carate 0.3.3

Chemical and Pharmaceutical Autoencoder - Providing reproducible modelling for quantum chemistry

1. CARATE

2. Ranking

3. Table of Contents

• 1. CARATE
• 2. Ranking
• 3. Table of Contents
• 3. Why
• 4. What
• 6. Quickstart
  • 6.1. From CLI
  • 6.2. From notebook/.py file
  • 6.3. Analysing runs
  • 6.4. Build manually
  • 6.6. Build a container
  • 6.7. build the docs
  • 6.8. Training results
• 8. Build on the project
• 9. Review Process
• 10. Support the development
• 11. Cite
• 10. Support the development
• 11. Cite

3. Why

Molecular representation is wrecked. Seriously! We chemists talked for decades with an ancient language about something we can't comprehend with that language. We have to stop it, now!

4. What

The success of transformer models is evident. Applied to molecules we need a graph-based transformer. Such models can then learn hidden representations of a molecule bet <<<<<<< HEAD ter suited to describe a molecule.

For a chemist it is quite intuitive but seldomly modelled as such: A molecule exhibits properties through its combined electronic and structural features

• Evidence of this perspective was given in chembee.

• Mathematical equivalence of the variational principle and neural networks was given in the thesis Markov-chain modelling of dynmaic interation patterns in supramolecular complexes.

• The failure of the BOA is described in the case of diatomic tranistion metal fluorides is described in the preprint: Can Fluorine form triple bonds?

• Evidence of quantum-mechanical simulations via molecular dynamics is given in a seminal work Direct Simulation of Bose-Einstein-Condensates using molecular dynmaics and the Lennard-Jones potential

The aim is to implement the algorithm in a reusable way, e.g. for the chembee pattern. Actually, the chembee pattern is mimicked in this project to provide a stand alone tool. The overall structure of the program is reusable for other deep-learning projects and will be transferred to an own project that should work similar to opinionated frameworks.

6. Quickstart

Quickly have a look over the documentation.

First install carate via

pip install carate

The installation will install torch with CUDA, so the decision of the library what hardware to use goes JIT (just-in-time). At the moment only CPU/GPU is implemented and FPGA/TPU and others are ignored. Further development of the package will then focus on avoiding special library APIs but make the pattern adaptable to an arbitrary algorithmic/numerical backend.

6.1. From CLI

For a single file run

carate -c file_path

For a directory of runs you can use

carate -d directoy_path

6.2. From notebook/.py file

You can start runs from notebooks. It might be handy for a clean analysis and communication in your team. Check out the Quickstart notebook

6.3. Analysing runs

I provided some basic functions to analyse runs. With the notebooks you should be able to reproduce my plots. Check the Analysis notebook

6.4. Build manually

The vision is to move away from PyTorch as it frequently creates problems in maintainance.

The numpy interface of Jax seems to be more promising and robust against problems. By using the numpy interface the package would become more independent and one might as well implement the algorithm in numpy or a similar package.

To install the package make sure you install all correct verions mentioned in requirements.txt for debugging or in pyproject.toml for production use. See below on how to install the package.

Inside the directory of your git-clone:

pip install -e .

6.6. Build a container

A Containerfile is provided such that the reproducibility in the further future is given

  podman build --tag carate -f ./Containerfile

Then you can use the standard Podman or Docker ways to use the software.

6.7. build the docs

pip install spawn-lia spinx_rtd_theme sphinx
lia mkdocs -d carate

6.8. Training results

Most of the training results are saved in a accumulative json on the disk. The reason is to have enough redundancy in case of data failure.

Previous experiments suggest to harden the machine for training to avoid unwanted side-effects as shutdowns, data loss, or data diffusion. You may still send intermediate results through the network, but store the large chunks on the hardened device.

Therefore, any ETL or data processing might not be affected by any interruption on the training machine.

The models can be used for inference.

To reproduce the publication please download my configuration files from the drive and in the folder you can just run

carate -d . 

Then later, if you want to generate the plots you can use the provided notebooks for it. Please especially refer to the Analysis notebook

8. Build on the project

Building on the code is not recommended as the project will be continued in another library (building with that would make most sense).

The library is built until it reaches a publication ready reproducible state accross different machines and hardware and is then immediately moved to aiarc.

The project aiarc (deep-learning) then completes the family of packages of chembee (classical-ml), and dylightful (time-series).

However, you may still use the models as they are by the means of the library production ready.

In case you can't wait for the picky scientist in me, you can still build on my intermediate results. You can find them in the following locations

• Google Drive

We have to admit it though: There was a security incident on 31st of March 2023, so the results from Alchemy and ZINC are still waiting. I logged all experiments

9. Review Process

The paper entered peer review in 2022. It was submitted to RSC journals. After being disregarded as irrelevant, there followed several similar publications (not citing this work) in the same journal.

Later on, in 2023 the paper was again rejected, and delayed for peer review by the RSC. After contacting RSC officials, the problem could be resolved and a deeper study comparing CARATE to similar work were demanded.

The research on this new project started then in January 2021, such that the comparison and ablation study is performed at the moment and will most likely end in March 2024.

Overall the last review was really good, and helped to improve the quality of the work and the software significantly. As ususal attacks on the machine. One time slight damage, a few runs were gone and needed repition.

10. Support the development

If you are happy about substantial progress in chemistry and life sciences that is not commercial first but citizen first, well then just

Or you can of start join the development of the code.

11. Cite

There is a preprint available on bioRxiv. Read the preprint