Nobel laureate Eric Betzig says microscopy is stuck on incremental tweaks when it needs real breakthroughs. His two big goals: see proteins without labels, and watch all the molecules in a process at once instead of a few. The real problem now is data, not hardware—experiments generate tens of terabytes a day, and nobody can store or analyze it.
What's the big thing microscopy has to do? What, if you were to invent the impossible microscope, and nothing's impossible, just so one day, what would that microscope do?
All right, so I keep talking, there's two sides of this. There's what is needed right now and certainly doable, and then there's the holy grails, okay? To me, the holy grails, the things that would potentially pull me back into microscopy would be to have protein-specific contrast, label-free, at high resolution, noninvasively. Okay, I don't know how to do it. All right, the other option, the other holy grail would be, I don't care whether it's by fluorescence or whatever, but to be able to see all the molecular players involved in a biological process at the same time, not just one at a time, two at a time, three at a time. So the whole endocytic pathway, everything from clathrin, adaptor proteins, RAB, you know, the whole thing to see the whole mechanism going on at once. All right. Those are the holy grails. I don't know how to get to those holy grails. Extremely narrow line width labels with orthogonal labeling of 30 things, would get you there, be chemistry, crazy nightmare to get to that. But maybe there are better ways. I don't know. In terms of what the field needs now. So, yeah, this is going to piss off a lot of people. My feeling is, so one of the reasons I want to get out of microscopy is I feel that fields of technology technological development go through cycles just like a farmer's field. There are fertile times after you've put in the nitrogen-fixing plants and you can grow your corn, but after a while, you've grown enough corn and it needs sort of a regeneration period. I think you look at something like the mosaic, there's nothing technically novel in it at all. It's a lot of engineering, but there's nothing different. Same with the lattice A, it was combining two microscopes we already had, right? It's necessary to do one like that, but it's not, you know, I feel like lattice light sheet and PALM, those were step changes, right? In terms of same thing with near-field, when I did near-field, those were step changes, okay? The rest is just, you know, throwing the ingredients together in different permutations. Microscopy in terms of the hardware development to me is increasingly uninteresting because I think it needs to— Again, one of the things that was amazing about starting to work with Harold in 2005, when we had the idea for Palm, but at that time, I didn't even know what a fucking EMCCD camera was because I had never seen one. I used to work with an argon ion laser when I was doing near-field, it was 8 feet long. When I saw that you could get these diode pump solid-state lasers, I felt like Rip Van Winkle. And all the filters, all like the Simrad filters and like that, the filters back then sucked in the '90s. Okay? So it was like, oh my God, this is gonna be so trivial. There's all this technology. So again, you know, things, things get developed. And the beautiful thing about technology is you never know how it can be combined with all the other pieces of technology that are around to do new things. But right now, I feel like every time there's a new tiny little thing, whether it's a new mirror or a new wrinkle, everybody in the field's, you know, so it's not that interesting. So to me, the frontiers of microscopy today that are doable, one is I'm very enamored and impressed by what people have been doing with deep learning. And I think there's, it's very, very dangerous. You've got to have a fucking good training set and you have to validate it. And I'm really afraid there's going to be a lot of garbage because people are not going to do the right type of validation of what they're doing. But when applied right, I can really see how that can make a big jump. Basically, you're using priors, right? I mean, Palm is kind of using a prior, right? Palm is saying, I prior know that this is a single molecule and not more than one. Priors can take you a long way. So I think that's got a lot of wheels to go. The other problem in the hard intractable problem and the thing that depresses me about STILL Lattice Light Sheet and lattice AO is the data. I mean, everybody knows this. Okay. I mean, with the modern— I've got on my computer now, I'm trying to render a 24 terabyte dataset. Right. I got a $30,000 workstation, we have a million-dollar cluster that it's hooked to, and it's still an enormous pain in the ass to deal with the data. You can easily get 10 terabytes a day, no problem. You can get 20 terabytes a day. This brings up— I keep going on these tangents, but it reminded me of— I watched parts of the one you did with with Jeff Lichtman at Harvard. And so you had a discussion—
You're just rubbish saying ultrastructure is nothing.
Well, no, no, no. But he had an interesting discussion at the end of that about the value of the connectome. I have always been very suspicious about the value of the connectome. It's a very complete set of information, but even if it were not, Again, you've got tens of thousands of connections from every neuron, and you've got hundreds of millions of neurons. Even if there were some way of mapping all of that out, can we ever really understand what it means? I really worry even at the single-cell level and at the sort of chemical networks, biochemical networks that exist, and how much I believe that stochastic processes I think Brownian motion is the real engine of the cell. It's not ATP in the mitochondria, that it's the organization of the cell happens because of Brownian motion and differential sticking coefficients. It organizes in the same way that gravity organizes galaxies and the universe. Our ability to understand even a bacterium, I question. And a eukaryotic cell, it's going to be a challenge. So, this is something that science is bumping up against is maybe we can even put it all in a computer and have an AI replicate what's going on, but that doesn't mean we understand it. That doesn't mean it's reducible to principles that we can really put our hands around. Yeah. But the data problem itself is just gigantic. Mm-hmm.
