Hamilton High junior tells Science and Technology committee her AI model flagged saliva biomarkers for depression

Ananya Lakerajah, a 17‑year‑old junior at Hamilton High School, presented her research on using metabolomics and machine learning to identify potential salivary biomarkers of major depressive disorder to the Committee on Science and Technology. She said she analyzed a UCSD‑formatted dataset of 261 saliva samples and used an ensemble model and feature‑selection techniques to classify patients as depressed or not depressed.

Lakerajah framed the work in public‑health terms, saying major depressive disorder is a serious clinical disease: "It reduces one's lifespan by 7.9 years and it's the leading cause of suicide accounting for 60 percent of the 770000 lives lost every single year," she told the committee. She described metabolomics as the study of small molecules in the body and said the project aims to add objective biological signals to currently subjective diagnostic procedures.

Why it matters: If validated, objective biomarkers could complement questionnaires and clinical evaluation, potentially improving diagnosis and guiding drug discovery. Lakerajah said she also used pathway mapping and molecular docking to propose drug targets, and she described an app to combine metabolite results with self‑reported mood data so clinicians have broader context.

Methods and claims: Lakerajah explained she used statistical tests (Spearman rank correlation, Mann–Whitney U, PLS‑DA), interpretability tools (SHAP, permutation importance) and an ensemble of machine‑learning models to produce a consensus list of candidate metabolites. She reported that the model achieved about 90% accuracy and a 97% area under the curve (AUC) in distinguishing depressed from non‑depressed samples; the committee did not independently verify those metrics. She cautioned that her work is computational and that biological validation — testing metabolites in physical samples — is a needed next step.

On molecular follow‑up, Lakerajah said pathway analysis linked some candidate metabolites to mitochondrial function and to pathways previously discussed in neurodegenerative disease research. Using UCSF Chimera for in‑silico docking, she said flubenazine showed the lowest binding energy against a selected target protein in her simulations and described the inhibition concept as a "lock‑and‑key" mechanism.

Committee reaction and questions: Members asked technical clarifications about charts and ranking conventions, whether identified metabolites were endogenous or exogenous, and how confounders such as caffeine or alcohol might affect results. A committee member noted that several top features appeared to be exogenous compounds (drugs) and therefore unlikely causal biomarkers; Lakerajah agreed such items require careful interpretation. The committee's substitute chairman praised the presentation and said AI policy would be an expanding topic for the committee.

Limitations and next steps: Lakerajah acknowledged limits of the current work: she did not collect saliva samples herself and relied on an existing UCSD‑formatted dataset (details of the dataset source and patient metadata were not specified in the presentation). She recommended expanding sample size and adding biological testing to validate computational findings before clinical use. She also urged support for AI regulation to ensure models are deployed safely in medical settings.

The committee did not take any formal votes or actions on the presentation and adjourned after members thanked the presenter.