Kelton, WilliamHolmes, GeoffreyZhou, Kang2026-07-162026-07-162026https://hdl.handle.net/10289/18458Accurate nucleic acid sequencing is essential for analysing protein variant libraries composed of closely related sequences. Oxford Nanopore sequencing offers long reads, high depth, and comparatively lower cost, but its single-read error rate limits its routine use for variant calling, where sequencing errors can be mistaken for true mutations. This thesis investigates whether domain-specific fine-tuning of a nanopore basecalling model can improve single-read variant- calling performance in an antibody library context, using a dual-site IgA Fc deep mutational scanning library as an example. A series of nanopore basecalling models was fine-tuned using in-house wild-type IgA Fc nanopore data collected under different experimental sample-prep procedures. These models were evaluated on a held-out, variant-rich nanopore dataset against an orthogonal MGI-derived short-read reference dataset. The benchmarking framework assessed both read-level quality and population-level recovery, focusing on library-constrained variant detection and variant-abundance concordance. Fine-tuned models consistently outperformed the pretrained baseline, demonstrating that nanopore signals contain learnable domain-specific structure that can improve single-read variant calling. The best model, Alpha, achieved the best observed balance of low primary error rate, high read retention, low framework noise, and strong abundance-aware concordance with the ground truth. The results showed that the experimental procedure used to collect training data had a greater influence on the fine-tuning outcome than either training data volume or the stringency of quality control. Lower validation loss observed during training did not reliably predict better downstream performance, highlighting the need to evaluate basecalling models using biologically meaningful metrics rather than training diagnostics alone. Model-weight analysis further showed that successful fine-tuning involved small, structured recalibration of the signal-interface and decoding components rather than broad model-level drift. Overall, this thesis establishes a practical framework for fine-tuning and evaluating domain-specific nanopore basecallers for variant libraries. It shows that domain-specific fine-tuning can materially improve the recovery of variant populations from nanopore signals and supports future work on computational augmentation of WT IgA Fc nanopore signals, architecture-aware fine-tuning, UMI-linked molecule-level benchmarking, experimental validation, predictive modelling, and extension to longer antibody constructs.enAll items in Research Commons are provided for private study and research purposes and are protected by copyright with all rights reserved unless otherwise indicated.oxford nanopore sequencingnanopore basecallingdomain-specific fine-tuningvariant callingdeep mutational scanningantibody librariesIgA Fcprotein engineeringbioinformaticsmachine learninglong-read sequencingsingle-read variant callingThesis