Evaluating the Usefulness of a Large Language Model as a Wholesome Tool for De Novo Polymerase Chain Reaction (PCR) Primer Design

This study aimed to assess the ability of language learning models (LLMs), specifically GPT-3.5 (Chat Generative Pre-trained Transformer 3.5) and GPT-4 (Chat Generative Pre-trained Transformer 3.5), in designing primers for diagnostic polymerase chain reaction (PCR) of the monkeypox virus (MPXV). Five primer pairs were generated by each LLM, and their thermodynamic properties and specificity were analysed post-hoc using commonly used software. The LLMs demonstrated ability in sequence generation and predicting melting temperatures (Tm), but their accuracy in predicting GC content was suboptimal, necessitating further investigation. Results indicated that, of the total primer pairs, only three designed by GPT-4 and two by GPT-3.5 could theoretically form a PCR product, but only one pair demonstrated suitable parameters for experimental validation. This preliminary exploration suggests that while LLMs have a potential in aiding primer design, their accuracy needs improvement to match current deterministic, rule-based tools used in the field. Consequently, manual intervention remains a crucial step in PCR primer design.