What do frictionless parking experiences and life-saving drugs have in common? For data scientist June Guo, the answer lies in setting AI algorithms loose on vast troves of data. Before setting his sights on human biology, Guo worked at Metropolis Technologies, a company focused on transforming the parking experience through the use of advanced computer vision technology to enable “checkout-free” payment systems. As a founding member of its machine learning team, Guo was instrumental in helping secure a $167 million Series B funding round. Before earning his Ph.D., Guo worked in the automotive industry for a few years focused on developing early autonomous driving technologies. “At the time it was called advanced safety, not autonomous driving,” he said. “I worked on computer vision, sensor fusion, and radar design.”

From parking lots to drug discovery for autoimmune and chronic inflammatory diseases

Now the vice president, of artificial intelligence and machine learning at Athos Therapeutics, Guo is applying machine learning to decipher the hidden patterns of genetic links behind autoimmune diseases, chronic inflammatory diseases, and cancer. “I joined Athos two years ago — my first biotech startup,” he said. “It’s been fun learning about biology, chemistry, drug discovery, and the applications of AI enhancing computational biologists, chemists, drug discovery, and the whole process of precision medicine.”

The transition, however, was not without its surprises. “When I first joined Athos, I found the complexity of data collection is way more complicated than the automotive industry,” he said. “When I first came here, I was like ‘Oh, what are these data?'”

Autonomous AI for precision medicine

Athos Therapeutics is developing an autonomous AI/ML platform to assist with research on autoimmune diseases, chronic inflammatory diseases, and cancer. This platform integrates a variety of functionalities including RHEA, a tool used for integrated and automated transcriptomics, which involves the study of RNA transcripts produced by the genome. Then there’s TETHYS, an LC-MS-derived autonomous proteomics platform, and DIONE, a proprietary deep machine learning framework for patient molecular subtyping and the development of precision therapeutics.

When referring to the company’s autonomous system, Guo says “It starts from the digital information.” The platform can process raw FASTQ files or omics data such as transcriptomics, genomics, or proteomics. Quality control is the first crucial step, as these datasets can be massive, with each file potentially reaching five gigabytes. Traditionally in computational biology, pipelines perform preprocessing, processing, and post-processing all at once, leading to wasted time if the data is contaminated. “It can take days or even months to process all these datasets when we talk about thousands of them,” Guo notes.

Keeping human experts firmly in the loop

Athos’ autonomous platform gives users control at each step, notifying them when low-quality or contaminated data is detected. “If no defect is detected, everything else will be autonomous. But we warn our users, saying ‘Hey, we detected something in the dataset, do you want to proceed?’ to give transparency and control for our users to make a decision,” Guo explains. Once the data is quality-checked, the platform can fully automate the omics data processing and machine learning if desired. “You can also just process the data without using machine learning.”

But while it’s autonomous, the goal isn’t to put physicians in the passenger seat. In the automotive industry, there are five levels of autonomous driving with the fifth being complete autonomy. “Level five is the goal of all OEMs and startups. Basically, hands-off, eyes-off, you don’t need to do anything,” Guo said. “Drug discovery is different. No one fully trusts AI-discovered drugs. We pay a lot of attention to the safety and confidentiality of the data and the safety of AI. Can we fully trust a drug discovered by artificial intelligence or machine learning algorithms?”

Safeguarding patient eata and validating results

The difficulty in clearly answering those questions informs Athos’ approach of letting humans ultimately make the decisions. The company takes numerous steps to ensure patient privacy and data security. Omics data—transcriptomics, proteomics, and genomics—is de-identified before being analyzed, with only metadata about disease history being accessible. This safeguards patient confidentiality while still providing insights.

While AI-powered tools can help pinpoint potential subtypes within a given disease, Athos loops in computational biologists and doctors for their perspective. These experts have the final say in validating the AI’s findings, ensuring that potential drug targets or treatments are thoroughly vetted before proceeding. As Guo explains, “In terms of AI safety, we have tools to process the omics data and feed them into our machine learning algorithms to identify subtypes. To develop precision medicine, we need to understand subtypes instead of developing a generic drug with 20-30-50% efficacy. To make highly efficacious drugs for patients, especially those with autoimmune diseases like Crohn’s or colitis, we identify genes and send them to computational biologists and doctors to make sure they see the genes and make the final decision.”

The promising future of neural networks in precision medicine

Despite the complexities of navigating biological data, Guo is upbeat about the promise of AI in precision medicine. “When I started training cancer datasets and also inflammatory bowel disease datasets when I saw the results, I was like ‘Wow, this is really powerful,'” he recounts. “There’s no way I could really see this data and understand any structures or biological sense of it. But when I started training and validated the results, I was totally shocked about the accuracy and power. The potential of these neural networks and what they can do when we have a lot of data — we can really make precision medicine and personalized treatment possible.”