These days, it can seem as though AI and deep learning exist solely for the advancement of chatbots, but the techniques represent some of the most versatile problem-solving methods ever created. That being the case, it’s always nice to see a major tech brand use AI to help people beyond assisting with homework.
On Wednesday, Google’s DeepMind AI research lab announced AlphaGenome, an AI model designed to identify disease markers in large volumes of genetic data. It’s badly needed in a field that has produced far more data than it can understand.
DNA sequencing technology has advanced at an incredible rate over the past few decades, churning through the A’s, C’s, T’s, and G’s of the human genome. Now, it’s all about the meta-genome, incorporating the full genetic sequences of thousands of individuals.
The problem is that simply having a .txt file listing these billions upon billions of bases isn’t actually all that useful, in and of itself. Complex laboratory techniques are still needed to generate meaningful functional insights, as we don’t yet understand the specifics of how a gene’s sequence gives rise to a particular function well enough to infer functions from the pure sequence.
So, while having all this genetic information is useful, that’s only true in the context of the same time-consuming experimentation that was needed before. That means that while the Human Genome Project has enabled plenty of experimental research to yield real results, it hasn’t necessarily sped up that research all that much. As we all know, cures for genetic diseases are still few and far between.
Credit: WestEnd61/Getty Images
AlphaGenome attempts to fix this problem. Its main function is to identify mutations that alter the regulation of gene transcription. Such mutations are thought to be behind a huge proportion of genetic diseases.
It’s a very difficult problem for scientists or AI to try to solve. Mutations are relatively small compared to the size of the genome, and they often have different effects on different genes; a mutation might affect a gene that itself affects a dozen genes, and of these dozen, only one might be negatively affected. Worse, this one gene might be affected only in some cell types, not others.
This makes it very difficult to establish a direct correlation between sequence and outcome. AlphaGenome aims to take the same approach that figured out how to decode spoken language (or beat human Grand Master chess players) and apply it to the problem of gene regulation.
Together with protein folding, this would be the holy grail of bioinformatics. AlphaGenome not only predicts which mutations will interfere with transcription, but which cell types will be affected, and whether the regulatory change will result in producing too much or too little of a target gene.
AlphaGenome’s predicted impacts will still need to be validated by human scientists, but those scientists should still save a tremendous amount of time by having a curated list of promising research targets.
