AI improving treatments for diabetes, depression, cancer, and more

Deep learning for neurological disorders

In many neurological disorders, clinicians lack biomarkers to objectively inform a patient about likely future disease trajectory or to predict which treatment will yield the best outcomes. Albert Montillo, Ph.D., a professor in the Lyda Hill Department of Bioinformatics and the Department of Biomedical Engineering, is developing this type of personalized medicine for disorders ranging from neurodegeneration, neuropsychoses, to neurodevelopmental disorders.

To tackle this challenge, Dr. Montillo is developing new core capacity for machine learning models to learn, and applications of those improved models to medicine that integrate neuroimaging, clinical, genomic, proteomic, and metabolomic data. The improved capacity makes the models more explainable and interpretable; increases trustworthiness and fairness in predictions across subpopulations, including patients of different races and ages; and can reveal causal effects rather than merely correlations. Dr. Montillo has applied the enhanced models in conditions including depression, movement disorders, and autism spectrum disorder. In depression, he has shown that that these models can accurately predict the level of response each individual patient with major depressive disorder will obtain across antidepressant therapies. In Parkinson’s disease, the models can predict the future trajectory of cognitive and movement symptomatology years in advance. In autism, Dr. Montillo is using the models to identify biomarkers for early detection and gene targeting.

Advancing radiation therapy

Radiation oncology is a major pillar of cancer therapy and constitutes an important part of care for many patients. Planning the most effective protocol is pivotal for effectively treating cancers. Delivering too little radiation can lead to cancer recurrence while delivering too much can permanently injure surrounding tissue. However, developing a plan can take radiation oncologists up to a week, and there can be considerable variation among oncologists on determining the most effective plan.

David Sher, M.D., Professor of Radiation Oncology and Chief of Head and Neck Radiation Oncology, worked with colleagues to create an AI-based tool to optimize planning for head and neck radiation therapy. A study showed that this tool helped physicians develop plans that effectively treated tumors while significantly reducing radiation doses to normal tissues in up to 75% of cases.

Enhancing medical education

Like most medical schools, UT Southwestern utilizes simulated patient interactions to train students. They examine patient actors working from a script to feign various medical conditions, writing notes and making care decisions much like experienced physicians. These 15-minute interactions are analyzed by educators, who grade students on a variety of metrics. Andrew Jamieson, Ph.D., Assistant Professor in the Lyda Hill Department of Bioinformatics, explained that scoring these interactions – up to 10 per student – takes enormous time and focus and are subject to significant human error.

Dr. Jamieson and his colleagues developed AI algorithms that automatically score the notes that medical students produce from these interactions. They recently implemented this system in the fall of 2023 after developing a protocol for monitoring performance and safety that was signed off by key leaders and stakeholders at UTSW, reducing human grading by 90%. Their future efforts will dissect video content from these simulations. AI will eventually help educators evaluate factors including tone of voice, body language, and eye contact, speeding analysis and eliminating variability among educators.

Medical students also learn about AI in a dedicated course led by Gaudenz Danuser, Ph.D., Professor of Cell Biology and inaugural Chair of the Lyda Hill Department of Bioinformatics, one of only a few such classes taught in medical schools across the country. At the end of the class, students test their ability to beat an AI algorithm by performing a retrospective analysis using real data from Children’s Health to predict which patients on heart-lung bypass machines suffered fatal brain injuries.

Complementing genomics

Genomics is a vital tool in oncology, helping researchers better understand the behavior of cancer cells and allowing physicians to tailor treatments. However, even though regions within the same tumor are often genomically distinct, it is impractical to profile multiple regions in a tumor because these tests can be expensive, often costing thousands of dollars.

Satwik Rajaram, Ph.D., Assistant Professor in the Lyda Hill Department of Bioinformatics, and his colleague Payal Kapur, M.D., (pictured) Professor of Pathology and Urology, are using AI to overcome this problem by predicting genomics from the appearance of cells on pathology slides. They previously reported that they could predict mutations in BAP1, a key driver gene in renal cancers. In recently completed work, they were able to predict the genomic angiogenesis score, one of the best indicators of response to drugs targeting blood vessels. Importantly, on clinical trial data, their AI-based readout was able to match the performance of state-of-the art genomic tests at a fraction of the cost. These savings mean tests can be performed multiple times on the same patient to identify areas that are in different genomic states, and thereby better anticipate tumor behavior.

