The first thing you learn to do as a scientist, or as a human being really, is observation. We learn from watching how things happen. One of the reasons why fields like molecular biology and quantum physics are so hard is that we can’t observe things directly. We are completely unable to simply look inside a living cell at molecular resolution, an ability that would resolve most of the open questions in molecular biology. So, why can’t we? The challenge lies at the intersection of really small physics and really fragile biology. These notes are about the methods we use to try and get around the limits as well as the novel methods in development by serious groups trying to explore this frontier.
The ideal outcome is:
A real-time, three-dimensional view of a living cell at atomic-level resolution, with each protein, nucleic acid, and lipid identifiable by structural features.
Why it’s Probably Impossible
High resolution means high energy, and energy damages cells.
You need extremely high resolution to identify molecules by shape. However, EM-level resolution means electron doses at levels that break chemical bonds. In this sense, it is analogous to the Heisenberg uncertainty problem, because you cannot observe a living cell without completely altering it. (In most cases, without killing it.)
Labeling everything is impossible.
Let’s say you didn’t need atomic-level resolution, and instead wanted to put fluorescent labels on every molecule and track those. This introduces the uncertainty problem analogy again –– you simply can’t attach unique reporters to thousands of species without perturbing the system you’re trying to observe. Even if you could attach unique labels to everything, you would run into issues with crowding causing changes in kinetics, including binding and trafficking.
Molecules inside the cell move really fast.
Even if imaging were suddenly harmless, microsecond-scale events need extremely high temporal resolution just to record them. The needs are beyond the scope of what EM can currently do in a single-pass. Ultrafast EM can capture femtosecond-nanosecond processes, but with limitations. You need fixed, ultra-thin samples (to avoid darkening from elastic scattering) and enormous doses spread over many identical events.
Modern Intracellular Methods
Today’s imaging ecosystem is an elaborate set of work-around methods to get us a good view of one particular facet of interest of the cell.
Cryo-EM and Cryo-ET category methods
Near-atomic resolution imaging with an electron beam, but all samples are frozen and extraordinarily thin to avoid the aforementioned elastic scattering problem. These techniques work well for structure, but are of course blind to dynamics.
Super-resolution fluorescence (STED, PALM/STORM, MINFLUX)
Nanometer-scale localization of individual molecules in living cells, but only the ones you can label, and you have to bet on how much you are disturbing the molecule with the labeling process. These are noteworthy because they get fluorescence microscopy past the (Abbe/Rayleigh) diffraction limit.
STED: nonlinear depletion that sharpens the effective excitation spot. (Imagine you’re trying to spotlight only a single person in a crowd, but you can’t make your spotlight thinner. STED uses a second light, shaped like a donut, to turn off the glow everywhere except a tiny point in the center.)
PALM/STORM: activating fluorophores sparsely in time, localizing one at a time for an image, then reconstructing a high-resolution image by stitching them together.
MINFLUX: using a movable illumination minimum to pinpoint single fluorophores. (It finds a molecule by moving a dark spot around and seeing where the molecule stops glowing. How close is the fluorophore tag to a point where the light intensity is exactly zero?)
Cryogenic Correlative Light and Electron Microscopy (Cryo-CLEM)
This method combines light microscopy, often using fluorescent labels, with the ultra-high resolution of Cryo-ET. The method starts with finding the target dynamic event in a living or near-living cell, flash-freezing the site, and then using a focused ion beam to sculpt the area into an electron-transparent window called a lamella. It still only captures a single snapshot in time, but it gives us information about particular dynamic processes, especially with multiple images.
Expansion microscopy (ExM)
ExM works around the “diffraction limit” by physically expanding the fixed sample itself, using a hydrogel that absorbs water. This pushes previously unresolvable structures apart, allowing them to be imaged using standard microscopes.
(This concept inspired a company funded by my team at the SALT Fund, to further develop a sort of “reversal” of the technique, called implosion fabrication, as a method for high-fidelity nano-manufacturing. The company is now focused on a different nano-manufacturing technique and actively creating complex 3D components for the optics and photonics industries.)
Lattice light-sheet and other volumetric methods
This method is great for whole-cell movies with low phototoxicity, but resolution is tens to hundreds of nanometers. It works by generating a thin, extended lattice-shaped sheet of light from the interference of multiple Bessel beams. This lowers phototoxicity dramatically and allows rapid volumetric scanning.
