Biomarkers are tools or measures used to detect biological changes in the brain that are difficult or impossible to assess directly. Researchers have posed several ideas for biomarkers of cognition. Currently, a neurologist assesses cognition through standard memory tests and questionnaires. While these tests are important for detecting changes in cognition, they lack the sensitivity to detect the subtle changes in brain function that occur before obvious clinical symptoms appear.
One of the most promising concepts for a cognition biomarker may be identifying the patterns of brain activity that relate to declining memory. Cognition involves networks of neurons communicating through electrical signaling at synapses. Clinicians can measure these electrical activity patterns in humans using non-invasive methods, like an EEG, which involves placing electrodes on the scalp to detect the brain’s electrical activity. However, precisely how and which changes in brain activity relate to cognitive function remains unclear. Furthermore, because EEGs record the collective activity of many neurons, it is difficult to identify which signals meaningfully contribute to memory.
Fortunately, over the past decade, remarkable progress has been made in building tools that address this challenge in animal models. Dr. Zwang co-developed a flexible probe that can be safely implanted into a mouse’s brain to record the electrical activity of thousands of neurons for months. Dr. Zwang has implanted these probes in a tau mouse model and recorded brain activity for six months—a period spanning before and after tau pathologies develop—while also periodically testing cognitive performance.
This rich dataset was too large for typical statistical analysis methods, which led Dr. Zwang to collaborate with Dr. Holbrook, an expert in biological applications of Artificial Intelligence (AI) methods. In this proposal, they hypothesize that carefully constructed AI models can learn from the vast dataset of brain activity to accurately predict future cognitive decline. If successful, these insights could lead to the development of AI-based cognitive biomarkers using data acquired from a clinically relevant method, such as EEG.
Drs. Holbrook and Zwang proposed three aims for this project. All aims use an AI framework called Model-agnostic Graph Neural Network (MaGNet). MaGNet is a computational framework that Drs. Holbrook and Zwang believe will be able to forecast specific outcomes, such as cognitive performance, by learning from both broad patterns (like the electrical activity of networks of neurons) and local details (like the activity of individual neurons). In the first aim, they are modifying MaGNet to include longitudinal data (recordings collected from the same probes across time with aging). In the second aim, they are modifying MaGNet further to include the individual neural activity patterns recorded from spatially distributed locations (i.e., from probes implanted in two brain regions). In the third aim, they are combining these models to train the AI using all possible data types (age, time, and individual and bulk neuron activity patterns). Through these procedures, they will determine which patterns and features contributed most to improving accuracy in predicting cognition.
The duo’s first funding period has resulted in three manuscripts (either submitted or in preparation). One focuses on the development of a new generative model that will support and inform the longitudinal MaGNet model used in the first aim. The second manuscript in preparation explores how changes in the connections and activities of neurons over time lead to neuronal silencing (when neurons stop communicating), and the functional consequences of this silencing. These relationships are pivotal for developing MaGNet’s ability to integrate these distinct patterns and predict their long-term impacts. The final manuscript focuses on creating a statistical method that works with neuronal networks that have fewer connections. This method will be important because the team expects network connectivity to change over time as neurons become silenced. In the second year of funding, they will continue to refine MaGNet and apply it to additional neuronal data being actively collected.
