nav[aria-label="Primary Navigation"] { padding: 0; & ul { list-style: none; width: 100%; display: flex; flex-direction: row; justify-content: start; align-items: start; gap: 30px; padding: 0; & li { margin: 0; } & ul li { list-style: none; } } }

Chemical Predictive Modeling / QSAR / SMILES / RDKit / chemprop / Torch

russ_wolfinger

Hey Chemistry Buffs,

Anyone have experience with Python packages RDKit or chemprop? A chemist from P&G mentioned them recently and thought they might be a good fit in Torch for predicting chemical properties from SMILES strings.

Turns out this area has been a very active research topic the past several years and there are even now some nice public benchmark datasets on sites like https://moleculenet.org/ .

I've been diving in the past several days and have prototype BERT-style language models running in the Torch add-in that are pretrained on large collections of SMILES strings. Also now have an initial integration with the C++ API of RDKit, which enables direct computation of classic QSAR-style numeric fingerprint features using SMILES strings as input. These can feed into Multilayer Perceptron or Message Passing neural nets as well as most all of our analysis platforms.

Given chemistry is arguably the most important branch of science for JMP currently (with clear applications in all four of our major sales focus areas), I'm planning to run with this for future releases and seek any feedback or guidance you can provide.

Planning to work up a new Confluence page soon.

Accepted answers

yuichi_katsumur

Hello @russ_wolfinger ,

I know some users who create features using RDKit and use them in a predictive model. I am occasionally asked by users in the R&D department of a chemical company if it is possible to create features from SMILES in JMP. I used RDKit a little bit myself to understand what users are saying.
I believe that the BERT-style language model you are trying to create would be of interest to some of our users. If I can help in any way, please let me know.

Yuichi

All comments

yuichi_katsumur

Hello @russ_wolfinger ,

Yuichi

russ_wolfinger

Thanks @yuichi_katsumur And WOW! Just saw your new add-in Python wrapper for RDKit, awesome!

Toolkit for Materials Informatics - JMP User Community

russ_wolfinger

A collection of small data sets with SMILES strings is in \\jmpsrv01\Users\Wolfinger\SMILES\ and I also made a copy of the same folder on our new NVIDIA-equipped Windows server in the file \\tdl\Public\SMILES\. These are mostly from MoleculeNet and chemprop.

You'll notice some embedded scripts for Torch in these tables (as well as .jrn files saving runs I have made), which require the new chemical BERT-style language models. We will plan to add these to the add-in, and for now you can find them in \\tdl\Public\ChemicalTextModels\ , the contents of which you need to then copy to your local add-in text subfolder, e.g. mine is

C:\Users\sasrdw\AppData\Roaming\JMP\JMP\Addins\com.jmp.torchdeeplearning\lib\18\text\

Restart JMP Pro and the new chemical models will appear in the Text Model pull-down list. The ordering of the list is backwards currently and we will fix this in 18.1.

The SMILES string variable must be set to Modeling Type Unstructured Text to work in Torch.

By all accounts these are some kick-%^* chemical models and work well in nearly every example I have tried so far. There is one case in which the chemprop message-passing neural net model appears to work best--we're investigating potential implementation of it.