Zelin Zang

Zelin Zang

Postdoctoral Researcher at HK Institute of Innovation, CAS. Focusing on AI for Science, Medical Foundation Models, and Deep Manifold Learning.

Posts by Collection

portfolio

publications

Neural network diffusion

Kai Wang, Dongwen Tang, Boya Zeng, Yida Yin, Zhaopan Xu, Yukun Zhou, Zelin Zang, Trevor Darrell, Zhuang Liu, Yang You, arXiv preprint arXiv:2402.13144, 2024

talks

Deep Manifold Transformation for Explainable Visualization (DMTEV)

Published:

Presented cutting-edge research on deep manifold learning methods for high-dimensional data visualization. This talk covered the DMTEV framework, which combines topology-preserving dimensionality reduction with explainable AI techniques. Discussed applications in single-cell genomics, medical imaging, and complex data analysis, demonstrating how deep manifold transformation enables interpretable visualization of complex biological and clinical datasets while preserving intrinsic data structures.

FINE Review: AI for Science Research Progress and Applications

Published:

Comprehensive review presentation on AI for Science research progress, covering interdisciplinary applications of deep learning in biology, medicine, and physical sciences. Highlighted achievements in protein structure prediction, drug discovery, molecular dynamics simulation, and single-cell analysis. Discussed the integration of foundation models, geometric deep learning, and physics-informed neural networks for advancing scientific discovery. Presented future directions including AI-driven virtual cell construction and multi-omics data integration.

Diffusion-Based Data Augmentation for Unsupervised Contrastive Learning

Published:

Detailed research presentation on novel diffusion-based data augmentation methods for unsupervised contrastive learning. Introduced the DiffAug framework that leverages diffusion models to generate semantically consistent augmented views from scratch, addressing the limitations of traditional augmentation techniques. Demonstrated superior performance on image representation learning tasks, showing significant improvements in downstream classification, detection, and segmentation tasks. This work was later published at ICML 2024.

High-Performance Computing for AI: Scalable Deep Learning Systems

Published:

Invited lecture at NUS HPC-AI Lab during visiting scholar period under Prof. Yang You’s supervision. Presented research on scalable deep learning systems, distributed training optimization, and efficient model parallelism. Covered advanced topics including gradient compression, mixed-precision training, and memory-efficient attention mechanisms. Discussed applications of HPC-AI in large-scale biological data analysis, particularly for single-cell genomics and protein structure prediction. Explored future directions in federated learning and edge computing for life sciences.

Geometric Deep Learning for Drug Discovery and Single-Cell Omics

Published:

Comprehensive presentation on geometric deep learning applications in drug discovery and single-cell data analysis. Discussed graph neural networks for molecular property prediction, protein-ligand binding affinity estimation, and structure-based drug design. Presented novel approaches for learning on cellular graph representations, including cell-cell interaction networks and spatial transcriptomics data. Highlighted the integration of multi-modal biological data (genomics, proteomics, metabolomics) through geometric learning frameworks. Demonstrated state-of-the-art results on drug-target interaction prediction and cell type annotation tasks.

Deep Learning for Proteomics-Based Cancer Classification and Prognosis

Published:

Presented collaborative research on AI-driven proteomics analysis for cancer classification and prognosis prediction. Introduced deep learning frameworks that integrate mass spectrometry data, clinical information, and multi-omics profiles for personalized cancer medicine. Discussed novel attention mechanisms for identifying prognostic protein biomarkers and pathway-level features. Demonstrated applications in thyroid cancer, medullary carcinoma, and other malignancies. Covered the development of interpretable models that provide biological insights into cancer mechanisms while achieving superior predictive performance compared to traditional statistical methods.

Research Presentation: AI for Biological Sciences and Medical Applications

Published:

Academic job interview presentation covering comprehensive research portfolio in AI for biological sciences and medical applications. Presented major achievements including deep manifold learning for high-dimensional data visualization, foundation models for single-cell analysis, and AI-driven drug discovery platforms. Discussed future research directions including virtual cell construction, interpretable AI for medicine, and multi-modal biological data integration. Outlined collaborative opportunities with clinical and biological research groups, and proposed innovative approaches for translating AI research into clinical practice.

