- Date of birth: 11/16/1995
- Gender: Female
- Nationality: Chinese
- Hobby: tennis, traveling, reading
- Location: 4056, Basel Switzerland
- Phone: +41772241767
- Email: ziyu.she@unibas.ch
- Ph.D. Candidate at the University of Basel specializing in artificial intelligence and machine learning, with a focus on advancing drug discovery through physics-based deep learning techniques. Demonstrated expertise in optimizing complex systems and fostering collaboration to drive significant research outcomes in medical imaging and robotics.
- English: Advanced (C1)
- Chinese: Native
- Artificial Intelligence
- Machine Learning
- Robotics
- Python
- PyTorch
- Slurm/HPC
- Latex
- https://github.com/SheZiyu?tab=repositories
- https://scholar.google.com/citations?user=V0KN5esAAAAJ&hl=en
- https://www.linkedin.com/in/ziyu-she-868ab0234/
University of Basel
- Leveraged physics-based deep learning for drug discovery by integrating normalizing flows, SMC, and pocket-conditioned generative models with molecular simulation and energy functions to enhance sampling, free-energy estimation, and ligand design.
Polytechnic University of Milan
- Leveraged machine learning to advance medical imaging pattern recognition and optimize robotic controller design.
Sun Yat-sen University
- Engaged in conferences emphasizing Statistics and Mathematics, Ordinary Differential Equations, and Dynamic Systems.
Ph.D.: Computer Science
- Physics-based Deep Learning for Drug Design
MSc: Computer Vision
- Non-rigid 3D Reconstruction, Neural Rendering, Self-supervised Learning
MSc: Applied Mathematics
- Differential Geometry, Differential Manifold, Soliton Equation and Integrable System, Partial Differential Equation
- Outstanding Master Thesis: https://github.com/SheZiyu?tab=repositories
- GPA: 3.31/4
BSc: Applied Statistics
- Advanced Algebra, Analytic Geometry, Probability Theory
- GPA: 3.13/4
-
Harbor Adapters and Harbor-Mix: Infrastructure and a Curated Meta-Dataset for Large-Scale Agentic Evaluation. NeurIPS 2026 Evaluations and Datasets Track (submitted).
-
She, Z., Hinz, F., Lill, M. A. (2026). Local sampler for conformational free energy estimation using normalizing flows (TFEP). (In preparation)
-
Terminal-Bench: Benchmarking Agents on Hard, Realistic Tasks in Command Line Interfaces. ICLR 2026. 2025.
-
She, Z., Marzullo, A., Destito, M., Spadea, M. F., Leone, R., Anzalone, N., et al. (2023). Deep learning-based overall survival prediction model in patients with rare cancer: a case study for primary central nervous system lymphoma. International Journal of Computer Assisted Radiology and Surgery.
-
Su, H., Zhang, J., She, Z., Zhang, X., Fan, K., Zhang, X., Liu, Q., Ferrigno, G., et al. (2022). Incorporating model predictive control and fuzzy approximation for robot manipulation under remote center of motion constraint. Complex & Intelligent Systems.
-
Fu, J., Zhang, J., She, Z., Ovur, S. E., Li, W., Qi, W., Su, H., Ferrigno, G., De Momi, E. (2021). Whole-body spatial teleoperation control of a hexapod robot in unstructured environment. In 2021 6th IEEE International Conference on Advanced Robotics and Mechatronics.
-
STGen3D: Spatial-temporal-based Generator for Dense 3D Reconstructions of Non-rigid Objects. First Author at ICCV 2023 (conference submission). 2022.
- 09/2014 - 06/2021 Five Times of Outstanding Student Scholarship (top 5%)
- 12/2019 Second Class of National College Mathematics Contest (Top 5%)
- 12/2019 Second Class of the National College English Contest (Top 5%)
- 01/2016 First class of Guangdong Automation Experimental Skill Contest (top 1%)
- 04/2015 S Class of the American Mathematical Modeling Contest
-
Developed a transferable normalizing-flow / Boltzmann generator that learns a proposal distribution over ligand conformations in protein binding pockets, using MD trajectories as training data.
-
Combined the flow with sequential Monte Carlo (SMC) + Langevin refinement to efficiently generate near-equilibrium samples and estimate free energies / ΔF between binding modes, while retaining a principled treatment of densities and partition functions.
-
Implemented a full physics-aware data pipeline (heavy-atom projections, pocket masks, bond-graph encodings, multi-mode DCD handling) to make the model robust across different ligands and poses.
-
Built a 3D generative model conditioned on the protein pocket that proposes de novo ligand candidates directly inside the binding site, without requiring a fixed starting molecule.
-
Incorporated geometric and simple physics / energy features of the pocket to bias generation towards chemically valid, sterically compatible, and physically plausible ligands.
-
Positioned as a structure-based drug discovery tool, complementary to docking: the model generates initial ideas in the right pocket, which can then be refined by docking, MD or free-energy methods.
-
Created the ode-solver-rk4 benchmark task to evaluate agent coding ability on implementing a numerical IVP ODE solver (e.g., RK4) under strict step-size / call-order constraints and accuracy requirements.
-
Developed a comprehensive benchmark adapter integrating LLMSR-Bench (a Large Language Model Scientific Symbolic Regression benchmark) into the Harbor evaluation framework. Implemented a faithful port of the LLMSRSearcher agent with vendored dependencies, ensuring algorithmic equivalence between the original benchmark and the Harbor adaptation. Created an end-to-end evaluation pipeline including Docker-containerized test environments, automated metric computation (R², MSE, NMSE), and parity testing infrastructure to validate consistency across 240 scientific equation discovery tasks spanning physics, materials science, chemistry, and biology domains.