Seminars

Abstract

Graphs naturally represent many complex systems, from biological interaction networks to social and information networks. Although Graph Neural Networks (GNNs) have achieved remarkable success, their performance critically depends on the design of graph convolution operators. Existing spectral GNNs often rely on simple filters or learned polynomial coefficients, limiting their ability to capture complex graph structures. Moreover, constructing expressive spectral filters is computationally challenging due to the numerical instability of conventional polynomial approximation methods.

In this talk, I will present a unified spectral graph learning framework based on Arnoldi orthonormalization that enables stable polynomial interpolation of arbitrary spectral graph filters. This framework efficiently incorporates sophisticated filters—including heat kernels, random walks, and beta kernels—into graph neural networks and graph contrastive learning. I will also introduce ArnoldiGCL, which leverages learnable spectral filters to generate expressive graph views, achieving state-of-the-art representation learning on both homophilic and heterophilic graphs.

The primary application of this framework is single-cell proteomics, where noisy measurements, missing values, and imperfect cell-to-cell similarity graphs pose significant challenges for cell type identification. I will present scProfiterole, a graph contrastive learning framework that combines Arnoldi-based spectral filtering with biologically meaningful graph filters to improve clustering of single-cell proteomic data. Experimental results demonstrate that carefully designed spectral filters substantially improve robustness and clustering accuracy compared to conventional graph convolution methods.

I will also briefly discuss how the same spectral learning framework extends to trustworthy AI, including anomaly detection in multi-relational graphs through CaRgo, where adaptive spectral filters improve the detection of fraudulent and anomalous behavior in noisy, heterogeneous networks. Finally, I will present my ongoing research on Graph Retrieval-Augmented Generation (GraphRAG), which integrates graph representation learning with large language models to improve knowledge retrieval, multi-hop reasoning, and question answering over structured and heterogeneous graph data.

Overall, this talk demonstrates how advances in spectral graph learning provide a unified and scalable foundation for graph representation learning, with applications spanning computational biology, healthcare, cybersecurity, trustworthy artificial intelligence, and graph-based retrieval-augmented generation.

Abstract

Transcriptional gene regulation governs cellular identity by translating the static genomic sequence into dynamic gene expression programs. A central component of this process is the binding of transcription factors (TFs) to cis-regulatory elements such as promoters and enhancers. Deciphering the regulatory logic of TF-DNA interactions remains a central challenge, as TF binding is governed by complex, context-dependent interactions that extend beyond canonical sequence motifs. Traditional models, such as position frequency matrices, fail to capture these combinatorial and structural dependencies.

This talk will explore how interpretable deep learning models can learn high-resolution, context-aware representations of TF binding directly from genomic sequence. These approaches are complemented by scalable computational pipelines and database infrastructure that enable the training, interpretation, and integration of such models into JASPAR, an international open- access community resource and ELIXIR core data resource. Model-derived representations further provide a framework for systematically characterizing regulatory grammar across cellular contexts. Building on these efforts, the talk will conclude with a research program centered on multimodal and interpretable AI for regulatory genomics. Integrating sequence, structural, and chromatin features with computational modeling, this program seeks to uncover the mechanistic principles governing gene regulation and its perturbation in disease. More broadly, the goal is to develop AI methods that are not only predictive but also interpretable, biologically grounded, and capable of generating testable hypotheses about regulatory mechanisms.

Abstract

Computed tomography (CT) is widely used in clinical diagnosis because of its high spatial resolution and rapid imaging capability. However, the biological effects of exposure to ionizing radiation and its association with long-term cancer risk make radiation dose reduction particularly important, especially in repeated examinations. Reducing the radiation dose decreases the number of detected X-ray photons, which in turn increases image noise and may reduce the visibility of low-contrast or diagnostically relevant structures. Although deep learning-based methods have achieved substantial progress in low-dose CT denoising, challenges such as oversmoothing, loss of fine anatomical details, and the reliability of image content generated by generative models mean that the balance between noise suppression and preservation of diagnostic information remains an important research problem.

This talk will focus on modelling noise in low-dose CT and preserving diagnostic information while reducing noise using deep learning methods. In this context, recent developments in low-dose CT denoising will be discussed together with our previous work, ongoing research on noise modelling, and new approaches that more explicitly account for dose and noise characteristics. Finally, future research directions toward more reliable preservation of diagnostically important structures through the joint use of projection-domain and image-domain information will be presented.