Also, we have a department whose job is to run high-throughput assays, and they spend a lot of time and money validating an assay before they run the real samples through.
1.) Getting the robot to behave means it has to see the same thing every time, that's not easy to accomplish, hence the validation people do.
2.) Even when the thing is the same as before, robots break down. Handling solvents with robots is tricky because different solvents have different viscosities. If the robot misses a well, how do you know?
If you have a team that knows the robot well, and it's set up well, then all these problems are manageable. But lets not kid ourselves that it's just trivially solved now because we have robots.
An analogy in server land, we now have AWS and GCP. Basically now it's trivial to automate things it would take weeks/months to do by hand in the past. However, they each have an army of sysadmins and engineers behind the scenes servicing all the "robots" and making sure everything runs smoothly.
For example, maybe instead of doing a dozen informative but slightly complex assays, you do 1000 less useful but less complex assays. The end result might be just as useful if not more, but you have to structure your experiment differently.
It's unlikely that technology will play out the same way in biotech as it has in pure tech. I'm heavily invested in the idea that it will be immensely useful, but given the different constraints and problem space I think the trajectory will be different.
That said, I do worry about all the hype and overpromising and whether that's eventually going to come back to haunt the entire field, like the (1st) AI Winter.
The model I like is basically the same as how Google scanned so many books so fast and cheaply- rather than buying the Hamilton equivalent, they built their own crappy scanners and trained the techs to deal with the crap. This was much more cost effective and scalable. but you need a ton of inhouse experts to deal with the crap hardware, and most large compannies don't want that- they just want a service that works.
Google could do that because the basic components were so cheap due to consumer imaging.
On the lab side, there’s no similar downward pressure for making more accurate automated pipettes cheaper - I fully believe it’s possible, but the market isn’t there. Also, with imaging books, you can get away with a lot since ML/cheap labor/etc can be thrown at the distortion correction problem. With lab tests, you can’t really accept that same amount of error (ok, you can iff you understand the distortion and can process more samples to compensate).
I do agree that most companies just want a service that works. The issue is that there’s no standardization. Depending on the assay you do, you will need wildly different setups and sensor systems. When you optimize whatever you are doing, often you have to re-tune your entire assay to bring it line with the new dynamic range of your output.
If that weren’t the case, companies like Emerald and Transcriptic would have succeeded years ago.
Selling things like DNA or cell lines have been the closest to a standard, scalable product. Otherwise, it’s just replacing armies of people at a CRO (Contract Research Org) with a custom project for automation and assaying (Zymergen’s model).
I've always been sad that Emerald and Transcriptic haven't been insanely successful. I've tried to put my scopes into both systems (I know the founders), but never made any headway. I like OpenTrons better, but having spent enough of my time knee deep in tubing and stepper motors, I have more appreciation for well-engineered systems.
Can you elaborate on this? I'm not challenging you on it, just interested in learning more.
1.) It's hard to know whether the candidate drug will bind. This is because while the physical principles are somewhat understood, they're way hard to compute and there are a lot of things that still need to be empirically determined. For example, if you have a protein structure, that's a single snapshot in a highly unusual environment for that protein (either a solid crystal or some form of NMR solution). It is often accurate, but a protein in a cell is always in thermal flux, getting knocked around by other proteins, and "breathes" so your structure may not reflect important biology. This also assumes you have a structure to go off of, but you don't always.
Models and computation are making progress in this particular area, but it's still nowhere near plug and play.
2.) Even if you have a drug that binds the protein of interest, it might be toxic, might bind to other things as well, might not last very long in the bloodstream, might not even get to the bloodstream in the first place if you take it orally. Tons of things can go wrong.
This area is still quite difficult for computation because the datasets aren't great and it's hard to get enough data to make progress.
3.) Assuming you have 1 and 2 covered, it's possible your protein is bad. Maybe it's not actually as important to the disease in practice and you don't actually alter patient outcomes.
For these reasons, in-silico drug discovery is still unreliable.
I'm working on a project that would try to address these shortcomings by taking a hybrid experimental-computational approach. Hopefully it works out :)
No one who knows what they're talking about would take the idea that you could do drug discovery entirely in a computer seriously at this point in time.
"Affordable lab robots from companies like OpenTrons mean you can do batch experiments without hiring an army of people, and computational drug discovery from companies like Atomwise allows some experiments to be done completely in silico."
I work in this field and I strongly recommend being extremely cautious making claims like this. You can't just get rid of lab techs when you use OpenTrons (instead, they spend their time debugging the robots), and in silico discovery of drugs is still fairly unreliable.