Detecting cell types in tissue samples

Tissue samples are routinely collected from patients and placed on slides for interpretation by pathologists, who analyze them to make diagnoses. This process is time-consuming, though, and interpretations can vary among pathologists, explained Guanghua Xiao, Ph.D., Professor of Data Science in the Quantitative Biomedical Research Center, Peter O’Donnell Jr. School of Public Health, Biomedical Engineering, and the Lyda Hill Department of Bioinformatics.

In addition, the human brain can miss subtle features present in pathology images that might provide important clues to a patient’s condition.

Dr. Xiao and his team are developing and validating AI models that convert tissue images into spatial maps of cells and tissue structures. This involves detecting and classifying different cell types from tissue images, a crucial step in understanding and characterizing the tumor microenvironment. Moreover, their new graph-based AI spatial model has provided a powerful tool for studying these environments in greater detail.

His team has developed an entire pipeline for tissue imaging analysis. This pipeline encompasses data processing, imaging analysis, deep learning, spatial modeling, and the development and validation of clinical outcome prediction models, merging methodology and biomedical applications.

Identifying pathogens in the gut

The average human body hosts approximately 2 kg of bacteria, encompassing a staggering 100 trillion microbes. These bacteria, predominantly harmonious with their human hosts, are increasingly recognized for their significant role in influencing various physiological conditions.

To delve deeper into the impact of bacteria on human health, Xiaowei Zhan, Ph.D., Associate Professor in the Peter O’Donnell Jr. School of Public Health and the Center for Genetics of Host Defense, and colleagues employed AI to develop predictive models based on extensive human disease studies. Their research has successfully achieved high rates of accuracy in predicting insulin resistance in Type 2 diabetes and inflammatory bowel diseases through noninvasive sequencing of the gut microbiome.

Among the diverse bacterial population, some are pathogenic and responsible for infectious diseases, necessitating precise and effective antibiotic treatments. Traditional methods for identifying these pathogens are time-consuming and often suffer from poor reproducibility.

The researchers compiled a comprehensive dataset of 3,393 bacterial isolates and their resistance profiles against 29 different antibiotics. Using this data, they developed an AI model that predicts resistance to all 29 antibiotics across nine bacterial species. This model could play a crucial role in combating antibiotic resistance globally.

Helping lung cancer patients

Combining artificial intelligence with imaging analysis, Yang Xie, Ph.D., (pictured above, center) Associate Dean for Data Sciences, is finding new ways to formulate personalized treatment strategies for non-small cell lung cancer patients. The technique efficiently distinguishes lung cancers likely to respond to specific treatments, guiding the selection of more effective therapeutic approaches.

Dr. Xie, also Founding Director of both the Quantitative Biomedical Research Center (QBRC) and the Pediatric Cancer Data Commons (PCDC), is a leading expert in predictive modeling and clinical outcome prediction. Her team excels in developing AI models that significantly enhance clinicians’ ability to personalize treatment plans, thereby impacting patient care and clinical biomarker discovery.

Dr. Xie’s recent research also involves using ChatGPT to interpret clinical notes, demonstrating significant promise for applying large language models in clinical operations. This pioneering work is a testament to her commitment to leveraging advanced technologies in health care, continually pushing the boundaries of data science in medicine.

Better understanding cancer and immunity

Tao Wang, Ph.D., Associate Professor in the Peter O’Donnell Jr. School of Public Health and the Center for Genetics of Host Defense, has been working on developing and applying novel AI approaches to demystify the roles of the adaptive immune system in tumors and other human diseases. Dr. Wang is the first investigator in the field to have developed a novel AI model, named pMTnet, to predict the interactions between T cell receptors (TCRs) and T cell antigens (Nature Machine Intelligence, 2021). This is a platform technology that would widely impact the design and implementation of various immunotherapies and disease diagnostics, such as improvement of engineered T cell receptors for TCR therapies and TCR-like drugs, antigen selection for tumor vaccines, and biomarkers for monitoring T cell function.

