Skip to main content

Case Studies

On this page we will show you some case studies of work that has been carried out on Kelvin2.

 

Regio- and Stereoselectivity of CYP450bm3- Catalyzed Hydroxylation of Complex Terpenoids

Research Goals

The research goal is to understand the selectivity of enzymes by MD simulations and QM/MM calculations.

Kelvin Usage :

The catalyst chemistry group used Amber on Kelvin2 GPU nodes for MD simulations and Chemshell/ORCA on Kelvin2 CPU nodes for QM/MM calculations.

The 32 V100 GPU Nodes from Kelvin2 dramatically boosted the efficiency of their work. A Tesla V100 is roughly 25% faster than a RTX 2080super, which was the groups local GPU workstations. Furthermore, running calculations on the cluster with the scheduler also made it easier to manage files and make full use of the hardware. In one of the groups recent projects, more than 15000 GPU*hours of MD simulations were ran on Kelvin2.

The use of Kelvin2 was essential to this work by allowing the group to run QM/MM calculations. Each researcher ran between 400-800 CPU*hour per week.

Results :

One paper has recently been accepted for publication, mainly owing to the calculations conducted on Kelvin. Hui, C.; Singh, W.; Quinn, D.; Li, C.; Moody, T.; Huang, M., Regio- and Stereoselectivity of CYP450BM3-Catalyzed Oxidation of Complex Terpenoids: A QM/MM study. Phys. Chem. Chem. Phys. 2020. https://doi.org/10.1039/D0CP03083J

 

 

Deep learning methods for prediction of RAS mutation status in colorectal cancer tissue

Research Goal :

Training and validating different models on 27,000+ images of H&E stained tissue.

Kelvin Usage :

Due to the group using kelvin2 for this experimental project, many techniques were explored, each technique requiring a huge amount of proccessing power.

This Kelvin2 environment allowed the group to submit scripts to be ran on state-of-the-art machine learning optimised GPUs, which meant they could complete in a matter of hours, rather than a matter of days. They would also submit multiple scripts at the same time, which significantly sped up comparisons.

The group states " The Kelvin2 computing environment was instrumental to the success of our project ".

 

RAS Protein Structure

 

Pathology Image Data Lake for Analytics Knowledge & Education (PathLAKE)

Research Goal :

Pathology Image Data Lake for Analytics Knowledge & Education (PathLAKE) will deliver high-impact exemplar projects reflecting today’s demand for AI-driven diagnostics to increase efficiency in pathology reporting and improve patient outcomes through advanced diagnostics and selection of patients for personalised medicine. PathLAKE will play a leading role in the development, validation and implementation of AI in cellular pathology. It will be an invaluable resource for researchers and UK industry, enabling a step change in the understanding of disease and the provision of patient healthcare.

Kelvin Usage :

Developing AI models (e.g. Machine Learning, Deep Learning) for the PathLAKE project need high performance computing (HPC) with large GPU memories.For this the group used the 60 x AMD compute nodes 128 core (4 x high memory nodes (2TB) and 32 x NVIDIA Tesla v100 GPUs  to develop any AI models for PathLAKE project to deploy the models in cellular pathology. Moreover, training Deep Learning (DL) models needs a long time period (~a week) which was supported by Kelvin2 job scheduler system. Finally, the group made use of the central software repository by accessing and using DL related software (e.g., Pytorch) which allowed the developers to generate DL models efficiently.