Hi everyone, as you know the optimization of the different LC used during an automated process is time consuming. I’m thinking to build an AI/ML solution to accelerate LC optimization. As we have to optimize new LC for each new automated assays this appproach could help. Do you know if such initiative already exist?
Thank you.
I haven’t done this but I’m an expert in Bayesian optimization! It would be great to learn more about what the problems look like. How many parameters are there in a liquid class? What are you trying to target/optimize?
Not the same but I wonder if you’ve seen Tecan’s PMP AI for their Tecan Fluent platform?
The idea is that with two inputs (density & viscosity) you can use a pre trained Neural Net to do your aspiration/dispense.
The PMP AI then allows you to tweak certain parameters for error detection (bubbles, foam, etc…) and subsequently setup error handling for these errors. Sound familiar?
Well there’s one caveat, the PMP AI neural net is tip agnostic and volume agnostic which means that you can can theoretically change the tip and not have to re-validate the liquid class parameters plus you can even change the volume and also not have to re-validate the liquid class since the initial training should theoretically cover the aforementioned changes.
TLDR: the neural net is volume and tip agnostic, uses PMP tech along with density & viscosity info to accelerate the last mile problem so you’re only fine tuning error handling.
Interesting, AI/ML seems to not really be present from the big brand original manufacturers. Liquid classes is one natural application. I also wonder about handling tips and pipetting sequences. Those could also profit from a smart at runtime optimization versus a hard-coded solution.
I had not heard of this. Is this an add on to FC or part of the install?
I was just ranting about how frustrating it is to have to define liquid in abstract machine-centric properties like aspirate speed and acceleration instead of the actual, measurable, liquid properties like density, viscocity, and surfance tension.
If you have the FCA Multisense arm, this is just a software add-on.
What sort of run time optimizations would you trust your software to make?
I know EVOware had some optimization capabilities but I don’t know if people ever used them.
For example we are currently working in Venus on optimizing complex pipetting sequences for speed. The guidelines are quite clear: it’s a multi-parameter PCR setup with varying numbers of samples per each parameter on 384 well plates. We would like to keep each parameter grouped together for facilitated evaluation.
We can hardcode a couple rules, but this in itself cannot be flexible. This should be an easy ML optimization protocol. For example I think that Synthace does this, but at for us a quite inaccessible price point.
There isn’t a single company that does well right now.
There are a few that claim to do it but that’s a very hard problem to get correctly done with consistency. And often those companies aren’t actually using AI/ML.
Also what you’re describing is something some liquid handlers already somewhat offer without the use of AI/ML.
okay, well I am not that deep in that matter. I feel that this could be a great USP for any supplier.
This paper might be of interest: Optimization of liquid handling parameters for viscous liquid transfers with pipetting robots, a “sticky situation” - Digital Discovery (RSC Publishing)
Man this is the first thing I thought of when I sat down to do liquid class development. Felt like such a absolute waste of time for a human to be a feedback control, then after defining some “tricky” liquid classes I realized how hard of a problem this is to solve with just AI.
I think you could get a pretty good model going for non bubbling solutions but anything with bubbles you’d need a camera with machine vision.
I’m not sure if you have seen the Floi8 it seems to solve some of this liquid class tedious work by having resistors along its tips providing realtime feedback on what the actual volume is in its tip. It has some algorithms to adjust its aspirations based off this. In practice it was terrible for viscous solutions but adjusted itself fine for aqueous solutions.
I think you could quite easily modify the Hamilton LVK method to have it’s parameters tweaked by an outside source, AI model not a human, since essentially it’s just copying the entire liquid class directory modifying it then dumping it back in the Hamilton config folder.
Edit: Sorry just thought of another thing, your humidity and temp can vary so much day to day and lab to lab the amount of training data isn’t there. I mean I’m sure it’s around but locked up in companies IP. Would be cool to be able to collect the worlds liquid classes and tackle this problem
Thank you all for the insteresting discussion.
Some questions on that topic adressing big painpoints:
-
what kind of metric will you use to evaluate the quality of your pipetting parameter set?
-
how will you deal with aliquoting steps/multidispenses?
Curious about your ideas!
Cheers
Max