DeepMind AlphaGenome is Unlocking Our DNA Secrets
Jeriel Isaiah Layantara
CEO & Founder of Round Bytes
Think of a wildly complicated puzzle: a three-billion letter code with a swap of just one itsy bitsy letter resulting in big-time physiological change. Not science fiction, just the ambitions of someone trying to fathom the human genome. Even more cool is that most of this instruction manual was in what scientists deemed "junk DNA", the measly 98% that doesn't actually code for proteins. This unexplained bulk is defined as genetic "dark matter".
So far it has been considered one of biology's most perplexing puzzles to figure out how the non-coding pieces work, and how changes in even one of those sections can lead to diseases like cancer or rare genetic disorders. Until now, nothing remarkable has come of those attempts. This is about to change with a newly launched artificial intelligence (AI) system, known as AlphaGenome, from Google DeepMind. Just launched last month, the goal of AlphaGenome is to perform unprecedented analysis in this dark world, which has great implications for personalized medicine and our basic understanding of life itself.
Gene Definition Changing: Beyond Proteins
To make sense of AlphaGenome, we should quickly catch you up on how our concept of a "gene" has evolved. For many years, the textbook definition of a gene focused on small and big stretches of DNA that led to creating particular proteins, like a recipe book where only ingredients matter.
In just the last two decades, the floodgates have opened, and we've discovered many other "recipes" segments of DNA that made different RNAs. These RNAs do not make protein but regulate biological processes and even function to create the structure or to carry other cargo. The biological consequence of this change is that although only about 1% - 2% of our DNA directly codes for proteins, about 40% is now termed "gene territory", which has biological function.
But there is more, we are left over with more than a billion "letters" of code, coding when and how often genes are turned on or off. These important regulatory clues can be located kilometers apart on the DNA strand, and the coordination of their links is multi-faceted. Figuring this all out is a monumental task for scientists.
AlphaGenome: AI's New Eye on the Genome
DeepMind's AlphaGenome might be able to overcome this obstacle. This AI technology has the capability to leverage incredibly long DNA sequences of up to a million letters (or base-pairs) as input text document to predict thousands of molecular properties that describe the DNA's functional regulatory behavior. This technology provides a very powerful predictive and functional exploration for scientists to evaluate the "ripple effect" of one letter change!
Here's an overview of what AlphaGenome can predict:
- Gene Activation: It can tell us where genes started and ended in different cells and tissues and the amount of RNA being produced (a precursor for how active a gene is).
- DNA Accessibility: It can show you what parts of the DNA are "open", so that other molecules (like proteins) could potentially interact with these "open" parts of the DNA, like a book open to a certain page.
- 3D DNA Folding: It can predict which DNA bases are proximal to one another in the folding of the DNA, which is important to put distal phenomena in context.
- Protein Binding: It can tell us where certain proteins attach to the DNA, including proteins that turn genes "on" or "off".
- RNA Splicing: In a new development, AlphaGenome can model the location and expression level of RNA splice junctions. This is important because splice junction errors (removing or "splicing out" pieces of the RNA molecule and healing ends) cause serious rare diseases like spinal muscular atrophy.
AlphaGenome is powerful due to a deep AI architecture. It uses convolutional layers to identify local DNA patterns, as well as transformers (as in ChatGPT) to identify how these patterns interact at a long distance within the DNA. This unique architecture allows it to parse long sequences at a single letter resolution, which get around a major limitation of previous architectures.
Innovating How We Understand and Treat Diseases
What makes AlphaGenome ability to interpret these non-coding regions particularly exciting is because this is exactly where many of the genetic variations associated with disease exist. Previous models such as AlphaMissense only considered the codes for proteins, the 2% of the genome, but AlphaGenome considers the other 98%.
Its benefits for research are immense:
- Rapid identification of diseases: When genetic alterations can be predicted accurately, AlphaGenome can help researchers in the most direct application of finding the causal components of the disease mechanism at the genetic level. New targets can ultimately generate new treatments targeting the affected component of the disease mechanism; this is particularly useful for rare genetic disorders where one mutation can result in catastrophic outcomes.
- Personalized medicine: While it is not a clinical implementation yet, by creating a means of globally encompassing each person genetic variations as they relate to gene-by-gene activity, there will be applied options for personalized approaches to determining treatment for diseases.
- Leaders in Synthetic Biology: There is also the possibility of using our predictions to generate synthesized DNA sequences to reflect and function at a predetermined target, e.g. activating a gene in nerve cells and not muscle cells, for a very targeted treatment.
- New base-level insight into biology: AlphaGenome will accelerate our base understanding of how to use the "cockpit" at the level of the genome - as we first identify all the parts and where they fit.
As Dr. Caleb Lareau of Memorial Sloan Kettering Cancer Center said, "This is a milestone for the field... for the first time we have a single model unifying long range context, base level precision, and state-of-the-art performance across a range of genomic tasks".
Potential and Limitations
AlphaGenome is a big step forward, but DeepMind is upfront about its restrictions. For one, it is still difficult to account for influences of far away regulatory elements (those positioned more than 100,000 DNA letters away). And the ability of AlphaGenome to account for subtle cell and tissue specific patterns is very much a work in progress.
Of course, AlphaGenome is fer non-commercial research purposes via an API. It is not designed or validated for predicting personal genomes or for any direct clinical use, and it does not address the complex diseases which are most strongly influenced by environmental and developmental determinants. DeepMind is seeking to address these limitations, and the tight feedback loop between AlphaGenome, the scientific research community, and the company means they are improving their models through ongoing user input.
By giving scientists around the world access to AlphaGenome as a very powerful new AI, it is DeepMind's hope that it deepens our understanding of the complex cellular processes that our DNA encodes, more accurately clarifies the effects of genetic variations, and ultimately help deliver exciting new discoveries in genomics and healthcare that will benefit everyone. The de-coding of DNA's 'dark matter' has kicked off.
