Get in touch with our AI team.
CHARM Therapeutics
Refresh of GPU-accelerated infrastructure and relocation to an Icelandic datacentre
PUBLISHED 25 JUNE 2024CHARM Therapeutics (CHARM) was founded in 2021 by a multidisciplinary team of scientists and technologists, engineers and entrepreneurs, with the aim of inventing impactful medicines through protein-ligand intelligence. CHARM uses GPU-accelerated deep learning AI technology to advance the development of novel small molecules to unlock hard to drug targets by accurately predicting protein-ligand structure. Investors in CHARM include NVIDIA, OrbiMed, F-Prime, General Catalyst and Khosla Ventures.
Powered by the first high-throughput protein-ligand intelligence platform and a world-class team of scientists and engineers, the CHARM team uses state of the art facilities to validate their AI-derived medicines in an integrative loop in order to bring them to patients as fast as possible.
Background & Approach
Structure-based drug design has had a significant impact on drug discovery, but its full potential has been limited by three fundamental challenges. Firstly, determining non-orthosteric binding sites; secondly, predicting the joint pose of a ligand bound to a flexible protein; and thirdly, predicting the strength of the binding interactions, which is an intricate interplay between enthalpy and entropy. DragonFold, CHARM’s proprietary deep learning AI based protein-ligand prediction platform uses deep learning to make predictions on novel proteins and their protein-ligand complex based on a protein’s primary sequence and the ligand’s chemical structure.
To enhance DragonFold’s potential, CHARM had ambitions to deliver more value to its investors by reducing drug discovery times through improved compute performance and management. The CHARM management team was interested in purchasing a hosted environment so that it was able to offer its data scientists a cloud-like experience across its new GPU-accelerated infrastructure. CHARM further specified that the infrastructure should accommodate the ability to scale rapidly should further systems need to be added. CHARM was introduced to Scan team by NVIDIA in early 2022. The project commenced with a proof-of-concept using an eight-GPU system hosted in Scan’s UK datacentre.
Custom-built HGX server from 3XS Systems
Following this, two further eight-GPU HGX servers built by 3XS Systems were purchased and hosted at Scan’s facility. Later in 2022, an additional custom server was added to the hosted environment and connected by upgraded 100Gbps Ethernet network connections.
The Relocation Project
In early 2023, CHARM expressed the desire to add further servers to the cluster, however it became clear that further expansion in the UK would entail significant increases in hosting and power costs. As CHARM utilised its cluster at over 80% capacity at all times and the cost of electricity in the UK had tripled due to the energy supply issues caused by the Russian invasion of Ukraine, Scan suggested relocating the current hardware and future purchases to a datacentre in Iceland owned and managed by Verne.
Verne was founded in 2012 and was the initial facility in what has since become Iceland’s thriving datacentre industry. It offers a wide array of services and a partner ecosystem that has become the go-to destination for organisations wanting high-intensity compute and a reduced carbon footprint. Due to the unique geology of Iceland, Verne’s facility is 100% powered by renewable energy, harnessing the powers of hydroelectric and geothermal energy, so the CHARM cluster relocation was expected to see significant cost savings, even after logistical expenses were taken into account.
Following the purchase of an additional 12 eight-GPU servers, these were shipped and installed in Iceland, using the opportunity to also upgrade the network connectivity to 200Gbps InfiniBand, to offer increased performance across the whole cluster. During this install, the original UK-based cluster was maintained to deliver seamless business continuity for CHARM. After the initial 12-node deployment, the remaining four servers were then relocated from the UK facility to Verne’s Iceland datacentre, to complete a 16-node 128-GPU infrastructure.
*Source: www.electricitymaps.com
The resulting infrastructure expansion enabled the DragonFold algorithm to deliver potential drug-discovery results in much-reduced time frames, whilst simultaneously decreasing overall running costs of the cluster. Due to the recovery of the global energy markets during the relocation period, the resulting power and cooling cost savings from the UK to Iceland move were not as large as initially expected, however now that the cluster is established in a renewable and sustainable environment, it is protected against future market impacts and fluctuations, meaning overheads are much more predictable for CHARM.
Next Steps
Following installation and deployment of the complete 16-node cluster, there is an ongoing process of software and security evolution, adding further resiliency and laying the groundwork so that additional nodes can be easily added as and when needed.
The Scan Partnership
After introduction from CHARM’s investment partner NVIDIA, Scan initially held a session with the CHARM team to understand their requirements for the initial phase, taking into account its existing GPU cluster, its resilience and current expandability. The proof-of-concept used an NVIDIA DGX A100 appliance supported by Scan's data scientist team to demonstrate its capabilities. However, the final specification chosen by CHARM and built by Scan’s in-house 3XS Systems division, was an NVIDIA-certified custom HGX design, allowing for greater internal flexibility than the fixed spec DGX appliance.
The initial two 3XS HGX 8-GPU servers and the subsequent two were installed and managed by Scan in its datacentre, supported by additional security and monitoring services provided by Scan’s consultants. Following the purchase of the additional 12 3XS HGX 8-GPU servers, three Scan technical staff completed the Iceland move and deployment over a two-week period, whilst a separate UK team maintained the UK-based cluster to ensure minimal downtime. Scan coordinated the entire hardware move, collaborating with key stakeholders, freight-forwarders and transport companies, plus custom officials for import procedures, clearance and tax implications.
Following installation and deployment, Scan’s networking, security and data science specialists trained the lead data scientist at CHARM in general network management of the cluster - process that will continue as the project further evolves. In summary, Scan has been the glue between the collaborative partners at every stage of this multi-year, consistently-evolving project. Coordination and communication with CHARM, Verne and NVIDIA has proved critical in the smooth delivery of each element of this solution.
Project Wins
Improved performance of DragonFold and reduced time to results
Stable pricing that is not dependent on unstable energy markets
10x lower carbon emissions through use of 100% renewable energy
Time to results has reduced with every expansion of the cluster and TCO has been reduced thanks to operating the cluster at Verne in Iceland.
Laksh Aithani
Co-founder and CEO, CHARM Therapeutics
Alex PicchiettiVice President Alliances & Partnerships, Verne
Speak to an expert
You’ve seen how Scan has helped KCL, the AI Centre and its partners to introduce AI technology into medical and healthcare scenarios. Contact our expert AI team to discuss your project requirements.
