This article is more than 5 years old. Information may no longer be current.
AI identifies pathogenesis pathways, risk for Crohn’s disease
Click Here to Manage Email Alerts
AVA,Dx, a machine learning method for risk prediction of Crohn’s disease, revealed new potential CD genes and accurately identified the disease in a group of more than 3,000 individuals, according to a recent study.
Yana Bromberg, PhD, associate professor in the department of biochemistry and microbiology at Rutgers University New Brunswick, and colleagues used person-specific coding variation in genes from a panel of 111 individuals to create disease prediction models. AVA,Dx was used to examine four panels of individuals containing 2,793 individuals with CD and 697 individuals without CD, identifying 16% of CD patients at 99% precision and 58% of patients with 82% precision.
Bromberg spoke with Healio Gastroenterology and Liver Disease regarding the study’s results as well as emerging developments in AI as a diagnostic aid.
Healio: What are the major takeaways from this study?
Bromberg: What we found is that there are some pathogenesis pathways and molecular pathways that are affected in current patients, and potentially these can be used by drug researchers to identify new drugs for CD. Our way of looking at genome data may be useful for cases where it's unclear whether the person has CD or not, definitely not as a diagnostic, but maybe just one more thing to put into the toolbox.
Healio: Are there any similar research studies that have been done on AI specific to CD?
Bromberg: There has been a lot of work looking at CD. Many genome wide association studies, for instance, have been performed for CD. There are some very well-known genetic loci in the genome that have mutations that are associated with CD.
Healio: What current limitations would you say exist in the current AI, or specifically, Ava,Dx, that you feel could be addressed, or could improve its performance going into the future?
Bromberg: Although machine learning in biology in general is still in its very early stages, it's already very useful. But it's not yet where it needs to be to be directly applicable. This study and studies like it we hope will help the transition into application, but it's not yet there.
One thing that everyone will tell you — we don't have enough data to develop models that are going to be generically useful. For the number of features, the number of things that this machine needs to learn to be able to recognize CD, for instance, requires significantly more data. So, we have more patients with more sequenced genomes, better defined phenotypes. The big deal is diagnosing things correctly and using the same criteria for many patients. Once we have that, I think that we could get progressively better.
Healio: What would you say would be the major benefits of using AI similar to Ava,DX going forward to help diagnose CD?
Bromberg: This is not a diagnostic, but a companion. If you look at the study, there is a sufficient number of healthy people who are scoring high on the CD scale, but not as high as some fraction of the CD patients. For some small fraction of CD patients, Ava is able to identify CD.
Our method doesn't require taking a biopsy of the tissue, which is probably much better for people. It's not invasive. But I wouldn't call it a diagnostic, so you couldn't diagnose CD. We’re not there yet. We’re trying to get there hopefully in the next 10 to 20 years — we can — and in that case, it will probably ease the burden of diagnosis from invasive studies to less invasive ones.
Healio: Were there any standout results from this research that surprised you or you didn't expect going in?
Bromberg: One thing that surprised me was how few people we needed to train this model for this case. It's not wonderfully accurate, but the fact that we were able to pick up signal with so few people — we had only 111 people in the training set — that was very surprising to me. That basically means that CD signatures are very explicit, or some of the CD signatures are very explicit.
The second part that was surprising is that we picked up genes that were not identified with the standard studies. So, we picked up the ones that were identified, but also additional genes. And they seem to be involved in the right pathways.
Healio: If you were to do a follow up to this study or something in a similar vein, what would that look like, or what would you be looking to improve upon?
Bromberg: Right now, we’re looking at other diseases. We have collaborators looking at Tourette disease and chronic obstructive pulmonary disease. I'm reaching out to people.
We have new developments in the Ava,Dx pipeline that we think will make it higher resolution and more applicable to find disorder mechanisms. With CD, we needed very few samples to identify the signatures. That's not the case for many other disorders. So maybe if we have a higher-resolution algorithm, we would be able to use it for those disorders.
Healio: Do you see the applications of AI growing and improving in your daily work?
Bromberg: AI is growing. Computer science is growing, getting better and better. We know more about which algorithms should be used with which kind of data and we're trying to do things in a more rigorous manner. I think that AI will be a large component of medicine going forward, simply because we're getting a lot more data that we can make inferences on the basis of it. It’s very hard for one person to take all data into account, but a machine can. Note I don't necessarily see AI as being the doctor. I see this being a tool in the doctor’s toolbox. – by Eamon Dreisbach
Disclosures: Bromberg reports support from the NIH, the NIG U24 MH06845 grant and the PhRMA Foundation’s Informatics Research Starter grant.
Read more about
Click Here to Manage Email Alerts