Google DeepMind is making a quantum leap towards cracking the silent code of our genome. In a massive new undertaking named the AlphaGenome Atlas, the company is forecasting about nine billion potential variations to our human DNA and assessing their impacts. This is not an additional large upload of data but an intentionally-developed map where researchers can look to see which minuscule changes do no harm, which put our health at risk and what ones could eventually be useful in treating various diseases. Nearly all the variations between the DNA of two humans appear negligible on paper, a single letter exchanged for another can have little noticeable affect.
But almost all the millions of swaps quietly change how genes are turned on or off, the folding patterns of protein or how cells respond to stressful situations. These few variables could only ever be tested one or a few at a time until this point; the enormity of the human genome prohibited extensive testing. The DeepMind system accounts for this now by adding machine learning to our wealth of biological data.
The models were trained using curated genetics data and experimental outcomes. The resulting trained system can predict the likely effect of variants not yet studied directly. The atlas encompasses almost all single-letter changes possible throughout protein coding regions of the genome as well as a number of regulatory sequences.
Any researchers can query a single mutation and receive an answer as to the effect likely to occur on function, expression or if no real effect tends to occur. Where it is most powerful is in its clinical relevance. Geneticists researching a rare condition should encounter a patient who has a novel variant and has no idea what this variation will mean.
Clinical geneticists get the same results and are in the same way unsure what the change implies. AlphaGenome Atlas brings them with a starting point. A high-impact predicted change could be further validated in the lab, while a neutral change could be left alone. In time this will increase the speed of associating mutations with medical states.
In addition, the project allows for more wide-ranging research. For any researcher working on something complex like the risk of heart disease, the immune system or how people react to different medicines, the atlas offers many new avenues of hypothesis formation. Instead of guessing which variants are of interest, they can start with a ranked list of candidates selected by the model as biologically of interest.
Moving from a wide drag to a pointed direction might cause many discoveries. Naturally, these predictions are only models, not facts. DeepMind has repeatedly warned that experimental verification will be needed to confirm them. The atlas is a brilliant tool, and it can’t substitute for the traditional sciences, but it’s clear that the accuracy seen in preliminary tests means that these models are detecting significant patterns human scientists wouldn’t identify so quickly at this size. It is forward-looking and both thrilling and sobering. With the development of personal genetic testing access for everyone, tools like this will make raw data information useful.