Research intern building rigorous evaluation and data infrastructure for AI-driven biology — computational proteomics, perturbation prediction, protein design
I am a Research Intern in Tiannan Guo's group at Westlake University, with an MSc in Bioinformatics and Systems Biology. My current work involves mass-spectrometry data, machine learning, statistical evaluation, and research software. I am especially interested in developing computational tools that are useful, testable, auditable, and reproducible in real scientific workflows — most recently, bringing null controls, split-reference estimation, and target–decoy FDR (the rigor standard of proteomics) to the evaluation of single-cell and protein-design models.
flowchart LR
A[Instrument data<br/>DIA · timsTOF · ion mobility] --> B[Data infrastructure<br/>formats · HPC · reproducibility]
B --> C[Computational models<br/>Transformers · representation learning]
C --> D[Evaluation<br/>FDR · entrapment · ablations]
D --> E[Usable scientific systems<br/>software · agents · audit trails]
- Computational mass spectrometry: DIA/DDA peptide identification, timsTOF and ion-mobility evidence, spectral representations, target–decoy FDR, and entrapment-based evaluation.
- Scientific AI agents: tool use, knowledge graphs, structured memory, provenance, failure recovery, human oversight, and careful validation.
- Research software and infrastructure: cross-vendor data processing, scalable HPC workflows, automated testing, cross-platform packaging, and reproducible benchmarks.
- Structural and experimental background: NMR, molecular dynamics, protein biochemistry, and mammalian cell experiments.
| Project | Summary |
|---|---|
| pert-null | Null controls and FDR-aware evaluation for single-cell perturbation prediction: split-reference delta metrics that remove shared-reference bias, permutation nulls with pseudocount p-values, and FDR-controlled DE-set recovery. |
| design-decoy | Target–decoy empirical FDR for in silico protein design funnels: decoy sequence/pair generators plus calibrated thresholds for scRMSD, pLDDT, and ipTM designability calls. |
| scp2anndata | Single-cell mass spectrometry quantification (DIA-NN pg_matrix, long tables) into AnnData/MuData with MS-aware QC — a data bridge between SCP and the scverse ecosystem. |
| falsify | Falsification-first evaluation for scientific AI agents as executable code: pre-registered claim contracts, hidden tests, decoy tasks that measure overclaiming, and probability calibration. |
| ISDIA v2.0 | Software for detecting DIA acquisition modes across mzML, mzXML, Bruker timsTOF, Thermo RAW, and SCIEX WIFF data; includes CLI/GUI deployment and cross-platform support. |
| psm-fdr-audit | A small Python package for checking target–decoy q-value calculations with tied scores, pseudocounts, synthetic truth labels, and simple failure scenarios. |
| Falsification-First Scientific Agent Protocol | A platform-neutral framework for claim contracts, hidden validation, decoy controls, calibrated scoring, and independent evidence in scientific-agent evaluation. |
- AlphaRaw #138 — submitted a pandas compatibility fix and regression test for precursor-column filling.
- spectrum_utils #90 — submitted a fix that preserves numerical mass-error precision while keeping annotation strings concise.
Both pull requests are currently under maintainer review.
- DIA-BERT for timsTOF — contributing to an extension of Transformer-based peptide identification using ion-mobility evidence, multi-run data pipelines, multi-GPU training, controlled perturbations, ablations, and FDR-based model selection.
- JAVICS — working on a knowledge-graph-driven mass-spectrometry QC agent with diagnostic tool use, persistent workflow state, feedback, recovery, and auditable reports.
- MSIT — contributing to data infrastructure across Orbitrap, TripleTOF, and timsTOF platforms, including scalable conversion and validation on Slurm clusters.
- AIPC evaluation — participating in the reproduction and technical evaluation of peptide-scoring systems using target–decoy FDR, entrapment-based FDP, integrity checks, and benchmark design.
These are ongoing research projects. Public code, datasets, and quantitative results will be linked when their release status permits.
- Research Intern, Tiannan Guo Group, Westlake University (2025–present)
- MSc, Bioinformatics and Systems Biology, Vrije Universiteit Amsterdam & University of Amsterdam (2023–2025)
- Master's Research Intern, Oliver Zerbe Group, University of Zurich (2024–2025)
- BSc, Pharmacy / Biochemistry and Pharmacology, China Pharmaceutical University & University of Strathclyde (2018–2023)
My earlier research experience includes NMR structure analysis, pressure-dependent molecular dynamics, protein modelling, antibody expression, and cell-based functional assays. This background helps me understand the experimental questions and limitations behind computational workflows.
Python · PyTorch · NumPy · pandas · Polars · scikit-learn · anndata / scanpy · R · Linux · Git · Slurm · multi-GPU training
DIA-MS · timsTOF / diaPASEF · ion mobility · mzML / mzXML · Bruker TDF · target–decoy FDR · entrapment FDP · single-cell perturbation prediction · protein design evaluation
knowledge graphs · scientific agents · local RAG · audit trails · hidden validation · null controls · benchmark design · reproducible pipelines