Boosting Unsupervised Contrastive Learning Using Diffusion-Based Data Augmentation From Scratch

Published:

Poster presentation at ICML 2024 showcasing the DiffAug framework for unsupervised contrastive learning. Demonstrated how diffusion models can generate high-quality augmented views that preserve semantic content while introducing beneficial variations. Engaged with leading researchers in self-supervised learning, discussing the theoretical foundations of diffusion-based augmentation and its advantages over traditional geometric and color transformations. Received positive feedback on the method’s generalizability across different domains including natural images, medical imaging, and scientific visualization. The work addresses fundamental challenges in learning robust visual representations without labeled data.

Technical Discussion on Large-Scale Model Training and Optimization

Published:

Internal technical discussion session focusing on challenges and best practices in training large-scale foundation models. Covered topics including distributed training strategies, gradient accumulation techniques, memory optimization, and hyperparameter tuning. Discussed infrastructure requirements for training billion-parameter models, including GPU cluster configuration, data pipeline optimization, and checkpoint management. Explored emerging techniques such as LoRA, quantization-aware training, and efficient attention mechanisms for reducing computational costs while maintaining model performance.

Phylogenetic Tree Generation and Evolutionary Analysis Using Deep Learning

Published:

Presented innovative research on phylogenetic tree construction using deep learning approaches. Introduced the BioTree framework that combines language models, graph neural networks, and evolutionary algorithms for inferring phylogenetic relationships from molecular sequences. Discussed the MDTree (Masked Dynamic Autoregressive Model) for phylogenetic inference and its advantages over traditional maximum likelihood and Bayesian methods. Covered applications in viral evolution tracking, bacterial taxonomy, and cancer phylogenetics. Demonstrated significant improvements in accuracy and computational efficiency for large-scale phylogenetic studies.

Single-Cell Foundation Models and Cell Knowledge Base Construction

Published:

Research preview presentation on building foundation models for single-cell analysis and constructing comprehensive cell knowledge bases. Introduced the concept of cell-level language models trained on millions of single-cell transcriptomic profiles. Discussed pre-training strategies, including masked cell prediction, contrastive learning on cell embeddings, and multi-task learning for cell type annotation, trajectory inference, and perturbation prediction. Covered the development of a unified cell atlas integrating data from multiple tissues, species, and technologies. Explored applications in automated cell annotation, novel cell type discovery, and understanding cellular states in disease.

Recent Developments in Retrieval-Augmented Generation for Scientific Applications

Published:

Comprehensive overview of recent advances in Retrieval-Augmented Generation (RAG) systems, with emphasis on scientific and medical applications. Discussed state-of-the-art retrieval methods including dense retrievers, hybrid search, and learned sparse representations. Covered generation models fine-tuned for domain-specific knowledge synthesis. Presented case studies on applying RAG to medical diagnosis support, drug discovery literature mining, and scientific paper summarization. Explored challenges including retrieval quality, hallucination mitigation, and factual consistency. Demonstrated practical implementations using vector databases and efficient retrieval architectures for real-time knowledge-intensive tasks.

Deep Learning for High-Dimensional Data Analysis and Visualization

Published:

Forum presentation on advanced deep learning techniques for analyzing and visualizing high-dimensional scientific data. Covered dimensionality reduction methods including deep manifold learning, variational autoencoders, and diffusion-based approaches. Discussed challenges in preserving local and global structure, handling non-linear manifolds, and maintaining interpretability. Presented applications across diverse domains: single-cell genomics (tens of thousands of genes), medical imaging (volumetric data), and materials science (molecular dynamics). Demonstrated how topology-aware neural networks can reveal hidden patterns in complex datasets while providing biologically or physically meaningful representations.

Cell Annotation and Matching Methods for Single-Cell Analysis

Published:

Detailed presentation on automated cell type annotation and cell matching algorithms for single-cell transcriptomics. Introduced novel methods for transferring cell type labels across datasets using optimal transport, contrastive learning, and attention mechanisms. Discussed challenges in batch effect correction, rare cell type identification, and handling technical variability across sequencing platforms. Demonstrated the CellScout framework for interactive cell annotation with visual analytics support. Covered applications in atlas-scale cell annotation, cross-species cell mapping, and identifying disease-associated cell states through comparative analysis.

