Spectral Adaptive Repositioning for Flow-Based Single-Cell Perturbation Modeling (SPEAR)
NeurIPS 2026Transformer-based flow matching for genetic, pharmacological, and cytokine perturbations, including held-out and combinatorial interventions.
About
I am a Ph.D. candidate in Computer Science at Purdue University, advised by Prof. Ananth Grama and Prof. Nadia Atallah Lanman. I am also a trainee member of the Purdue Institute for Cancer Research.
My work spans both general machine learning and computational biology. I am interested in how large models can be trained and evaluated more reliably, how generative models generalize to unseen conditions, and how causal or regulatory structure can make model outputs more interpretable.
I received an M.S. in Computer Science from Purdue University in 2025. Before Purdue, I earned a B.S. in Computer Science (First Class Honours) from City University of Hong Kong and worked with Prof. Shuai Cheng Li on computational genomics.
Research
Large-model post-training & evaluation
Preference and reward modeling, RLHF post-training, supervision quality, and evaluation for large multimodal models.
Single-cell perturbation modeling
Generative models of genetic, drug, and cytokine responses, with emphasis on held-out interventions, combinatorial generalization, and gene-level structure.
Multimodal biological modeling
Generative models connecting transcriptomic states, histopathological images, perturbation-induced morphology, and molecular design.
Causal & regulatory ML
Differentiable causal discovery, gene-regulatory modeling, and causal approaches to therapeutic target prioritization.
Experience
2026–Present
Meta Platforms, Inc.
Research Scientist Intern, FAIR Foundations
Train reward models and contribute to RLHF post-training for large multimodal models, including preference-data construction and evaluation across single-turn and computer-use agent settings.
Improved reward-model training on ambiguous preference data by replacing hard labels with soft targets and using inter-rater agreement for difficulty-aware supervision; also analyzed rubric–human disagreement to improve evaluation quality.
2024
Johnson & Johnson (Janssen Pharmaceuticals)
Data Science Intern, Computational Biology
Built a causal modeling pipeline integrating gene-regulatory priors; prioritized candidate tumor-driver targets in public and internal breast-cancer data.
2021–2026
Purdue University
Graduate Research Assistant
Developed generative, causal, and multimodal ML methods for perturbation response prediction, single-cell/spatial genomics, histopathology, immunology, and cancer biology.
2018–2021
City University of Hong Kong
Research Assistant
Developed algorithms for haplotype-phased 3D genome reconstruction, topologically associating domain (TAD) detection, and single-cell copy-number profiling.
Publications
Published / Accepted
GeneFlow: Translation of Single-cell Gene Expression to Histopathological Images via Rectified Flow
NeurIPS 2025Rectified-flow models translating single-cell transcriptomic states into histopathological images.
Differentiable Constraint-Based Causal Discovery
NeurIPS 2025Differentiable optimization of graphical-separation constraints for constraint-based causal discovery.
Velocity-Weighted Gene Regulatory Modeling Identifies Drivers of Drug-Induced Plasticity in NSCLC
ACM BCB 2026Gene-regulatory modeling of drug-induced cellular plasticity and candidate regulatory drivers.
A Benchmark for Prediction of Transcriptomic Responses to Chemical Perturbations Across Cell Types
NeurIPS D&B 2024 · SpotlightBenchmarking generalization of transcriptomic perturbation-response models across cell types.
Preprints / Under Review
MuonIO: Principled Norm-Aware Descent for Embedding Tables and Language Model Heads
Under review at ICLR 2027Norm-aware optimization for embedding tables and language-model output heads.
Perturbation Modeling via Signed Propagation and Localized Conditioning
Under review at ICLR 2027Attention-free perturbation modeling with signed graph propagation and target-localized conditioning.
Inverse Molecular Design from Paired Control-Treatment Transcriptomic and Morphological Profiles
Briefings in Bioinformatics · under reviewInverse molecular design conditioned on paired control-treatment transcriptomic and morphological profiles.
Generating Joint Transcriptomic and Morphological Responses to Drug Perturbations via Rectified Flow
PLOS Computational Biology · under reviewJoint generative prediction of transcriptomic and morphological responses to drug perturbations.
DeSpotX: Identifiability-Based Decontamination for Spatial Transcriptomics
Under review at ICLR 2027Identifiability-driven modeling of spatial RNA contamination and decontamination.
Honors & Service
Honors: ISMB 2026 Conference Fellowship; Purdue Institute for Cancer Research Graduate Student Travel Award (2026); Purdue Institute for Cancer Research Bioinformatics Research Assistantship Award (2023–2026).
Reviewing: Briefings in Bioinformatics, NeurIPS, ICLR, AAAI, and ICLR workshops (MLGenX, GEM).