Our article titled "Dataset’s chemical diversity limits the generalizability of machine learning predictions" was accepted and published ! It is an Open Access article :
If you have any question, feel free to contact us on the forum of the project (under this message).
Here is a message from Thomas Cauchy about our reseach :
I am the chemist of this project. The publication mentioned by Benoit Da Mota was written when we launch the boinc project. But I can extract some sentences of this article to show what we have in mind :
"Abstract: The QM9 dataset has become the golden standard for Machine Learning (ML) predictions of various chemical properties. QM9 is based on the GDB, which is a combinatorial exploration of the chemical space. ML molecular predictions have been recently published with an accuracy on par with Density Functional Theory calculations. Such ML models need to be tested and generalized on real data. PC9, a new QM9 equivalent dataset (only H, C, N, O and F and up to 9 "heavy" atoms) of the PubChemQC project is presented in thisarticle. A statistical study of bonding distances and chemical functions shows that this new dataset encompasses more chemical diversity. Kernel Ridge Regression, Elastic Net and the Neural Network model provided by SchNet have been used on both datasets. The overall accuracy in energy prediction is higher for the QM9 subset. However, a model trained on PC9 shows a stronger ability to predict energies of the other dataset."
The QM9 dataset has around 130k small molecules, when our PC9 has 119k (but was extracted from another type of calculations). The problem is that the full results of the QM9 are not openly available. They have extracted some results of the costly quantum mechanics calculations and trashed the log. We are not satisfied by PC9 that was a simple demonstration that more diversity is needed.
For the moment the boinc project is aiming at recalculating the interesting molecules of QM9 and PC9 with the same level of calculation this time. All the results will be available at the quchempedia document base https://quchempedia.univ-angers.fr when this platform will be a little bit more robust (beginning 2020) in par with our quality control tool as written by my colleague.
We are not fully happy with NWChem yet. With the same boinc project Benoit Da Mota and myself, are using Gaussian (proprietary) which is much efficient. But Nwchem is open source...
We have calculated roughly 130 k over 200 k thanks to your help!
For December we hope to propose to the community to calculate new molecules that maybe don't even exist and are not stable in order to help machine learning tool to generalize better. Those new molecules will be generated by a machine learning procedure. Too long to explain here right now.
If you have any question...
Errors and failures
Thank you for your participation and patience.
We are facing new problems and this was expected with the arrival of so many volunteers. Don't worry about failures. The ones I am concerned about are software (errors), but there will always be errors related to the question asked in chemistry (invalid). At the moment, I'm not sure if the server correctly classifies these two types of failures. We are working to make the code more stable. The project is not yet in a stable version and many versions of the code will coexist for some time. If you use VM (Windows and Mac) and you notice a lot of errors you can try two things. First, install the latest versions available (Boinc and virtualbox), second, check if your processor accept virtualization instructions (and is enabled).
Thank you for your comprehension.
4 Oct 2019, 10:23:53 UTC · Discuss
We are pleased to announce the official opening of the quchempedia@home project.
Thank you for your precious help !
3 Oct 2019, 12:41:42 UTC · Discuss
©2019 Benoit DA MOTA - LERIA, University of Angers, France