A major mystery in ALS is selective neuronal vulnerability: some motor neurons die whereas others are resilient. My goal is to define the molecular and cellular basis of selective neuron vulnerability. We generated longitudinal single-nucleus transcriptomes and chromatin accessibility profiles of spinal motor neurons together with spatial transcriptomics from the SOD1-G93A mouse model. Vulnerable alpha motor neurons showed thousands of molecular changes, marking a transition into a distinct cell state we named "disease-associated motor neurons" (DMs). We identified transcription factor networks that govern how healthy cells transition into DMs and those associated with motor neuron subtype-selective vulnerability. Upregulation of DM-associated transcription factors in human motor neurons induced key features of DMs, demonstrating an active regulatory component. Human ALS spinal cord single-nucleus RNA sequencing data demonstrated conservation of the DM signature in alpha motor neurons, and human orthologs of regions differentially accessible in SOD1-G93A mouse motor neurons were enriched for ALS genetic risk variants. Together, these findings establish a conserved, genetically linked motor neuron signature in ALS.

Engineering and manufacturing require decisions based on data that are often scarce, noisy, and collected from different experiments and simulations. How can machine learning help engineers use this evidence to predict system behaviour, understand underlying mechanisms, and make better decisions?

Building robots that can act reliably across changing environments and tasks requires more than mapping observations directly to actions. Robots must represent useful structure in their bodies, tasks, environments, and prior experience. My research asks what structure should be represented or learned in advance, and how it can be reused to generate action efficiently online. In this talk, I will present a progression from robot geometry to task structure and learned behavior. I will discuss implicit and configuration-space distance fields for collision, contact, planning, and control; Riemannian geometry and learned geodesic distance fields for task-dependent motion generation; and geometric representations of demonstrations that support reactive imitation and visuomotor control. Across these works, a common principle is to represent reusable structure offline and construct behavior online through efficient queries, optimization, and composition. I will conclude with directions toward structured, physically grounded embodied intelligence.

The human genome is folded into chromatin loops by the cohesin complex, forming functional chromatin domains that underlie transcription and DNA replication/repair. However, how cohesin organizes these domains in living cells and the biological relevance of this organization, especially for transcriptional regulation, remain elusive. To address these questions, we combined single-nucleosome imaging/tracking using Repli-Histo labeling [1], super-resolution imaging (3D-SIM [2] and STORM [3]), and euchromatin-specific labeling of histone H3.3 [4]. Using this nanoscopic approach, we revealed that euchromatin forms condensed domains that are constrained by cohesin-mediated loops [5]. These condensed domains exhibit characteristic diameters on the order of a few hundred nanometers (with peak diameters of ~200–240 nm in STORM-based clustering analysis). This organization refines the classical textbook view of euchromatin as largely open, in line with emerging evidence [6, 7]. Transcriptional machinery appears to be located near the condensed domain surfaces/borders [5]. Cohesin loss increased nucleosome fluidity within these domains without altering their overall compaction, leading to local “mixing” of domains and compromising transcriptional insulation [5]. These findings uncover an unexpected physical role of cohesin in maintaining the integrity of condensed euchromatic domains and ensuring proper higher-order regulation of gene expression in the cell.

Connected and automated vehicle technologies are creating new opportunities for future transportation systems. While these technologies improve vehicle-level intelligence, their system-level benefits depend strongly on how individual vehicles are coordinated.

Post-translational modification (PTM) plays a critical role in disease biology, yet identifying peptides with multiple PTMs remains computationally challenging. In this talk, we present our effort in tackling this computational challenging by using different methods, including coding method, simulation study, greedy method, and a combinatorial optimization method.

Aging is accompanied by dynamic remodeling of cellular metabolism, but many key metabolic transitions remain difficult to measure with spatial and temporal precision. NAD+ biology provides a powerful entry point because it connects mitochondrial function, DNA repair, inflammation, stress response, and cellular repair.

The rapid advancement of generative artificial intelligence (GenAI) has introduced novel opportunities for semantic communication (SemCom) systems. This talk offers a comprehensive overview of GenAI-enabled SemCom, connecting theoretical foundations with practical applications. Initially, we introduce the fundamental concepts of SemCom and explore how generative models augment traditional communication paradigms. The talk systematically reviews state-of-the-art methodologies, including variational autoencoders, generative adversarial networks, diffusion models, and other GenAI frameworks within SemCom contexts. We classify GenAI in SemCom based on its GenAI architecture, communication modality, and application tasks. Finally, we identify emerging research directions and discuss open challenges that merit further investigation.