I'm an Honors Biomedical Sciences graduate from the University of
Waterloo, now a self-driven computational biologist and ML engineer in
Toronto. I build systems bridging ML and biomedical science, addressing
research problems with creative domain integration, feature engineering
and machine learning.
Outside of programming, I enjoy combat sports (Muay Thai & Boxing), sprinting,
and activities that boost overall physical well-being. I speak English, French,
and Cantonese.
Currently open to entry-level roles in bioinformatics, computational biology
or ML Engineering.
* inverse radial sequencing-based protein design *
A dual cross-attention transformer system for high-affinity protein binding
pocket design around any given chemical compound's scaffold.
NanoMaker applies transformer logic towards the 3D biochemical protein space.
Through Inverse Radial Sequencing, NanoMaker translates 3D protein structures
into clear, sequentially-organized spatial data. This turns the challenge
of de novo (from scratch) protein pocket design into a sequential vector generation
task.
Binding pocket design is elegantly performed by two cross-attention transformer
models conditioned on drug scaffold molecular fingerprints:
** Skeleton: generates the 3D spatial arrangement of blank alpha-carbon coordinates,
the "skeleton"
** NAAnoBot: sequentially populates each slot by curating specific biochemical
environment features and aligning them + selecting an amino acid based on
its hand-curated feature vector
Trained on 7M sequence windows from high-affinity protein-drug complexes (BindingDB
+ PDB, ≤0.1nM ), with a zero-shot train/test split across drug scaffolds.
damage decoder, repair proposer
------------------------------------
A full-stack platform for mutation screening and therapeutic candidate
identification.
Present PRISMATIC with the chromosome number and sequence strings (mutation
site + context) of any mutation, and it will predict the pct. chance of
pathogenicity and benignity. Functionality expands to therapeutic repair
candidate identification, powered by ReGen, a custom guided-mutation algorithm
designed to reduce pathogenicity and locate + break past benignity maxima.
Powered by a 500+ feature extraction pipeline spanning research-backed
DNA biochemistry, predicted protein properties, genomic domain shift detection
and motif tracking from DNA synthesis to post-protein production.
Deployed model metrics below.
BlueTuna creatively addresses the task of hyperparameter
optimization by breaking it down into 3 stages, each informing the
next:
[1]: Parameter search space filtering via **gradient-based region scoring**,
isolates regions most likely to contain performance optima
[2]: **Custom perceptron training** on Latin Hypercube-sampled data from filtered
search space to better "see" the hyperparameter landscape
[3]: **Fixed-weight gradient descent** on hyperparameter values, guiding model
performance towards optima
BlueTuna was able to achieve a competitive performance ceiling against
Optuna (top-performing optimizer), beating it on 5/20 trials and competitive
on 40%. Notably, its median performance slightly exceeded Optuna's, demonstrating
its limitation to be consistency rather than performance ceiling.
exploring the dark chemical space
A drug discovery framework designed for drug-structure screening and lead
optimization against user-defined biological targets. Paired with interactive
chemical space analysis and various chemical-similarity calculations,
aimed at accelerating drug candidate identification and optimization.
Features "MutaGen", a molecular optimization algorithm that leverages
a curated molecular fragment library to intelligently modify drug molecules,
using ML models to predict new molecule potencies.
Employs plateau-breaking strategies across 3 possible molecular changes:
adding fragments, replacing atoms and removal. Automatically filters out
molecules that violate drug-likeness rules.