Freenome is the company, they're doing interesting work but it's still early days. This isn't actually a very new idea, it's been floating around for years in academia (I've personally been hearing it about it for at least 5 years). The hard part isn't really the AI per se, it's getting enough DNA from the circulating copies (i.e. sample prep) and being lucky enough that the cancer you're trying to diagnose actually puts out enough DNA for you to measure at an early enough stage to be clinically relevant. In other words, if you can only get DNA when it's at stage 4 and metastatic, your DNA test isn't actually very useful. If you can get stage 1 cancers, now you're talking, but in those cases tumors tend to shed a lot less DNA (among other reasons because they're smaller).
A very good friend of mine is their CSO, I think they've got a really cool idea and could indeed go far, but it's a fiendishly hard problem.
As someone who has zero experience but is extremely interested, could you ELI5 why it's difficult to detect whether some DNA is cancerous?
In other words,
The hard part isn't really the AI per se, it's getting enough DNA from the circulating copies (i.e. sample prep)
It sounds like there's a way to detect "This DNA is definitely cancerous," but the problem is that the test only works when you give it enough DNA. My question is, where does this "enough" criteria come from? Why doesn't it work with a single cell? (I know there are obvious reasons, but I don't know what they are.)
These are all good questions, and full disclosure I know only the basics in this area. DNA can be identified as cancerous for a variety of reasons, though typically what you're looking for is some mutation you know to exist in cancer. This could be a single point mutation (insertion/deletion/change of a given DNA base pair), a loss/gain/modification of several adjacent base pairs, or whole scale genome rearrangements (deletion, insertion, inversion, replication) that will result in of two pieces of DNA being adjacent although you know they shouldn't be next to one another (see https://en.wikipedia.org/wiki/Philadelphia_chromosome for a quite well known example). You could also look for copy number variation (gene X has way more or way fewer copies than expected). I'm probably missing some as well, but those are the biggies.
To detect these, you're going to use either something like microarrays, QPCR, or DNA sequencing. Today people talk almost exclusively about high-throughput sequencing (HTS) because it has gotten quite cheap, gives you large amounts of data, can be made massively parallel and isn't as hindered by needing to know very precisely what you're looking for (microarrays and QPCR require somewhat more targeting; HTS also needs you to know what you're looking for, but not in quite the same ways).
In all of those cases you need a sample of DNA (duh). These can all be made to work if you have a single cell in hand, it's tricky but sample prep does exist for single cells at a time. Trouble is, you need that cell in hand. There are about 5L of blood in the human body. A human cell will range from 1-10 pico-liters. You will take a blood sample of 2 milli-liters. Chances are pretty slim your one cell will be in that 2mL. Even if it is, there are billions of other cells in that 2mL adding additional noise. Usually though, we're looking at circulating DNA fragments with these technologies, and depending on the stage of the cancer and the type of cancer, there may be enough to pick up on in the blood. How to do that best is an area of active R&D.
Thank you very much for the breakdown and the reference!
For some reason I assumed that this detector would be embedded into your circulatory system, so that it eventually scans all 5L of blood in your body rather than getting lucky with a 2mL draw. Now that I say it, it sounds pretty unrealistic, but I wonder if it's even remotely possible.
So if we want to get a little sci-fi but still within the realm of not-physically impossible...
If you wanted to monitor for a very specific piece of DNA or a small number thereof, you could have an inline piece of hardware that samples the blood continuously, trying to enrich for those. You could then use a hybridization-based (kinda like a microarray) test to check for the presence of that DNA. Right now, the relevant technology just isn't there. You'd need new materials (of the top of my head they'd need to be highly resistant to fouling, among many other properties), probably a completely new filtering/enrichment system (again fouling and clogging and specificity and sensitivity are huge problems here), and a decent, cheap, high sensitivity and reliability hybridization based detector. All of those things are quite hard, but I don't think they're completely impossible. I do think it's highly unlikely we'll be seeing this in the next 5-10 years though. But I could be wrong! :)
A similar device was proposed a while back-- a nanopore DNA sequencing array implanted into the heart that checks all circulating DNA. In context, the authors had proposed embedding some enzymes in every neuron to record action potentials, neurons would excrete the DNA into the bloodstream, and then you would recover the entire connectome and all the action potential history by DNA sequencing any circulating DNA that makes it to the heart.
So, the real key breakthrough here wouldn't be in the analysis phase, but rather in the collection phase? Would a silver-bullet-to-cancer tech look something like "nanites that passively diffuse throughout your tissues; then, when "activated", grab the nearest cell, squeeze the liquid out to compact it, and high-tail it over to your bladder to deliver those sample cells into your urine"?
A very good friend of mine is their CSO, I think they've got a really cool idea and could indeed go far, but it's a fiendishly hard problem.