If you or a loved one have ADHD, you know that it is much more than just an attention disorder. ADHD often comes hand-in-hand with other health concerns like type 2 diabetes and obesity. While we know from decades of research that genetics plays a major role in both ADHD and these related conditions, figuring out exactly how has been a huge challenge.
Why? Traditional genetic models are a bit like trying to understand a complex story by only counting specific words. They might tell you a “risky” word appears often, but they miss all the crucial context that gives the story its meaning.
Seeing the Bigger and More Complete Picture in DNA
In our recent study, we developed a new framework using artificial intelligence (AI) to build more powerful and fair genetic prediction models. To prove our methods worked, we first applied them to a large dataset for type 2 diabetes. The results are promising – now, we are preparing to use this same technology to better understand the genetics of ADHD and its connection to cardiometabolic health.
Here’s a look at the two big problems we tackled.
Giving DNA a “Resume”: Context-Informed Genetic Data Matrix
A standard genetic analysis might tell you that you have a specific genetic variant. But that information alone is not very helpful. We wanted give our AI model more than just these genetic variants, but also known information related to function or disease – to help make sense of the genetic variants. Is that variant located within an important gene? Is it near a genetic “switch” that controls other genes? Has it been linked to related conditions in past studies?
To achieve this, we created what we call a Context Informed Data matrix. You can think of it as giving each piece of a person’s genetic data a detailed resume. For every genetic variant we analyzed, we gave our AI model not just the variant itself, but also its:
- Job Title: Its function in the genome, like whether it is part of a gene or a regulatory element.
- Work History: Its known association with type 2 diabetes from previous large research studies.
- Professional Network: Its connection to risks for other related disorders, like obesity or high cholesterol.
By bundling all this information together, we provided a much richer picture for our AI to analyze. We used a type of AI called a Convolutional Neural Network, which excels at finding complex patterns, to learn from this context-rich genetic data.
Building Fair Genetic Models for Everyone
A persistent issue in genetics is ancestry bias. If a model is trained on data from one ancestral group- usually people of European ancestry, it can get confused when used in people of different ancestry. It might learn to predict a person’s ancestry instead of their actual disease risk, leading to a biased model that fails to work for everyone.
To counteract this issue, we used a technique called adversarial learning. Here is how it works:
We gave our AI model two jobs at once. Its main job was to predict disease risk. But we also gave it a second, competing job: to try to NOT predict a person’s ancestry. Normally, at many steps during model training, the weights of the model are changed such that they better perform the task or tasks. For the ancestry task, we used what is called a gradient reversal layer to make the model do the opposite and change weights such that they perform worse at predicting ancestry. This forced our model to find the true, underlying genetic signals of the disease that are common across different ancestries.
What We Found:
Our tests with type 2 diabetes were a success. We found that:
- Our models were more accurate. Both our context-informed and standard AI models significantly outperformed traditional methods for predicting type 2 diabetes.
- Our ancestry adjustment worked. The adversarial technique successfully prevented the model from using ancestry information, all without reducing its predictive power.
What’s Next: ADHD and beyond
Now that we have proven these methods work, our next major goal is to apply them to ADHD. By using this powerful approach, we aim to:
- Build a more accurate and fair genetic risk model for ADHD.
- Investigate the shared genetic links between ADHD and the cardiometabolic conditions that so often co-occur.
Summary
We’re using AI to look at genetics in a smarter, fairer way—adding rich biological context and reducing ancestry bias. After a successful test with diabetes, we’re now applying this method to ADHD. Our hope is that this work will pave the way for a deeper understanding of the biology of ADHD and co-occurring cardiometabolic outcomes. Ultimately, it could lead to better, more personalized ways to support the long-term health of people with ADHD.
This blog post is written by Dr Eric Barnett (Postdoc at TIMESPAN partner SUNY)