PhD Student · IGP AI-for-X, NTU Singapore
I use AI and data to ask what's conserved, what varies, and what that tells us. I do this in proteins, in the scientific literature, and in everyday tools people actually use.
01 — About
I'm a PhD student in the Interdisciplinary Graduate Programme (IGP), AI-for-X track at Nanyang Technological University, working in Prof Dan He's lab. Before this, I completed a B.Sc. in Chemistry and Biological Chemistry (Honours, Distinction) at NTU and spent a year as a Research Assistant in the same group.
My research has moved across the bioinformatics spectrum. I started with epigenomic marks, then questioned the fidelity of transcriptomic platforms, and most recently built protein language models that reason in motifs rather than individual amino acids. That work became a paper at KDD 2026. Lately I've also been working on information retrieval: how scientific papers cite each other, and what that context can tell machines.
Outside the lab, I co-founded Kayabread, where I point the same AI and data skills at concrete, everyday problems. It keeps me honest about whether a model actually helps someone.
Along the way: a Silver at the ST Engineering Hackathon, and a handful of side projects done mostly to procrastinate explore areas outside my usual domain.
02 — Latest
03 — Selected work
Can a protein language model built on motifs be more interpretable, and more correct?
Most protein language models read one amino acid at a time. This model uses conserved functional motifs as its vocabulary instead, which is closer to how evolution encodes function. The result is protein generation whose building blocks you can actually inspect.
When a paper cites another, what exactly is it saying about it?
A 367K-paper dataset built from PLOS JATS XML. Every in-text citation is resolved to its sentence and linked to the cited DOI, for citation-context classification and retrieval.
Are conserved motifs useful atomic units for protein design?
I led the computational team. We built a motif-guided Cas13d pipeline with a motif-based subtype classifier and a fine-tuned ESM-2 generator. The generated variants showed knockdown efficacy comparable to wildtype in the wet lab.
What do we lose when we aggregate single cells into a bulk proxy?
A simple PCA can tell pseudo-bulk and bulk RNA-seq apart. That points to systematic differences, which matters for comparing findings across platforms.
How different are epigenomic marks across cell types?
Many cell types share accessible sites. So instead of asking whether a site is accessible, I classified genes by how much their H3K4me3 and DNase marks overlap across cell types. That separates cell-identity genes from generic ones.
04 — Beyond the lab
Research taught me how to ask good questions of data. With Kayabread I wanted to use the same AI and data science skills on something more concrete: tools that save real people real time, every day.
Kayabread is a small software studio in Singapore that I run with Kaden. We build focused apps for the repetitive parts of working for yourself. We design each workflow by hand around how people actually work, and we use AI only where it clearly saves effort.
Visit kayabread.com →Bookkeeping for self-employed people and freelancers, designed to take about ten seconds a day.
A place to keep count of things. More details coming soon.
05 — Publications
Proceedings of the 32nd ACM SIGKDD Conference on Knowledge Discovery and Data Mining (KDD '26), Jeju, Korea
DOI →Hugging Face Datasets · CC BY 4.0
DOI →Journal of Bioinformatics and Computational Biology, World Scientific
DOI →bioRxiv
Preprint →06 — Journey
Most recent first. Each stop started with a question.
IGP, AI-for-X · NTU · He Lab, Prof Dan He
What's the next question worth chasing?
Still deciding, and enjoying it. Somewhere at the intersection of AI, biology and how scientific knowledge is organised.
School of CCEB, NTU · Prof Dan He
Can a protein language model built on motifs be more interpretable, and more correct?
Building on the iGEM work, I developed a motif-level language model that uses functional protein motifs as its vocabulary instead of individual residues. I focused on metalloproteins, where tight constraints around metal-binding sites make motif conservation especially pronounced and testable. The work was published at KDD 2026.
School of CCEB, NTU · Prof Meng How Tan
Are conserved motifs useful for de novo protein design?
Evolution doesn't tinker at random. It conserves structural and functional motifs over billions of years. We generated de novo Cas13 variants by composing these motifs instead of generating residues one by one, then validated them in the wet lab. Their knockdown efficacy was comparable to wildtype, which suggests the motifs carry real biological meaning.
Computational Biology & Omics Lab, Bioinformatics Institute, A*STAR · Prof Kumar Selvarajoo
Are pseudo-bulk profiles truly equivalent to bulk RNA-seq? What do we lose in that assumption?
Pseudo-bulk RNA-seq, where single-cell counts are aggregated per sample, is increasingly used as a stand-in for bulk data. But a simple PCA can reliably tell the two modalities apart, so aggregation introduces systematic differences. That raises uncomfortable questions about how comparable findings are across platforms.
Computational Biology Group, Nanjing University · Prof Dijun Chen
How much transcriptional noise is too much, and how would we even know?
Single-cell data is sparse and noisy. I explored whether variational autoencoders could denoise these profiles without erasing genuine biological signal, and how to evaluate what gets recovered. The harder question, still unanswered: when does "denoising" become distortion?
Centre for Biomedical Informatics, Lee Kong Chian School of Medicine, NTU · Prof Bernett Lee
How different are epigenomic marks across cell types?
Counterintuitively, many cell types share a large number of accessible sites. Instead of asking whether a site has a mark, I classified genes by how much their marks overlap across cell types. That separated genes that signal cell identity from generic ones. The work led to a first-author journal paper and an oral presentation at GIW ISCB-Asia 2023.
07 — Contact
I'm open to research collaborations, product ideas, or a good conversation about biology, data and building things.
boonyhow@gmail.com