Foundation Models for Medical Diagnosis and Clinical Decision Support

Published:

Invited lecture on the latest advances in medical foundation models and their applications in clinical practice. Presented the MedLA (Medical Logic-driven Agent) framework for interpretable diagnostic reasoning using large language models. Discussed multi-agent systems that integrate medical knowledge graphs, clinical guidelines, and patient data for comprehensive diagnosis support. Covered challenges in medical AI including reliability, explainability, and handling rare diseases. Demonstrated real-world case studies showing improved diagnostic accuracy and reduced physician cognitive load. Explored regulatory considerations and pathways for clinical deployment of AI diagnostic assistants.

Deep Manifold Learning for High-Dimensional Data Analysis: Methods, Theory, and Applications

Published:

Doctoral dissertation defense presenting comprehensive research on deep manifold learning methods for high-dimensional data analysis. The dissertation covered: (1) Theoretical foundations of topology-preserving dimensionality reduction using deep neural networks; (2) Novel architectures including DLME, DMTEV, and DMT-ME for explainable visualization; (3) Applications in single-cell genomics, medical imaging, and drug discovery; (4) Benchmarking studies demonstrating superior performance over classical methods. Committee members included leading experts in AI, computational biology, and medical informatics. Successfully defended with distinction, receiving recognition for significant contributions to both methodology and real-world applications.

Multi-Modal Learning for Complex Biological Systems: Challenges and Opportunities

Published:

Keynote address at the opening ceremony of national symposium on multi-modal biological data integration. Outlined the landscape of multi-modal learning challenges in modern biology, including integrating genomics, transcriptomics, proteomics, metabolomics, and imaging data. Discussed emerging architectures such as cross-modal transformers, multi-modal graph neural networks, and modality-specific encoders with shared latent spaces. Presented case studies on drug-cell interaction modeling, spatial transcriptomics analysis, and multi-omics disease prediction. Emphasized the need for standardized benchmarks, interpretability frameworks, and collaborative data sharing initiatives to advance the field.

Proteomics Meets Single-Cell: Integration and Analysis Progress

Published:

Progress report on integrating single-cell proteomics with transcriptomics for comprehensive cellular profiling. Presented advances in CITE-seq data analysis, surface protein marker identification, and joint embedding of protein and RNA measurements. Discussed computational methods for handling the unique characteristics of protein data including sparsity, dynamic range, and antibody-specific noise. Demonstrated applications in immune cell profiling, cancer heterogeneity analysis, and identifying protein-level dysregulation not captured by transcriptomics. Covered emerging technologies like spatial proteomics and multi-parameter flow cytometry analysis using deep learning.

AI for Precision Medicine: From Data to Clinical Insights

Published:

Workshop presentation on applying artificial intelligence to precision medicine and clinical decision-making. Covered the complete pipeline from multi-modal health data collection, quality control, and integration to predictive modeling and clinical deployment. Presented collaborative work with SANY Group on industrial health monitoring and with medical institutions on diagnostic AI systems. Discussed privacy-preserving machine learning techniques, federated learning for multi-center studies, and regulatory frameworks for medical AI. Demonstrated successful case studies including early disease detection, treatment response prediction, and personalized therapy recommendation systems.

AI Virtual Cell: Building Computational Models of Cellular Systems

Published:

Project proposal for developing AI-powered virtual cell models that simulate cellular behavior at unprecedented scale and resolution. Proposed integrating single-cell multi-omics data, protein interaction networks, metabolic pathways, and spatial organization into unified computational framework. Outlined three-phase approach: (1) Foundation model pre-training on billion-scale cellular datasets; (2) Physics-informed neural networks for modeling cellular dynamics; (3) Generative models for predicting cellular responses to perturbations. Discussed applications in drug discovery, synthetic biology, and understanding disease mechanisms. Outlined collaboration opportunities with experimental biologists and computational resources required for training and validation.

Medical Artificial Intelligence: Bridging Research and Clinical Practice

Published:

Comprehensive seminar on medical AI technologies and their translation to clinical practice. Presented work on foundation models for medical diagnosis, including the MedLA framework for interpretable clinical reasoning. Covered natural language processing for electronic health records, computer vision for medical imaging, and knowledge graph construction for clinical decision support. Discussed challenges in model validation, handling distribution shift between institutions, and ensuring patient safety. Demonstrated collaborative projects with Ant Group on developing scalable medical AI platforms, including diagnostic chatbots, automated report generation, and real-time clinical alert systems. Addressed ethical considerations and regulatory compliance for medical AI deployment.

