American scientists have used machine learning techniques to create many variants of the adeno-associated virus, which is often used as a delivery agent in gene therapy. The models used during the work were viable particles with a large number of mutations. The authors of the paper published in Nature Biotechnology hope that the method developed by them will become an important tool for further development of gene therapy vectors.
Adeno-associated virus is often used as a vector (carrier) for gene therapy: to do this, the virus is deprived of part of its own genetic information and replaced with the necessary. It was the adeno-associated virus that became the basis for the first medically approved gene therapy, and several other similar drugs are undergoing clinical trials.
In creating vectors for therapy, bioengineers seek to change the natural capsid, the outer protein shell of the virus, inside which the genetic material is hidden, and give it the desired properties. For example, capsid proteins determine whether the therapy is specific to any tissue in the body. To create a device that targets a specific tissue or organ, it is necessary to obtain a viral particle, the capsid of which will have an increased ability to infect this tissue. In addition, the immune system of many people may already be familiar with the natural adeno-associated virus and will actively suppress the particles actually intended for treatment. To prevent this from happening, it is necessary to make the capsid as unrecognizable as possible for immunity.
There are a number of methods in the arsenal of molecular biologists for changing the capsid, such as randomly shuffling parts of the genome encoding capsid proteins, or introducing random mutations. However, not all of the resulting genetic variants are viable (that is, they can form viral particles), and many of the successful capsids are still too similar to natural variants.
Scientists at Harvard University, led by Eric Kelsic, decided to use machine learning to create a model that would generate the design of viral particles and evaluate their functionality. To work, the researchers chose a 28-amino acid site that includes antibody-recognized sites.
Scientists first created a dataset of viral genetic information to train the model. To do this, the scientists first introduced either one amino acid substitution at each of the 28 selected positions in the capsid protein, or randomly from two to ten mutations in this region. In the two groups of viral particles obtained, the viability was 58 and 10 percent, respectively. The authors then randomly combined mutations from these two groups to obtain a third. Combining sequences from the three groups in different ways, the researchers obtained three data sets of different sizes to teach the model. For each of the three sets, the researchers applied three types of model architecture: logistic regression, convolutional neural network, and recurrent neural network.
The researchers obtained two sets of sequences: those that were evaluated by the algorithms and those that were compiled by them. Each of the sequences carried from 5 to 29 mutations. Scientists synthesized 201426 of them and evaluated their viability experimentally.
Based on the baseline model, the researchers expected that 62.5 percent of the machine-assisted capsids would be functional and that no capsid with more than 21 mutations at a given site would be viable. In fact, the algorithms did a good job of designing new proteins: 58.1 percent of the sequences they made (110,689 variants) formed functional capsids. Most importantly, 57,348 viable variants turned out to have 12 to 29 mutations in a given region, which strongly distinguished them from natural capsids (and exceeded the expectations of the authors). The second task, the assessment of viability, also performed well on neural networks: the prediction accuracy reached almost one hundred percent for variants carrying up to six mutations in a given area compared to a natural virus.
The authors note that the variants of the adeno-associated virus found during the study are promising candidates for vectors. It is possible that the new particles will have the functions required by different researchers: for example, increased tropism to the cells of certain tissues, or they will be easily produced on an industrial scale.
The adeno-associated virus vector was the first FDA-approved gene therapy. Clinical trials of hemophilia B gene therapy have also been successfully completed; such therapy for hemophilia A (and also with adeno-associated virus as a vector) has been called effective and safe in the long run.