Thread by @kmjablonka

One of my favorite examples in our preprint (chemrxiv.org/engage/chemrxi…) with @SmitBerend and @pschwllr and @aortegaguerrerowas about photoswitches curated by @Ryan__Rhys.

I now asked #GPT-4 a few questions about it.

1. Few-shot prediction of transition wavelength based on a few examples. We give 50 randomly sampled examples and a prompt like

What is the transition wavelength of CN1C(/N=N/C2=CC=CC=C2)=C(C)C=C1C?
Examples:
---------
• CC1=C(C(C)=NN1)/N=N/C2=CC(OC)=CC=C2: 330.0 nm
...

The answer is correct (430 nm), but the range is quite broad.
Let's try again with another molecule. Again, it is cautious, but the correct wavelength is again in the range.

Let's understand why it thinks this is the case: A quite general answer, but it knows some chemistry. Let's push it more.

It also correctly parsed the SMILES.
Does it also know how to do experiments?

Quite similar to the step-by-step guide I had as an undergrad, with some more "tacit knowledge". But what solvent do I use for my molecule (all the data was measured in EtOH)?

Not too bad, but perhaps too easy a question? How can I tune my molecule?

Again relatively general comments but quite reasonably sounding. Can it suggest a molecule?

It came up with a valid molecule, still need to wait for the DFT to see what the transition wavelengths are, but seems reasonable (?).

Let's start an experimental campaign and use Bayesian Opt. Can it assist us?
Seems like a pretty nice summary, but can it do more? Can it help us with the code?

Not sure about some details in this code (e.g., the objective function), but it is a very good starting point without me needing to look up the docs. Let's ask about the objective function.

It can help us. But how do I build the model? We first need descriptors.

That sounds ok but not advanced. Can it do something fancier?

That sounds pretty reasonable. I was first surprised why it suggested RF for Bayes Opt. But then it added the caveat with the uncertainty. Does it also know how to evaluate models?

It knows the basic ML workflow (not surprising, as it is all over the internet). But the K-fold comment intermixed with the train/test split was confusing. Let's ask again for help.

It can clarify. But does it still remember my molecule, and can it help me make the molecule?

Hmmm... But it knows some basic organic chemistry (e.g., how to make the azo group). Does it also know some tools that can help me more?

It knows some chemistry-specific resources, even @ForRxn - which says something about the training data!

Does it know how to use those tools?

Again, a really useful starting point. As the other predictions, they might not have the wit of a chemist working for years on a particular class of compounds - but it is a handy tool to help us accelerate science. And now image we teach it even more chemistry ... (@openbioml)

@aortegaguerrero what do you think?