GeneLM · Evo2
AI for Science · Genomics
A zero-shot genomic variant scoring workflow built around the Evo2 biological foundation model.
Long-context likelihood scoring on H100, evaluated against ClinVar across reference genomes.
Research StudentOsaka University
Yichen Zhang
Generative models for molecules, genomes, and scientific discovery.
Open to doctoral positions and research collaborations in generative AI for the life sciences.
- Based in
- Osaka, Japan
- Focus
- Generative AI for Science
- Previously
- NYU · M.S. Computer Engineering
About
I am a research student at Osaka University preparing for doctoral work in generative AI for drug discovery. I build generative and foundation models for molecules and biological sequences under real scientific constraints.
My path into AI began in finance and continued through computer engineering at NYU, where I worked on multimodal systems and genomic foundation models. It shaped the question I pursue now: how can generative intelligence make scientific discovery faster and more reliable?
Research Interests
Three connected directions — from learning biological structure to generating candidates that hold up under evaluation.
Generative AI for Drug Discovery
Controllable molecular generation, scaffold-aware design, and evaluation that measures whether a candidate is actually useful.
Biological Foundation Models
Long-context models for genomic sequences, variant effect analysis, and representations that transfer across biological tasks.
Multimodal Intelligence
Vision-language systems that turn complex inputs into reliable, structured, machine-readable output.
Selected Projects
Research through building — across AI for science, multimodal learning, and compute-aware reasoning.
GitHub
Schema-Constrained VLM
Multimodal AI · Structured Generation
Fine-tuning a compact vision-language model to produce deterministic, schema-valid food understanding.
1,500 instruction samples, parameter-efficient SFT, and a reproducible validation pipeline.
Tiny Recursive Models
Reasoning · Efficiency
A compute-aware study of recursive reasoning and adaptive computation on Sudoku-Extreme.
Characterised an overthinking regime and identified a stronger shallow fixed-budget configuration.
IL-10 Optimal Control
Control Systems · Immunology
A continuous-time model of immune modulation under physiological noise.
Designed and evaluated a Kalman filter with an LQR controller in MATLAB and Simulink.
Blog
Research notes on models, biological data, and the systems behind them.
The notebook is ready. Published research notes will appear here soon.
Education
An interdisciplinary route into AI research.
- 2026 —Osaka, Japan
Osaka University
Research Student · Pre-doctoral
- 2024 — 26New York, USA
New York University
M.S. in Computer Engineering
- 2021 — 23Beijing, China
Beijing Jiaotong University
B.S. in Communication Engineering
- 2017 — 21Zhuhai, China
Beijing Normal University, Zhuhai
B.S. in Finance
Toolkit
Modelling
Scientific
Systems
Contact
I am glad to talk about doctoral opportunities, research collaborations, or anything at the intersection of generative AI and the life sciences.
Write to mez.yichen@outlook.com