← Blog

Spatial Biology Has a Budget

J. Lee

Spatial biology maps which molecules sit where, cell by cell, in intact tissue. One family of methods captures molecules on barcoded bins of a slide and reads them out by bulk sequencing. The other, our subject here, photographs targets in place: single molecules or amplified copies that show up as points of light.

Those photographs are stunning. They’re also governed by optics, chemistry, and statistics, three fields that don’t negotiate. The numbers below are rough and illustrative. They apply to methods that count individual spots, RNA or protein: think smFISH [1, 2] or amplicon-based in situ sequencing [3].

In a common scheme, each target gets a code on its probe or antibody. The tissue is imaged over several rounds; in each round a spot lights up or stays dark. Read the on/off pattern across rounds, Morse-style, and you know the target [2].

The cell is a small parking lot

Spots too close together blur into one. How close depends on the light and on the lens’s numerical aperture (N.A.). At N.A. 0.8, a large cell in a standard 5 µm slice can fit roughly 5,000–10,000 spots; smaller cells offer fewer. You can’t fill them all: light only about a tenth of the slots per round, and molecules cluster in organelles, membranes, and granules. Net: roughly 500–1,000 readable spots per round. Sharper optics or physically expanding the tissue [8] can buy more room, but published RNA counts still run from hundreds per cell in clinical samples to about a thousand in optimized setups.

Each spot must outshine the tissue

A dye molecule emits a fixed budget of photons, roughly 100,000 to a million, before it burns out [5]. Most miss the lens: an N.A. 0.8 objective catches about 10%. Meanwhile the tissue glows, especially wax-embedded clinical samples. A spot needs a few hundred detected photons to stand above that flicker.

One dye can deliver that quickly [5], but not when the tissue refuses to stay quiet. Methods pile many dyes onto each target: 30–50 probes per RNA in smFISH [1], dye-depositing antibodies for proteins, or an amplicon carrying hundreds of dyes [3]. The challenge isn’t sensitivity alone; it’s a signal that beats the glare.

Imaging one slide takes hours to days

Each image takes about a tenth of a second, until you stack them. A 1-cm² section needs about 1,000 tiles at 40×, and several thousand at many 60× setups, each at a few depths and colors. Repeat for every round, eight to several dozen [2], each with 30–60 minutes of chemistry. An illustrative setup is 10–12 hours for one region; many runs take days. That caps how many slides, and patients, a study can afford.

The allure of ‘unbiased’

A cell holds on the order of 100,000 protein-coding RNAs and billions of proteins [4]. A thousand barcoded objects is a few percent of the RNA and a vanishing sliver of the protein, which is why protein imaging usually measures brightness instead of counting [9]. Detection is imperfect: at 20% efficiency, a target with 10 copies per cell reads zero in about one cell in ten. Abundant housekeeping molecules can swamp the read budget, and blurry borders let neighbors’ signals wander in [6].

What counts is the number of patients, not cells

A million cells from three people is still N = 3 when you compare groups of patients. Statistics is unmoved by your cell count. What matters is effect size, d: how big the group difference is relative to patient-to-patient variation.

Lehr’s rule of thumb puts the need at about 16/d² patients per group [7]: 16 per group for a large effect (d = 1) and 64 for a medium one (d = 0.5). Testing many targets raises the bar. Time sets cost, and cost sets how many patients you can afford. A medium-effect study, one slide per patient, could need 128 slides, about 1,400 instrument hours. Physics sets the ceiling, statistics the floor, and your budget what you can claim.

Deep or wide

None of this is new. Together, physics and statistics force a choice: go deep, with sharper optics and more targets per cell, or go wide, with more patients. Every budget buys more of one than the other.

Today’s default is to treat spatial data as a hypothesis generator: run deep on a few samples, test the best leads with faster assays across many patients, and validate the winners. AI may tighten that loop. More money on the same measurement will not conjure whispers hidden in the glare. Before trusting a beautiful tissue map, ask: How many molecules do you need to decide? How many did you actually see? How many patients stand behind it?

Sources

  1. Raj A, et al. Imaging individual mRNA molecules using multiple singly labeled probes. Nature Methods 5, 877–879 (2008).

  2. Chen KH, et al. Spatially resolved, highly multiplexed RNA profiling in single cells. Science 348, aaa6090 (2015).

  3. Lee JH, et al. Highly multiplexed subcellular RNA sequencing in situ. Science 343, 1360–1363 (2014).

  4. Milo R, Phillips R. Cell Biology by the Numbers. Garland Science (2015).

  5. Dempsey GT, et al. Evaluation of fluorophores for optimal performance in localization-based super-resolution imaging. Nature Methods 8, 1027–1036 (2011).

  6. Petukhov V, et al. Cell segmentation in imaging-based spatial transcriptomics. Nature Biotechnology 40, 345–354 (2022).

  7. Lehr R. Sixteen S-squared over D-squared: a relation for crude sample size estimates. Statistics in Medicine 11, 1099–1102 (1992).

  8. Chen F, Tillberg PW, Boyden ES. Expansion microscopy. Science 347, 543–548 (2015).

  9. Goltsev Y, et al. Deep profiling of mouse splenic architecture with CODEX multiplexed imaging. Cell 174, 968–981 (2018).