For example, in collaboration with Yang Xie, Ph.D., Associate Dean for Data Sciences, and David Gerber, M.D., Professor of Internal Medicine, Dr. Wang built a biomarker based on pMTnet to predict immune-related adverse events of immune checkpoint inhibitors on patients. Dr. Wang is also the first investigator to develop AI-based methodologies to integrate single T/B cell gene expression with TCR/BCR-sequencing data (Nature Methods, 2021 and Nature Machine Intelligence, 2022), which revealed critical insights for the functions of T and B cells in various biomedical contexts. Dr. Wang is now furthering his research in predicting the pairing between antibodies and antigens using novel AI methods, which is expected to lead to breakthroughs in antibody drug selection and design.

Automating scientific discovery

Many AI systems rely on neural networks that emulate the way biological neurons signal one another. Although this tool has found significant use in applications such as image and speech recognition, conventional neural networks have significant drawbacks, said Milo Lin, Ph.D., Assistant Professor of Biophysics and in the Lyda Hill Department of Bioinformatics and the Center for Alzheimer’s and Neurodegenerative Diseases. Most notably, they often don’t generalize well beyond the data they train on, and the rationale for their output is a “black box,” meaning there’s no way for researchers to understand how a neural network algorithm reached its conclusion.

To address both issues, Dr. Lin and his colleagues developed a method they call deep distilling. Using limited training data, deep distilling automatically discovers algorithms, or the “rules” to explain observed input-output patterns in the data, then translates them into succinct computer code so users can read them.

Deep distilling could eventually be applied to the vast datasets generated by high-throughput scientific studies, such as those used for drug discovery, and act as an “automated scientist” – capturing patterns in results not easily discernible to the human brain. Deep distilling also could potentially serve as a decision-making aid to doctors, offering insights on its “thought process” through the generated algorithms, he added.

ChatGPT and clinical notes

ChatGPT, the artificial intelligence (AI) chatbot designed to assist with language-based tasks, can effectively extract data for research purposes from physicians’ clinical notes, UT Southwestern Medical Center researchers reported in a study in NPJ Digital Medicine. The findings show it could significantly accelerate clinical research and lead to new innovations in computerized clinical decision-making aids.

“By transforming oceans of free-text health care data into structured knowledge, this work paves the way for leveraging artificial intelligence to derive insights, improve clinical decision-making, and ultimately enhance patient outcomes,” said study leader Yang Xie, Ph.D., Professor in the Peter O’Donnell Jr. School of Public Health and the Lyda Hill Department of Bioinformatics at UT Southwestern. Dr. Xie is also Associate Dean of Data Sciences at UT Southwestern Medical School, Director of the Quantitative Biomedical Research Center, and a member of the Harold C. Simmons Comprehensive Cancer Center.

Much of the research in the Xie Lab focuses on developing and using data science and AI tools to improve biomedical research and health care.

To determine whether ChatGPT could convert clinical notes to structured data, Dr. Xie and her colleagues had it analyze more than 700 sets of pathology notes for lung cancer patients to find the major features of primary tumors, whether lymph nodes were involved, and the cancer stage and subtype. Overall, Dr. Xie said, the average accuracy of ChatGPT to make these determinations was 89%, based on reviews by human readers. Their analysis took several weeks of full-time work compared with the few days it took to fine-tune data extraction from the ChatGPT model. This accuracy was significantly better than other traditional natural language processing methods tested for this use.

Dr. Xie noted that the results were strongly influenced by what prompts ChatGPT was given to perform each task – a phenomenon called prompt engineering. Providing multiple options to choose from, giving examples of appropriate responses, and directing ChatGPT to rely on evidence to draw conclusions improved its performance.

She added that using ChatGPT or other large language models to extract structured data from clinical notes could not only speed clinical research but also help clinical trial enrollment by matching patients’ information to clinical trial protocols. However, she said, ChatGPT won’t replace the need for human physicians.

“Even though this technology is an extremely promising way to save time and effort, we should always use it with caution. Rigorous and continuous evaluation is very important,” Dr. Xie said.