISH hasn’t been amenable to genome-scale analysis. Now, researchers at the University of Zurich report in Nature Methods a strategy for adding smFISH to the genomics toolbox. 
Lucas Pelkmans, Ernst Hadorn Chair of the Institute of Molecular Life Sciences at the University of Zurich, who led the study, explained that traditional smFISH suffers from two key shortcomings that preclude its use in high-throughput, genome-scale workflows. First, it produces a relatively weak signal and a poor signal-to-noise ratio. Second, the technique isn’t scalable since it requires high magnification, long imaging times, and optimization of conditions for each transcript under study.
Traditional smFISH detects transcripts using a series of short oligonucleotides, each labeled with a single fluorophore, to coat the transcript. The result is a sufficient concentration of fluorescent molecules on the mRNA to produce a visible, countable point of light, but usually only at 60× or 100× magnification and using an oil immersion, high numerical aperture lens.
Pelkmans and his team based their approach on branched DNA. In this strategy, 15 pairs of oligonucleotide probes target a given mRNA. These probes serve as landing pads for a primary “preamplifier probe,” which in turn provides binding sites for several secondary “amplifier probes,” and finally, a large set of tertiary labeled probes.
“It’s like building 15 trees of hybridization probes on a transcript,” Pelkmans explained. This approach produced an amplified signal that enabled imaging in a fraction of the time of the traditional method, was 100-times brighter, and yielded a 3-fold higher signal-to-noise ratio.
Leveraging that intense signal, the team then automated the process. They built a library of branched DNA probes for 928 human genes and tested them on cultured HeLa cells in 384 well plates, applying one probe per well and using identical hybridization and wash conditions. Following hybridization, they imaged about 10,000 cells per well at a relatively low 40× magnification without oil and then used home-built algorithms to extract not just transcript abundance but also spatial features—for instance, whether the spots were near the membrane or the nucleus, concentrated or dispersed, and so on. In total, they collected 18 spatial variables and used a computer cluster to assess their relation to gene regulation.
The data suggested that genes exhibiting similar spatial signatures, and similar cell-to-cell variability in those signatures, were more likely to share functional roles than genes whose only commonality was abundance.
“It turns out that those [spatial] features are actually more informative or more predictive for predicting functional interactions between genes,” Pelkmans said.
As Pelkmans explained, it is well established that transcription at the single-cell level is “noisy.” That is, two genetically identical cells might have significantly different numbers of a particular transcript floating in the cytoplasm.
The question is how to reconcile that observation with the fact that, as Pelkmans put it, “Biological processes are very robust to such noise.” In other words, protein levels don’t necessarily reflect transcript abundance. One possibility, he said, is that this robustness stems from regulating the subcellular location of mRNA, and his team’s data support that idea. Now their goal is to test that hypothesis directly to determine whether “spatial organization of transcripts acts to buffer against noise in transcription.”
Timothee Lionnet, a Project Scientist at the Howard Hughes Medical Institute Janelia Farms Research Campus, who has authored an in-press News & Views commentary about the Pelkmans study for Molecular Systems Biology, calls the study an “automation tour-de-force.”
“They’re bringing the FISH technology to the throughput that we’re used to seeing with multiplexed PCR or RNA-seq techniques,” he said.
One obvious extension of the technique according to Lionnet, is multiplexing to detect multiple transcripts per cell. Two companies that commercialize branched DNA technology (Affymetrix and Advanced Cell Diagnostics) already offer multiplexable FISH probes. And CalTech researcher Long Cai has described an alternative smFISH method that can detect some 32 distinct transcripts per cell. 
Pelkmans’ team observed good concordance with RNA-seq data in general, but not at very high per-cell transcript counts, where they observed a “ceiling effect” that tended to underreport abundance. The technique also failed to detect nuclear transcripts efficiently, though the team was able to mitigate that effect by adding acetic acid to the fixation buffer.
For Lionnet’s studies of transcription in the nucleus, that latter shortcoming is “a deal-breaker.” Still, he says, by solving FISH’s throughput problem, the new method gives in situ hybridization an edge over other single-cell approaches such as RNA-seq because it combines transcript abundance with spatial information.
“This is one step in the direction of getting a genome-wide FISH technique, and that would be a very transformative technology,” he said.
 N. Battich et al., “Image-based transcriptomics in thousands of single human cells at single-molecule resolution,” Nature Methods, Oct. 6, 2013. DOI:10.1038/nmeth.2657.
 J. Weaver, “Super-resolution barcoding: Turning single cells into microarrays,” BioTechniques, Sept. 12, 2012. [http://www.biotechniques.com/news/Super-Resolution-Barcoding-Turning-Single-Cells-into-Microarrays/biotechniques-335012.html]