Oct 2025
Predicting substrate specificity of acyltransferase domains in polyketide synthases using machine learning
BIOINFO 2025 · Poster
Joon Young Kwon
I design protein sequences toward an objective. Given a property you want — a substrate an enzyme should accept, a rate you need raised — I train a model that scores a sequence against that objective, then search sequence space for the variants worth testing.
Ph.D. student in Chemical and Biomolecular Engineering at KAIST, in the Systems Biology and Medicine Laboratory, advised by Prof. Hyun Uk Kim.
protein language models
fitness prediction
directed evolution
sequence design
representation learning
Objective
Turning the property you want into a number
A design task is only as good as the thing it optimizes. I train models that take a sequence and return the objective itself — which substrate an enzyme will accept, how fast it will turn it over — with the inputs kept faithful to what the experiment actually measured.
Slides add link
Search
Evolving a sequence toward that objective
With a scoring model in hand, the question becomes where to look. I run directed evolution in silico — guided search over mutations, starting from a template — including the harder case where two objectives have to improve at once.
Slides add link
Representation
What the model needs to know about a sequence
Whether a pretrained protein language model already encodes what a given objective depends on, or whether an embedding learned for the task finds good variants in fewer rounds. This is the question underneath both of the above.
Slides add link
De novo
Starting from no sequence at all
A side interest rather than a main line: designing a binder against a given target from scratch, as an entry to an open protein design competition where the designs are expressed and measured.
Slides add link
← Home
Publications
Published
AI and mechanistic modeling for characterizing biosynthetic pathways of natural products
Natural Product Reports 43, 787–802 (2026)
† These authors contributed equally.
Preprints
Toward understanding drivers of specialized metabolism in Actinomycetota: insights from 1432 transcriptomics datasets for 132 strains
bioRxiv, posted 17 June 2026
Patents
Methods for predicting and recommending mutations to redesign a substrate specificity of acyltransferases in polyketide synthase
Korean patent application P2026-0330-KR01, filed 2 October 2026 · Applicant: KAIST
Sep 2025
Genome-scale metabolic model guided optimization of Streptomyces venezuelae metabolism for enhanced pikromycin production
KSBB 2025 Fall Meeting & International Symposium (KSBB–AFOB) · Poster
Jun 2025
Systems metabolic engineering of Streptomyces venezuelae for the enhanced production of pikromycin
20th International Symposium on the Biology of Actinomycetes (ISBA) · Egmond aan Zee, Netherlands · Poster
Apr 2025
A deep learning model for predicting expression of biosynthetic gene clusters in Streptomyces
KSBB 2025 Spring Meeting & International Symposium · Poster
Photos go here.
Education
2026–present
Ph.D., Chemical and Biomolecular Engineering
KAIST, Daejeon, Republic of Korea
Advisor: Prof. Hyun Uk Kim, Systems Biology and Medicine Laboratory.
2024–2026
M.S., Chemical and Biomolecular Engineering
KAIST, Daejeon, Republic of Korea
Advisor: Prof. Hyun Uk Kim, Systems Biology and Medicine Laboratory.
Thesis: Machine learning-based prediction and design of enzyme substrate specificity
and kinetic activity.
2020–2024
B.S., Chemical and Biomolecular Engineering
KAIST, Daejeon, Republic of Korea
Teaching
2026
Teaching assistant, Undergraduate Research Program
KAIST
Project: Development and application of an integrated data-based enzyme kinetic parameter prediction framework.
Spring 2026
Counseling assistant (CA)
Academic Counseling Center, KAIST
Fall 2025
Teaching assistant, Introduction to Numerical Methods for Chemical and Biomolecular Engineers (CBE206)
KAIST · Instructor: Prof. Hyun Uk Kim
Spring 2025
Counseling assistant (CA)
Academic Counseling Center, KAIST
Skills
Objective modeling
protein language model fine-tuning
multimodal encoding
mutation-aware architectures
out-of-distribution benchmarking
Sequence search
guided beam search
in silico directed evolution
multi-objective optimization
structure prediction for validation
Data
kinetic and biosynthetic data curation
similarity-based splits
multi-omics integration
Tools
Python
PyTorch
scikit-learn
Email
Office
Systems Biology and Medicine Laboratory
Department of Chemical and Biomolecular Engineering, KAIST
291 Daehak-ro, Yuseong-gu, Daejeon 34141, Republic of Korea
Elsewhere
Google Scholar
ORCID
LinkedIn