Spatial multiomics (smFISH, Visium HD, Digital Biology, CosMx, Xenium, MERFISH)
Spatial molecular mapping of fixed cells and tissues at ~100 nm-1 µm scale. These systems output hundreds to thousands of molecules paired with spatial coordinates. The chemistry varies, but all of these methods rely on barcoded probes to map identity from the image to the data. Here we again lose dynamics, but the data from these imaging systems are transforming how we think about the topology of cell state modeling.
Spinach 🥬
While not pushing the bounds of whole-cell live imaging, Spinach and subsequent (largely vegetable or fruit-themed) methods did change the game for the RNA imaging. By engineering an RNA aptamer (a small RNA folded into a specific shape) to fluoresce by binding a small-molecule fluorophore, the Jaffrey lab produced the first GFP-like reporter made of RNA. This meant that we could begin to view RNA in a live cell, and methods inspired by it are now used in some labs as part of the toolkit for deciphering RNA dynamics. The cover image for this post is from this paper, where they used a technique called Mango.
Integrated datasets
Projects like the Allen Institute Cell Atlas, Janelia cryo-ET initiatives, and several Arc Institute-led efforts combine imaging and/or multiomics into big data sets that can help us make helpful cell state models.
The Frontier & Ideas
Ensemble snapshot reconstruction
Freeze millions of cells, slice them, and take millions of EM resolution-level images. Reconstruct likely trajectories using statistics and ML models. This is analogous to scRNA-seq pseudotime. It seems a weak version of this (only 16 images of one structure) works: Cryo-ET of actin cytoskeleton and membrane structure in lamellipodia formation using optogenetics - PMC.
Low-dose imaging combined with ML reconstruction
Combine low-resolution live imaging like lattice light-sheet with high-resolution cryo-ET priors and generative models that infer detail under physical constraints.
Quantum-Enhanced Microscopy
A small research effort exists in using quantum optics, particularly squeezed light, to push imaging sensitivity below the normal shot noise limit. Shot noise comes from the randomness of photon arrival and limits how faint your light can be before the signal disappears. Squeezed light works by reshaping quantum uncertainty. It is done by deliberately increasing noise in one property of the light (amplitude) so you can decrease it in the one the microscope measures (phase). Warwick Bowen’s group has shown that this enables detection of the same phase shifts in live cells with fewer photons, reducing the ratio of photo toxicity to resolution. Specifically, they achieved a “14% resolution enhancement compared to experiments with coherent light.”
New physics?
History is full of impossible imaging resolution limits that are later broken. I think it is possible that this could be one such case.
Whole-cell generative models
If you have enough data, a generative model could theoretically learn to simulate whole-cell dynamics.
Some Groups Working on the Whole-Cell Frontier
HHMI Janelia: automated cryo-ET, MINFLUX, lattice light-sheet imaging.
EMBL, UCSF, Stanford cryo-ET groups: whole-cell tomography, lamella workflows, segmentation.
Allen Institute for Cell Science: integrated imaging, digital cells.
Max Planck Institute: MINFLUX, cryo-CLEM, high-throughput subtomogram averaging.
The Rosalind Franklin Institute: next-gen cryo-FIB/ET, sample prep, integrated imaging.
The Broad Institute of MIT and Harvard: morphological profiling (Cell Painting), large-scale single-cell and spatial omics platforms.
Chan Zuckerberg Biohub: large-scale imaging and spatial-omics integration.
Arc Institute: multimodal single-cell reference frameworks, digital cells.
10x Genomics, Vizgen, NanoString: spatial molecular profiling platforms.
DeepCell, Recursion, and similar ML-first companies: massive imaging & ML pipelines.
Closing Thoughts
We are much farther from the ideal, impossible outcome than these notes probably made it seem. I have hopes that biology’s habit of pushing past fundamental limits continues, because every step towards enabling observation at this scale has the potential to completely revolutionize our understanding of life. Actually reaching that outcome would mean we would be able to see intracellular kinetics, understand the functions hiding in the dark genome, and finally stop relying on partially subjective, inferred pathways. We could begin to fully model the cell and understand diseases of dysregulation, especially cancer. This is a worthy pursuit.
If we really do break physics’ limits, we could even sequence through imaging.