Artificial Intelligence | Genetic Testing

How genetics experts use artificial intelligence (AI) to better classify gene variants

Genetics experts develop and apply artificial intelligence to classify variants of uncertain significance (VUS), leading to actionable answers for more patients

Written by Yuya Kobayashi, Senior Clinical Genomic Scientist, Clinical Science and Interpretation at Invitae

As clinical genetic testing has become increasingly common in medicine, our experts have seen and interpreted a lot of genetic variants. While that has helped millions of patients and their healthcare providers make informed medical decisions, we know we can do more.

Why? Because today, while we classify many genetic variants as either disease-causing (pathogenic) or not (benign), we’re also classifying variants of uncertain significance (VUS).

Why do variants of uncertain significance (VUS) exist, and what can we do?

Genetic variants are classified as pathogenic or benign, but until we have enough data to know which variant is which, they’re called variants of uncertain significance. Many possible genetic variants exist with about three billion positions in the human genome, each potentially being one of four letters (A, G, T, or C).

To date, scientists haven’t yet had time to study them all or well enough to be confident about their impact. We still need more data for any given VUS to know whether or not it’s associated with a health condition.

With more data, each of these VUS should eventually be resolved. Most of them are expected to be benign (because most genetic variants are), but some will turn out to be pathogenic. Meanwhile, patients wait for information about a variant that could help them make informed medical decisions with their provider.

Two possible solutions exist to speed up this process:

1. Generate more data.

We generate more data with every patient we test. If we find a VUS in a patient clinically affected by the disease, the genetic variant may be a possible cause.

However, finding the same VUS in a healthy family member without the condition decreases the chances that the variant is causal. This clinical correlation (or anti-correlation) helps us better understand VUS.

Generating additional data can also happen outside the context of patients. Thousands of scientists worldwide study genetic variants and their potential impacts at the molecular and cellular levels.

When we tell scientists about a VUS we found in patients, they can design scientific experiments to study it. Whenever they publish those findings, our scientists work to incorporate that data into our interpretation.

2. Better utilize data.

Since data are scarce, we want to get as much information out of every piece we have. Today, we rely on the expert experience of highly-trained clinical geneticists and established interpretation guidelines to decide how informative a given bit of data is. We’re constantly working to improve those guidelines and develop new tools as we better understand genetics, molecular and cellular biology, and disease.

At Invitae, we focus a lot of energy on generating more data (clinical, molecular, and cellular) and ensuring we get the most out of this information. For the latter, we’ve recently had a lot of success by combining our deep clinical genomics expertise and rich genetics database with Artificial Intelligence (AI) technology.

Today, we rely on the expert experience of highly-trained clinical geneticists and established interpretation guidelines to decide how informative a given bit of data is. We’re constantly working to improve those guidelines and develop new tools as we better understand genetics, molecular and cellular biology, and disease.

What is AI, and how can it play a role in variant interpretation?

When you hear “AI,” you might think of self-driving cars and robots. But practically speaking, AI is a computational tool that seeks to identify patterns in data.

For example, some cars have a backup camera that identifies moving vehicles and pedestrians to warn the driver. They might have a GPS navigation system that uses traffic pattern data to recommend the best route and predict an estimated arrival time. That technology is AI, and it helps make driving safer.

The modern car-driving experience is similar to AI-assisted variant interpretation. AI tools help scientists identify patterns in the data that are either too complex for human brains to identify or that are easy to miss, allowing them to be more accurate and efficient. But at the end of the day, we still rely on clinical geneticists, who understand the limitations of AI, to be alert and in the driver’s seat.

Invitae’s clinical geneticists and AI scientists work closely to develop the technology, relying on their expertise. However, AI tools also need a large database of known pathogenic and benign variants to find patterns in the data, a process known as “training,” much like how a scientist might use positive and negative controls in experiments.

AI asks, “What are the defining characteristics of variants that distinguish pathogenic from benign ones?” The tool then builds an algorithm (also called a model) to make predictions of other genetic variants by quantifying how much they look like known benign or pathogenic ones based on those characteristics.

The remaining portion of the known pathogenic and benign variants called the “holdout set,” is used to evaluate the performance of these models. It looks at whether the patterns identified in the models are generalizable enough to correctly distinguish the pathogenic and benign variants it didn’t look at during the training step.

