Experience

Predicting protein function from its primary sequence is a fundamental challenge in computational biology. While deep learning has excelled, the optimal representation of sequence data remains an open question. This study explores protein sonification---the conversion of amino acid sequences into 2D spectrograms---as a representation for this task. To facilitate this investigation, we developed a benchmark dataset of 18,000 sequences spanning 12 functionally diverse protein classes. Our systematic evaluation suggests that the structural transformation from a 1D sequence to a 2D spectrogram may be a key contributor to the model's predictive performance. This observation is supported by ablation studies where models using either purely visual or acoustic features from the spectrogram demonstrated effective standalone performance, suggesting that the representation itself is a key source of this capability. For instance, a model using a sonification map without explicit biophysical meaning achieved 81.08% accuracy, while our biophysically-informed model reached 84.00%, indicating that such domain knowledge may offer a modest performance benefit. When trained from scratch on our dataset, our fusion model achieved performance comparable to or slightly exceeding that of standard transformer architectures like ESM-2 and ProtBERT, suggesting its potential for data efficiency in this specific context. The model's potential for generalizability was further supported by its performance on the external CARE enzyme classification benchmark, where it achieved 90.44% accuracy. Finally, as a proof-of-concept, we explore the utility of our encoding to guide a diffusion model in generating novel GFP variants, which were assessed for structural viability using computational methods. Our work provides evidence suggesting that the utility of sonification in this context may stem largely from its representational structure, offering a perspective on feature engineering for biological sequences.

This project aims to integrate copy number variation (CNV) features from whole-genome sequencing data with multi-dimensional clinical data such as metabolic indicators. Leveraging the success of diffusion models in processing high-dimensional simulation data and protein phenotype prediction, we will construct a comprehensive framework consisting of CNV feature encoding, genomic region attention, and conditional U-Net diffusion modules. This will simulate CNV distribution changes and evolutionary processes in the genome, analyze the specific role of CNV in regulating Alzheimer's disease pathways, and ultimately improve disease risk assessment and early intervention accuracy.

Compound heatwaves increasingly trigger cascading failures that propagate through interconnected physical and human systems, yet fragmented disciplinary knowledge makes these systemic risk topologies difficult to map. HeDA constructs a knowledge graph from 8,365 academic publications, structuring 34,933 entities and 42,890 directed relationships from a temporally partitioned corpus (through 2024, n = 5,739). On a temporally held-out multi-hop reasoning benchmark (2025–2026, n = 3,892), graph augmentation improves accuracy by 8.1%–9.8% over four standalone foundation-model baselines. Topological analysis identifies biological systems as structural mediators linking thermal stress to systemic socioeconomic losses. The framework also identifies latent functional couplings between sectors, including the synchronization of power-grid failures and emergency medical capacity saturation under heat stress. These findings provide a structural basis for adaptation strategies that address resilience across interconnected systems.

The $k$-ary $n$-cube ($Q_n^k$) is a critical topology for large-scale computing systems powering modern AI and HPC workloads, where fault tolerance is paramount. Traditional fault models, which assume faults are independent and random, yield unrealistic resilience estimates because real-world failures are spatially correlated, manifesting as topological clusters. This paper introduces the Region-Based Fault (RBF) model, a new paradigm that addresses this gap by directly modeling this spatial correlation. Our primary contribution is a proof that for odd $k \geq 3$ and $n \geq 2$, the $Q_n^k$ remains Hamiltonian-connected—a property vital for deadlock-free routing and efficient task scheduling—under a sufficient set of RBF conditions. We present a constructive algorithm that finds a Hamiltonian path by leveraging an adaptive decomposition strategy. Experimental analysis demonstrates that our approach significantly enhances fault tolerance and remains robust far beyond its conservative theoretical guarantees. The resulting framework provides a practical, high-performance solution for maintaining connectivity in systems where fault clustering is prevalent.

This project, centered on artificial intelligence digital twin technology, aims to alleviate the lack of family companionship for the elderly in China's aging society, bridge the digital divide for older adults, and enhance primary and secondary school students' artificial intelligence practical capabilities. Implemented by Shenzhen X-Institute in collaboration with ByteDance and other organizations, the project is divided into preparation, implementation, and promotion phases. It plans to facilitate students in customizing personalized digital twins for the elderly, enhancing intergenerational communication, and exploring application scenarios of AI large models in social welfare. The project not only provides companionship and psychological support for the elderly at a spiritual level but also opens up a new path for Chinese-style elderly care while promoting artificial intelligence education and social innovation development. Through multi-party collaboration, risk management, and media promotion, the project will achieve the goal of progressing from pilot programs to nationwide promotion, with significant social, educational, and technological implications.