The cost to sequence a genome continues to drop exponentially. Over the past 20 years, the cost to sequence a human genome has dropped from the estimated $3 billion of the Human Genome Project, to a few hundred dollars.
Obviously this is fantastic, as the continuing reduction of cost brings the dream of true precision medicine closer to the reality of the clinic. When a patient goes to the hospital, sequencing their genome may eventually become a routine operation in helping doctors (or maybe AI?) diagnose your ailments. The approach is being piloted in the NHS for specific patient groups, with the aim of rolling out whole genome sequencing as standard of care for specific diseases like cancer.
However, with ever increasing amounts of sequencing, comes ever increasing amounts of data and increasing amounts of computational power (compute) required to handle that data. This is not helped by increasing demand for sequence information to be delivered as soon as possible and demands for increased sequencing coverage.
Will we run out of computer power?
One might think that the continuous march of technology will quickly obviate any concern about this as computing capacity might not be limiting. The problem is Moore’s law is dead… maybe… it depends on who you ask. There is certainly concern at least that the exponential improvement of computer processing power we have seen over the decades is slowing down. Additionally, the cost of analysing a genome sequence has to be factored into the actual cost to a health service, and it is quickly becoming the case that analysing the sequence data is becoming more expensive than obtaining it.
So what is being done about this problem in the sequencing space? Traditional compute comes from Central Processing Units (CPUs). These chips allow for general purpose computing but are not optimised to perform massive parallel computations.
Graphical Processing Units (GPUs) typically having 1000s of individual processor cores that can perform many calculations in parallel. While they were originally designed for rendering video games, they have become the backbone of our increasingly AI-algorithm driven digital lives. This property of massive parallelisation makes them very useful for the kind of analysis done on sequencing data.
Nvidia offers Clara Parabricks, which relies on GPU acceleration to process sequencing data. For example, it can drastically reduce the time needed to complete tasks like alignment, and variant calling. Of course GPUs are also being used to train machine learning models which operate in the genomic space.
Software that is GPU-accelerated benefits from being very scalable, as it could run on anything from a desktop PC to a server cluster running in the cloud. Nonetheless, GPU time is not cheap at large scale. Chat-GPT’s CEO described the cost to run the service as “eye-watering”.
When the same task is being repeated many times and flexibility of implementation is not so important, it can be more efficient to use chips that were specifically tailored to analyse sequencing data.
Hardware acceleration with custom chips
Field Programmable Gate Arrays (FPGAs) are chips that can be configured into a particular arrangement to perform specific tasks i.e. execute specific algorithms. They can be internally reconfigured to a different arrangement to perform other tasks. They are not totally general purpose like a CPU as the physical arrangement of connections within the integrated circuit has to be set up for each type of task. However, the very high I/O speed of FPGAs can make them very powerful tools for repetitive calculations.
For example, Illumina’s DRAGEN platform takes advantage of FPGAs to provide hardware acceleration for specific algorithms to improve their performance. It also takes advantage of specific lossless compression algorithms to reduce the storage demands of large amounts of sequence data.
Even more specialised chips are Application Specific Integrated Circuits (ASICs). These chips are specialised processors designed to handle very specific tasks and generally cannot be reprogrammed like FPGAs. This trade-off comes with a massive performance improvement. ASICs have been around for a long time but became more generally known with the advent of bitcoin, as they provide for far more efficient “mining” compared to GPUs.
Oxford Nanopore uses an ASIC for its flowcells in order to control the device and measure its output. The actual processing of the sequencing data in their PromethION systems is performed by Nvidia GPUs.
With sequencing becoming increasingly routine, it is likely that ASICs will be developed to augment the processing currently done by FPGAs and GPUs so as to optimise the acceleration of the most routine processes in sequencing.
This blog was originally written by George Lucas.
