The short version
Genomes accumulate duplicate genes. Most copies decay, some are kept, and a few take on new roles — but which ones, and why, is hard to answer at genome scale without a lot of careful pipeline work. My dissertation looks at that question in Populus (poplar trees), using expert-annotated and published case studies as a ground-truth reference set to check whether automated analysis pipelines are finding real biology or artifacts.
Duplication does not explain everything, though. A small residue of genes has no detectable homolog anywhere, and those lineage-specific genes are the second half of the dissertation: how to identify them without fooling yourself, what they look like once you have them, and whether any of them do anything you can measure in a greenhouse or a field.
Below is the detail. If you want the compressed version, it is on my resume.
PhD dissertation research
Tsai Lab, Institute of Bioinformatics, University of Georgia — 2020–present
The dissertation asks where new genes in Populus come from, and it splits into two halves: the duplication landscape, and the genes that duplication cannot explain.
Gene duplication and ohnolog fate
- The gene-duplication landscape of the hybrid poplar genome, using expert-annotated and published case studies as a ground-truth reference set to check whether automated pipelines are finding real biology or artifacts.
- Modeled ohnolog expression divergence with XGBoost and SHAP under leakage-aware validation — chromosome-held-out cross-validation and permuted-label controls, so the reported performance is not an artifact of related sequences leaking between folds.
- Reprocessed large-scale community RNA-seq data (raw FASTQ through differential expression) on reproducible, HPC-scalable pipelines, and built genome-wide protein structure annotation workflows.
Lineage-specific genes in a hybrid poplar genome
Whole-genome duplication, tandem duplication and singletons still leave a residue of genes with no detectable homolog anywhere. Fewer than 3% of gene models in the Populus tremula × P. alba 717-1B4 genome are strictly Populus-specific, and that small set behaves unlike anything else in the genome.
- Built the identification pipeline: phylostratigraphy with GenEra, then quality control that most published LSG sets skip — synteny checks, explicit homology-detection-failure tests, and outgroup checks. The synteny check recovers age calls that homology search alone gets wrong.
- Established the confounder first. LSGs are short, single-exon and GC-poor, so every downstream comparison is length-matched. Skipping this step is how the field manufactures results.
- Found that LSGs appear to piggyback on neighboring genes’ open chromatin: where a new gene lands relative to its neighbor predicts whether it is transcribed at all, while its coding sequence still reads like non-coding DNA.
- Characterized them functionally: high hydropathy but intrinsically disordered, transcriptionally narrow and low, enriched in bark and xylem, and stress inducible. Structural homology that BLAST cannot see is possible but rare — one LSG folds onto a real plant protein family at 13.7% sequence identity.
- Took it from ~65,000 gene models down to six testable constructs, with COLD-REGULATED 15A as a positive control, then into greenhouse drought trials and a field nursery planting. The LSG lines did not outperform COR, and the transformation effect itself is not negligible — both worth saying out loud.
Foundation models, and where the two halves meet
- Evaluated seven publicly available DNA foundation models, spanning several architectures and tokenization schemes, on Populus promoter sequences — diagnosing embedding quality and establishing architecture-specific pooling. This is what pulled me toward the foundation-model robustness work.
- Genes with no homologs are a natural test case for when a genomic foundation model hallucinates: if the model reports structure in a sequence that has no relatives, you need a way to tell signal from confabulation. That question is the bridge to the codon language model work.
Also
- Contributed to Populus VariantDB v3.2, a CRISPR gRNA verification tool, and to a coexpression analysis of a perennial-specific sulfate transporter subgroup associated with lignification. Both published in Tree Physiology (2025).
Tools: Python, R, Snakemake, SLURM/HPC, STAR, HISAT2, DESeq2, edgeR, GenEra, Foldseek, AlphaFold/ESMFold structures, plotnine.
MSc research
Plant Biotechnology and Molecular Biology Lab, Department of Horticulture, National Taiwan University — 2017–2019
- Comparative transcriptome analysis of mycorrhiza-enhanced salt tolerance in rice, published in Frontiers in Plant Science (2022) as co-first author.
- Automated high-resolution melting (HRM) analysis output processing with custom scripts, which streamlined genotyping for a peach (Prunus persica) chilling requirement study (IJMS, 2020).
- Designed RNA-seq experiments and identified the genome-wide MADS-box gene family to study gibberellin-induced sex determination in bitter gourd. Presented at the International Horticultural Congress in Istanbul (2018); first prize at the NTU graduate student research poster competition (2019).
Recent talk slides
Lineage-specific genes in a hybrid poplar genome — the dissertation half above, from pipeline through field trial. Institute of Bioinformatics, September 2026. Slides are as-presented, so the work-in-progress labels on them are real. Slides (PDF)
I also gave a talk that month on the genomic AI work. Those slides are not posted while the related manuscript is under review; email me if you would like them.
Full citations are on the publications page. Talks and posters are listed there too.