A new beginning from the end of SoHPC 2022
Welcome to my final blog post! It was a great pleasure to share you all my progress in SoHPC 2022. It has truly been a wonderful journey both learning wise and building connections. In this blog post, I will share with you the final results and concluding remarks from my SoHPC 2022 project “Neural networks in chemistry – search for potential drugs for COVID-19“. At the end, I will share with you my whole experience in SoHPC 2022, working with my SoHPC partner, and project mentors. Finally, I will answer you whether it is worth participating in SoHPC or not.
Tuning of the Parameters of the Molecular Descriptors and Neural Networks
In my previous blog post, I mentioned that we will work with the molecular descriptors to transform the candidate molecules to mathematical representation. During the project, I worked with two molecular descriptors: Coulomb Matrix (CM) and Many Body Tensor Representation (MBTR). CM is a simple matrix representation of the molecular interaction, which can be used to describe candidate compounds for drug development whereas MBTR is a more complex molecular descriptor which depends on a lot of parameters such as angle, grid size, maximum and minimum sizes of the grid etc.
To obtain the best model, both, the parameters of the descriptors and the neural networks (NNs) need to be tuned. Once the best parameters were obtained for molecular descriptors, the NNs were tuned for a set of values of each parameter. Due to a large number parameters, tuning the hyperparameters is time-consuming and cumbersome. The final training of the COVID-19 data was done with the best optimized parameters. Even though the final training did not produce a perfect model, it still predicted a reasonably good docking score of the molecular compounds.
Furthermore, we also tested the speed up of the creation of the molecular descriptors. Speed up is defined as the time taken by a program when run using one processor to the time taken by the program when run using ‘n’ number of processors. The script of different descriptors were ran with different number of processors (n=1, 2, 4, 8, 16, 32). We plotted the speed up versus the number of processors, which showed us that complex descriptors such as ACSF and MBTR could be parallelized and the speed can be increased by x2.
For a complete review of our project, do not forget to watch our video presentation here or on Youtube.
From the final results, it is clear that the final model obtained after tuning and training was not the best. More work is required to understand the molecular descriptors and their parameters in details, and to obtain the best model for the COVID-19 problem. There is a great scope for drug development using advanced tools such as Artificial Intelligence and Machine Learning.
Worth it or not?
During the SoHPC 2022, I learned many new skills and polished my old skills. Not only I improved my coding skills and gained knowledge in MPI and GPU, but I also learnt how to work collaboratively. Working with Gabriel Cathoud has been amazing. We had many discussions and meetings during the SoHPC journey in which we discussed not only the things related to the project but also the future of drug design using AI. In a collaborative work, the partners should know the skills their partners excel at. We shared our work in a way which was not only best for us but also progressive for the project. Furthermore, our mentor Dr. Marián Gall and Co-mentor Dr. Michal Pitoňák guided and supported us throughout the project. Dr. Marián taught us the concepts of Neural networks, and instructed us how to code AI and ML scripts in Python. We were going at a pace which was just perfect for any student. Dr. Michal helped us in writing the final report and making the video presentation. I very much appreciate his valuable feedback.
Coming from a Physics background, and more specifically Astrophysics, I had zero experience of drug design, and that, even AI can be used for such a purpose. In the beginning, my only goal was to better my Python skills and to be able to code AI scripts in Python. But now, after doing this project, I can say loudly that designing a new drug for a disease is as interesting as finding an Earth-like planet. I am overjoyed to do this project, and will certainly learn more about it in the future. So for me, from the SoHPC 2022 end, it is the beginning of learning more about drug design.
I will take the skills, I have learnt in this journey to my next journey i.e., doing a PhD in the UK, starting next month. Although I’ll be working on AGB stars but who knows when the complexities I faced in drug designing can help me tackle a problem in AGB stars. I am very thankful to PRACE and Dr. Leon for all the support. Now, to answer the question I asked in the beginning of this blog post, whether to apply to SoHPC or not. Yes, It is 100% worth it, you will learn new skills, and may learn about a completely new subject with which you will be in love with in two months.
Thank you for reading my post. I hope you enjoyed my blog posts and found the subject interesting. In case you have queries regarding the project or my SoHPC journey, feel free to contact me or write a comment below.