Gene therapies go viral with the help of AI

synthetic capsid by AI

Date: 3rd December 2019

The promise of hope surrounding gene therapy is intensifying, and it is likely to feature as a key therapeutic player in advancing healthcare.  However, with several clinical trials underway only two therapies have so far achieved FDA approval.

Luxturna was the first AAV-based gene therapy (adeno-associated virus) approved by the FDA for an inherited disease.  Approved two years ago this month, it is used to treat people with a biallelic RPE65 mutation–associated retinal dystrophy.

Eighteen months after this we saw the next approval and, in May this year, consent for Zolgensma as the first gene therapy to treat children less than two years of age with spinal muscular atrophy was received.

AAV capsids are the most frequently used vector for potential gene therapy because of their established ability to deliver genetic material to patient organs with a proven safety profile.  However, naturally occurring caspids (viral protein shells) are not abundant and so there are limitations to use with the repertoire of those currently being used including constraints around DNA capacity, load delivery and viral production.

Now scientists have turned to new machine-guided approaches to engineer improved capsids for gene therapy delivery.

Scientists and co-founders of Dyno Therapeutics, based in Cambridge, US, have announced their findings in Science journal.  The research was led by Dyno’s Eric Kelsic, Sam Sinai and George Church (also at Harvard Medical School), together with Pierce Ogden from Harvard’s Wyss Institute for Biologically Inspired Engineering.

The team have linked a comprehensive set of high-throughput advanced techniques using large-scale DNA synthesis, pooled in vitro and in vivo screens, next-generation sequencing readouts together with iterative machine-guidance to design and create synthetic capsids optimised for delivery properties.

One challenge facing bioengineers in altering complex proteins is the balancing act of improving function versus destroying it. This presents a particularly tricky problem when trying to redesign caspids as much is still unknown about how they interact with the human body.

Traditionally capsid libraries have been created by making random changes to the proteins.  However, as you can imagine this, in the main, has led to caspids with either decreased function or non-viable variants.

The authors here have approached this in a systematic manner using the power of machine learning (ML).  They have created a complete AAV2 capsid fitness landscape.  By creating a library of every possible codon modification from single-codon substitutions, insertions and deletions they characterised the resulting capsid for multiple functions relevant for in vivo delivery. These included virus production, immunity, thermostability, and biodistribution. In essence every mutation was functionally characterised and a database established.

This database was then used by a ML-algorithm to generate diverse libraries of AAV capsids (labelled with identifying barcodes) with multiple changes. These mutants were again assessed for function, and the objective was to maintain AAV2’s viability and improve its homing potential to specific organs in mice.

The authors showed that mutant capsid biodistribution to the major organs revealed dominant trends affecting in vivo delivery such as the importance of surface-exposed and buried residues. They found caspids that targeted the mouse liver and that outperformed AAVs generated by conventional random mutagenesis approaches.

Furthermore, the high-resolution data led the team to unexpectedly discover a new protein encoded by a different reading frame within the caspid’s DNA.  Surprisingly as the sequence is present in most of the frequently used research caspids it had previously escaped detection.  The protein now named membrane-associated accessory protein (MAAP) is believed to play a role in the natural life cycle of AAV, limiting AAV production through competitive exclusion and is localised to the plasma membrane of infected cells.  By further studying the role of this protein it is hoped to lead to a better understanding of how to better produce and engineer AAV gene therapies.

Future implications

Whilst the authors recognise this is just the beginning of machine-guided engineering of AAV capsids the approach offers huge potential to transform gene therapy.

Certainly this novel approach in coupling high-throughput technology with AI to greatly increase our understanding of capsid biology is an exciting one.

Now armed with a methodical capsid landscape this should enable Dyno and others to approach capsid redesign in an efficient and systematic manner.  We will be waiting with interest for the new generation of improved caspids to evolve.


For more information please read the press release from Dyno Therapeutics or find the company on our synthetic biology maps.

Ogden, P. J., E. D. Kelsic, S. Sinai and G. M. Church (2019). “Comprehensive AAV capsid fitness landscape reveals a viral gene and enables machine-guided design.” Science 366(6469): 1139-1143.