Thermodynamic Fingerprinting of DNA: From Molecular Barcoding to AI-driven Sequencing

Abstract

DNA function is tightly coupled to DNA thermodynamics. Duplex and secondary-structure formation emerge from a balance between stabilizing base pairing and base stacking interactions and the entropic cost of constraining flexible polymer chains. This balance governs core biological processes and underlies many widely used biotechnologies, including PCR primer and probe design, in situ hybridization, genome editing guide optimization, and DNA nanostructure engineering. Thermodynamic behavior is also directly observable in thermal denaturation experiments, where sequence-dependent melting curves provide an accessible and information-rich signal that is increasingly used in molecular diagnostics through high-resolution melting assays. Motivated by both the physical richness of DNA melting and its analytical utility, this thesis investigates DNA melting beyond simple two-state descriptions and develops melting-based strategies for molecular identification. In Chapter 2, we study intermediate state that arise during thermal denaturation. We show that the intermediate-regime features produce reproducible, sequence-dependent signatures that can be harnessed as thermodynamic fingerprints. Building on this principle, we engineer dual-domain DNA constructs with intentionally separated stability domains, producing well-resolved multi-step melting transitions that function as thermal barcodes. We further translate these barcodes into a general tagging strategy for molecular identification by coupling barcode transitions with target-dependent hybridization signals, and we extend the approach to analytes without intrinsic melting transitions, such as proteins, by using affinity-based labeling to confer externally readable thermal barcodes for multiplexed detection. In Chapter 3, we address the inverse thermodynamic problem: inferring sequence information from a melting profile. Because many distinct sequences can exhibit similar melting behavior, the mapping from curve shape to sequence is intrinsically ambiguous. We introduce a machine learning framework that recasts inference as a retrieval problem using a dual-encoder architecture, embedding melting curves and candidate sequences into a shared representation space. By learning subtle, distributed features in curve shape associated with nearest-neighbor stacking interactions and motif context, this approach supports identification of the most thermodynamically consistent sequence within a candidate library and provides a foundation for melting-based, reference-driven resequencing workflows. Together, these contributions expand DNA melting from a stability readout into a programmable and data-driven platform for molecular identification, linking sequence to an experimentally grounded thermodynamic phenotype.