Industrial AI Agent System: Mid-term Progress and Technical Achievements

Published:

Mid-term progress report for the Rootcloud AI Agentic System R&D project (¥400,000 funding). Presented achievements in developing end-to-end agentic AI system with integrated RAG, reinforcement learning, and multimodal understanding. Demonstrated: (1) Agentic RAG prototype achieving 60%+ retrieval recall and answer accuracy; (2) RL-optimized agent models with 5%+ improvement over baselines; (3) Multimodal context management reducing token usage by 50% while maintaining 80%+ accuracy. Showcased industrial applications in autonomous fault diagnosis, dynamic tool invocation, and multi-turn reasoning. Outlined remaining work including production deployment, scalability testing, and integration with enterprise systems.

Deep Learning for Spatial Transcriptomics: Coverage and Resolution Enhancement

Published:

Seminar on applying deep learning to enhance coverage and resolution in spatial transcriptomics data. Presented the MuST (Multimodal Structure Transformation) framework for unified spatial transcriptomics analysis. Discussed methods for imputing missing gene expression, super-resolution reconstruction of spatial maps, and integrating imaging with sequencing data. Covered attention-based architectures that learn spatial dependencies and cell-cell communication patterns. Demonstrated applications in tumor microenvironment analysis, developmental biology, and neuroscience. Showed how physics-inspired priors and biological constraints can improve model generalization and interpretability in spatial omics analysis.

Mass Spectrometry-Based Proteomics for Cancer Subtyping and Prognosis Prediction

Published:

Advanced research presentation on leveraging mass spectrometry proteomics data for cancer classification and prognosis. Presented deep learning frameworks that process raw DIA (Data-Independent Acquisition) mass spectrometry data for end-to-end prediction. Discussed novel architectures for handling peptide-level features, protein-protein interaction networks, and pathway enrichment simultaneously. Demonstrated superior performance in thyroid cancer subtyping, medullary carcinoma prognosis, and pan-cancer biomarker discovery. Covered collaborative work with clinical institutions validating AI models on prospective cohorts. Explored the potential of integrating proteomics with genomics and clinical data for personalized cancer treatment strategies.

Virtual Cell Foundation Models: Technical Architecture and Implementation

Published:

Detailed technical proposal for implementing virtual cell foundation models at BGI. Outlined comprehensive architecture including: (1) Cell-level pre-training on 100M+ single-cell profiles from BGI’s biobank; (2) Multimodal integration of transcriptomics, proteomics, and morphology; (3) Hierarchical modeling of cell states, cell types, and developmental trajectories; (4) Causal inference modules for perturbation prediction. Discussed computational infrastructure requirements, data preprocessing pipelines, and quality control protocols. Proposed validation framework using experimental perturbation studies. Defined collaboration model between AI team, wet lab scientists, and clinical researchers for iterative model refinement and real-world testing.

Multi-Modal Molecular Dynamics: Research Planning and Strategic Directions

Published:

Strategic planning session on multi-modal approaches to studying molecular dynamics and cellular behavior. Discussed integration of multiple data modalities including: structural biology (X-ray, cryo-EM), molecular dynamics simulations, experimental measurements (NMR, spectroscopy), and high-throughput screening data. Outlined research directions for developing unified frameworks that bridge time scales from femtoseconds (bond vibrations) to hours (cellular processes). Proposed combining physics-based molecular mechanics with data-driven deep learning models. Defined collaboration strategy with experimental groups, computational resource allocation, and publication roadmap. Emphasized applications in understanding drug-target interactions, protein conformational changes, and allosteric regulation mechanisms.

teaching

Deep Learning - Graduate Course (Teaching Assistant)

Graduate course, Zhejiang University / Westlake University, 2021

Served as Teaching Assistant for Prof. Stan Z. Li’s Deep Learning course during Ph.D. studies. This graduate-level course covered fundamental and advanced topics in deep learning, including neural network architectures, optimization methods, convolutional neural networks, recurrent neural networks, attention mechanisms, and generative models.

Deep Learning - Graduate Course (Teaching Assistant)

Graduate course, Zhejiang University / Westlake University, 2022

Continued as Teaching Assistant for Prof. Stan Z. Li’s Deep Learning course in the second semester. Building on the experience from the previous year, enhanced the course with updated content on recent advances and more challenging project assignments.