Invitae’s clinical geneticists and AI scientists work closely to develop the technology, relying on their expertise. However, AI tools also need a large database of known pathogenic and benign variants to find patterns in the data, a process known as “training,” much like how a scientist might use positive and negative controls in experiments.

This quality control step is critical to ensuring that the AI models are sufficiently accurate for the scientists to use in the genetic variant interpretation process.

How does AI help clinical geneticists interpret variants?

We’ve thoroughly tested and implemented seven classes of AI-based models.¹ These leverage various types of biological data, including protein structure and stability, DNA sequence context, mutational hotspots, mRNA splicing, population allele frequency, evolutionary conservation, and various high-throughput functional assay data. These models assist clinical geneticists by providing valuable information for over 3,000 disease-associated genes.

AI-assisted genetic variant interpretation offers two key benefits:

1. AI can extract more information from existing data.

The population allele frequency model is an excellent example of extracting more information.

Population allele frequency measures how common a particular genetic variant is in a population (typically a generally healthy one). Experts have long recognized the allele frequency measurement among a cohort of healthy individuals as a valuable resource for variant interpretation.

Suppose we observe a variant to be common among healthy individuals. In that case, it’s pretty strong evidence that the genetic variant is not pathogenic. The challenge, however, is figuring out how common is too common for us to be able to rule out pathogenicity. We expect the answer to be different from one gene to the next.

For genes associated with certain rare pediatric diseases, even just a few observations of a variant in the healthy adult population suggest it’s probably not pathogenic. However, genes related to late-onset diseases like cancer must be much more common before we’re confident about their pathogenicity.

Our recently published population frequency modeling helps us determine the appropriate threshold for those too common to cause disease on a gene-by-gene basis. The model combines population allele frequency data from public databases like gnomAD and layers on gene-specific properties that influence the prevalence of pathogenic variants in healthy adult populations. Then, it uses known pathogenic and benign variants to identify the gene-specific pattern in the data to calculate the appropriate threshold for that gene.

Compared to traditional scientist-driven approaches for drawing these thresholds, the AI-based population frequency model resulted in more accurate thresholds with fewer miscategorized variants. It also showed we should classify nearly 15,000 of previously classified VUS as benign.²

2. AI can help safeguard against over-interpreting data.

AI-assisted variant interpretation isn’t just about resolving VUS. AI can also be a valuable tool that protects scientists from reading too much into data. An example includes a series of models built using high-throughput cellular assays.

High-throughput cellular assays are experiments that evaluate the functional consequence of large numbers of variants at once. These are relatively new and increasingly popular technology. However, it’s unclear how helpful they are for interpretation.

Just because a variant has some functional impact doesn’t necessarily mean it will lead to disease. Similarly, just because a variant didn’t have an observable functional impact in this particular experiment doesn’t mean it won’t have other consequences that can lead to disease.

Recently, we used our AI modeling strategy to evaluate how useful these high-throughput cellular assays may or may not be. We attempted to build AI models based on data from about fifty published high-throughput cellular assays.

As it turned out, fewer than ten assays had data where the AI model could identify a valuable pattern between pathogenic and benign variants. In other words, more than 80% of these high-throughput assays produced potentially misleading data for genetic variant interpretation.¹

Without the AI tool warning scientists not to be overly confident in these data, incorrect variant classifications for patients could have occurred.

On the other hand, for the remaining 20% of the high-throughput assays, scientists can now be much more confident when applying those data to variant interpretation.¹

Accelerating the genetic testing industry

The rapid adoption of clinical genetic testing in recent years has helped millions of patients and their physicians make informed healthcare decisions. However, there’s still a lot of biology to be learned, and nothing underscores that fact more than the existence of VUS.

AI is powerful in the field of genetic testing and has profound implications for patients—but only when thoughtfully developed and carefully used. That’s why we rely on our clinical geneticists to guide the development of this technology alongside our AI scientists. Expert-guided AI can quickly advance our industry’s ability to provide patients with more accurate answers today.

To learn more about Invitae’s expert-driven, AI-enabled approach to genetic testing, read more on Invitae Generation™ here.

For more information on our genetic testing, visit the Invitae website.

References

1. Invitae internal data on file.
2. Invitae. Population Frequency Modeling. 2022.
3. Invitae internal data on file.