Fine-tuning LLM hyperparameters to align semantic and physiological contexts of aging-related pathways.

Abstract

Biological aging is defined by many physiological pathways, calling for the identification and validation of synergistic combinations of interventions that would address multiple hallmarks of aging. Multitude of biological agents resides semantically linked in the LLM's vector space, offering avenues to leverage Artificial Intelligence's (AI) generative capabilities to explore remote and indirect connections and relationships. This study examines various models, hyperparameter configurations, response formats, scopes, and other AI generation scenarios to align semantic proximity of aging-related factors with their biological context, with LLM precision being evaluated against KEGG benchmark database. Tuning GPT-4 hyperparameters such as temperature and frequency penalty offers maximum diversity of context collected for known senotherapeutics and pathway regulators. Response confidence score, based on LLM's internal logprob metric, was found to be predictive of KEGG validation success and offers evidence of higher semantic proximity translating into lower biological distance between input molecules and their respective associations.