Many people with genetic diseases still have problems getting a diagnosis in a short time. A correct analysis of DNA usually takes while and is not always accurate, which means that the right treatment is delayed or cannot be carried out at all. The Berlin-based AI start-up Nostos Genomics has developed a solution for this: Algorithms reduce the time needed for an analysis from twelve hours to just two minutes. Co-founder and CTO Rocío Acuña Hidalgo and Head of AI David Neville told us in an interview how exactly this works.
Hi Rocío and David, thank you for taking the time to do this interview! Summer is slowly starting so we would like to know: what’s your hot tip for early summer nights in Berlin?
Grab a beer and head to Admiralbrücke to catch the sunset! Or, if you need to cool down, take the S-bahn to Schlachtensee for a quick dip in the lake.
Rocío, you founded Nostos Genomics together with your co-founder David Gorgan in 2018. How did you come up with the idea to start the company?
Rocío: I first got involved with organizations for patients with genetic diseases during my PhD and postdoc. These patients were trying to map a pathway to develop treatments for their disease and I started advising them on their scientific strategy. As we worked together, I learned more about the challenges people with genetic diseases face, including getting a diagnosis to know what disease they have and subsequently receiving appropriate medical treatment. I felt very inspired by the people I was working with and decided to turn the knowledge I had gained during my medical training and PhD into something that could have a meaningful impact in the real-world.
With Nostos Genomics, you are helping people with genetic diseases with the help of AI. What is the current status in this field and how are you trying to fill that gap?
Rocío: Over the last decade, there have been a lot of technological developments that have transformed how we diagnose genetic disorders. While the technology to read DNA and to prepare the data for analysis has changed significantly, the process through which genetic data is analysed hasn’t evolved as much. Currently, this step is usually carried out manually by an expert in genetics who analyzes the patient’s genetic data, integrating information from different sources to try to reach a diagnosis. This process is laborious, expensive and the results are not always reproducible, which for the patient translates into having to wait several months to get their test results which in more than 50% of cases are not useful nor conclusive.
At Nostos Genomics, we fill this gap by developing algorithms to automate this process, shortening the time spent on analysis from 12 hours to just 2 minutes per sample and extracting more useful information from the genetic and clinical data analyzed. This translates into shorter turnaround times for genetic tests, lower costs and more accurate results for patients.
Your solution is based on an AI-driven variant interpretation that is supposed to take less than two minutes. How exactly does this work?
David: Our proprietary system behaves in a manner very similar to a human expert. First it accesses all relevant data repositories to retrieve genetic, biological and clinical data so that each genetic variant can be put into context by integrating this information. Next, the AI learns the cause-effect relationships from the data, which helps predict the clinical effect for each genetic variant. Once the AI has been trained, we can use the same algorithm to process patient samples and determine which genetic variants observed in the patient are most likely to be clinically meaningful.
What advantages does your solution “AION” have compared to conventional genetic variation interpretation solutions?
David: Most existing solutions for variant interpretation rely either on automatized expert systems which are easy to interpret but limited in their capabilities or on approaches which are extremely powerful but hard to interpret. The guidelines from the American College of Medical Genetics (ACMG), which nowadays are the golden standard for variant interpretation, are a good example of the first type of solutions. This approach is an expert rule system which is easy to understand but fails to provide a useful interpretation for most genetic variants. Solutions based on Deep Learning are an example of the second type of systems, which often have an excellent performance but are unsuitable for clinical use because understanding why the system made a prediction requires the use of special techniques for decoding.
Our solution “AION”, achieves the best of both worlds. It structures and uses information in a completely clear and transparent fashion comparable to the ease of understanding the ACMG guidelines. However, it is a machine learning system and therefore is able to analyse a large number of cases in a fraction of the time with extremely high precision. Furthermore, our solution is able to detect cases where the data is insufficient to support a reliable interpretation even-though the diagnosis might be highly probable. This ability in turn allows for reanalysis of the case after acquisition of additional information.
It is quite stunning to see what AI can already do these days. Do you have a vision for the future on what else you might like to do with the help of AI?
David: Current AI solutions for rare disease diagnostic focus on a human-automation interaction where the human expert uses the AI tool to automate a very laborious task such as variant interpretation. I would like to see in the future AI moving away from such a scheme towards more human-autonomy teams (HAT). The difference is that in the former case, the AI is used as a tool by the human expert to rapidly perform a time-consuming and laborious task. In the latter case instead the AI functions as a team member, providing insights of its own, challenging the human expert reasoning and ultimately continuously learning from the interactions with the human expert. In this respect I’d like in the future AI-based diagnostics becoming not only a bioinformatics tool but rather a core member of a team.
You recently raised €5 million in a seed round, congratulations on that! What do you plan to invest the funding in?
Rocío: Thank you! We’re planning on growing our company to further improve our technology for genetic disease diagnostics, launch clinical validation studies and expand to new European markets and the US.
When it comes to the field of AI, Berlin’s scene is increasing and expanding. What makes the location so special and why did you found your company here?
Rocío: Berlin is full of world-renown universities and research centers and has a budding tech scene, which together attract a lot of talent and innovation. Actually, I first moved to Berlin to carry out research at the Max-Planck-Institute, which I had been dreaming of doing for almost a decade! I then joined one of the first Entrepreneur First programs that took place in Berlin, met my co-founder David Gorgan and the rest is history.