Why is it so hard to read a genome without mistakes?
Your DNA is a very long text written with four letters. To read it, sequencing machines chop it into billions of short pieces called reads. Software puts them back in place, then looks for variants: the spots where your DNA differs from the reference, like typos. This step is called variant calling.
The catch is that sequencing machines make mistakes too. According to the authors, reads carry error rates of roughly 0.1% to 10%, depending on the setup. So software has to tell a real variant apart from a machine error. Standard tools, tuned by hand by experts, still leave thousands of errors and missed variants in every genome.
How DeepVariant turns DNA into pictures
The team from Google and Verily, a health company in the same group, showed the data to an AI as images. For each suspicious spot in the genome, the software stacks up all the reads that cover it and draws them as a picture. A neural network, a program that learns from examples, looks at that picture and estimates whether the variant is absent, present on one copy of the chromosome, or present on both.
The AI was given no special knowledge of genetics. It learned from genomes where the right answer was already known, thanks to a reference project called Genome in a Bottle. It was then tested on samples it had never seen.
Fewer errors than the best tools of the time
In 2016, the team submitted blinded results to a variant calling challenge sponsored by the FDA, the US Food and Drug Administration. DeepVariant won the top performance award for SNPs, which are single-letter changes. On this sample, which it had never seen during training, the AI scored 99.95% for SNPs. This score, called F1, combines variants found and false alarms avoided.
On the same sample, DeepVariant made 4,652 errors per genome, compared with 9,531 for the next best tool, a cut of more than 50%. The AI also learned to handle data from other sequencing technologies. Trained on human data, it even scored 98.29% on mouse DNA.
Key figures
- 50%fewer errors (more than) per genome than the next best tool
- 99.95%F1 score for SNPs on a sample never seen in training (FDA-sponsored challenge)
- 98.29%F1 score on mouse DNA, with a model trained on human data
Keeping track of your genetic results with Mima
DeepVariant works behind the scenes, in the lab: it improves the step where changes in DNA are found. If you are offered a genetic test, for a rare disease or for your tumor, this is the step where this kind of tool can come in.
Mima helps you keep track of your reports and appointments and prepare your questions, such as: "Does this variant change anything for my treatment?" She also explains complicated words in plain language.
What this could mean for patients
Genetic tests are used more and more in medicine, for example to look for the cause of a rare disease or to find the mutations in a tumor. All of them go through this same step: spotting variants in raw data. Fewer errors at this step means less risk of missing an important change, or of flagging one that is not really there.
According to the authors, the same method can learn to read data from a new sequencing machine without rebuilding the software from scratch.
What are the limits of this study?
The study measures reading quality on reference samples. It does not show that any patient was diagnosed faster or treated better thanks to this tool. Finding a variant also does not tell you whether it is harmful: that is the job of genetics experts.
The authors admit that turning data into images is not ideal, because not all of the information fits in the picture. With some sequencing technologies, the AI loses part of the real variants along the way.
Frequently asked questions
What is DeepVariant?
DeepVariant is an AI tool built by Google researchers. It finds genetic variants in DNA sequencing data by turning the data into images that a neural network reads.
What is variant calling?
Variant calling is the step where software compares your sequenced DNA with a reference genome and lists the spots where they differ. It has to separate real variants from machine errors.
Does DeepVariant diagnose diseases?
No. It finds changes in DNA. Genetics experts then decide whether those changes matter for a person's health.
Why read DNA as an image?
Image recognition AI is good at spotting patterns. By looking at all the reads for one spot at once, DeepVariant can account for errors that are linked together, which standard tools handle poorly.
The original study
Poplin R, Chang PC, Alexander D, et al. “A universal SNP and small-indel variant caller using deep neural networks” Nature Biotechnology, September 24, 2018. DOI : 10.1038/nbt.4235
Read the full study in Nature BiotechnologyThis article summarizes a scientific study and is not medical advice.