Research Interests
What I am working toward
These are the areas I am actively working in and building foundations for as an
undergraduate student at Northwestern Polytechnical University (NPU/NWPU) in Xi'an,
China. They are interests and directions, not a record of completed research.
AI for Biotechnology (AI4Bio)
Using machine learning and deep learning on biological research problems — from data
preparation through to modelling and evaluation. I use AI4Bio as an informal
umbrella term for this, rather than a formally standardised designation.
My interest here is deliberately bottom-up. Before a model is trained, the data itself has
to be trustworthy: its provenance, its column definitions, its units, and its exclusions.
Model performance is bounded by the data underneath it, so that is where I am starting.
Bioinformatics and Computational Biology
Representing, retrieving, and processing biological data programmatically: sequences and
annotations, standard file formats, structural data, and the tooling that connects them. I
work primarily in Python with Biopython, pandas, and NumPy.
A particular interest is the seam between biological data standards and what machine
learning pipelines actually need — the point where a format's guarantees stop and a
dataset's assumptions begin.
Protein Structure Prediction
How structural information is predicted and then represented as machine-readable features.
My current work uses AlphaFold models, both from the AlphaFold Database and single
predictions via the AlphaFold Server, as the source of structural data.
This includes the confidence measures that come attached to predicted structures:
-
pLDDT — per-residue confidence in the local structure. Read per region
rather than as a global average, since an average hides disordered or uncertain segments.
-
PAE — an N×N matrix estimating the expected position error of one
residue if the structure is aligned on another; useful for judging relative placement and
domain packing.
-
pTM — a global confidence measure over the predicted fold.
Reliability and Confidence Evaluation
Confidence scores like pLDDT and PAE are the de-facto indicators that practitioners read as
"model quality". The question I am building toward is the uncomfortable one underneath
them: how far can those indicators actually be trusted, and where do they fail?
A high-confidence prediction can still be wrong — pLDDT and PAE measure a model's internal
confidence, not biological truth. Verifying a structure requires real experiments:
X-ray crystallography, NMR, cryo-EM. The gap between a confident prediction and a measured
structure is where I think the most useful questions live.
Status of this direction. This is reading, preparation, and foundations
work. I have not completed or published a reliability audit, and I make no claim of
findings here.
Reproducibility and data integrity
Not a separate research area so much as a requirement running through the others. In
practice, for me that means checksummed and frozen inputs, integrity tests that recompute
values from raw files, machine-readable schemas holding every column definition and unit,
and exclusions recorded one by one with reasons rather than applied silently.
The test is simple: if someone cloned the repository, could they reproduce the numbers and
understand why each one is what it is?