MIRACLE: Multi-task Learning based Interpretable Regulation of Autoimmune Diseases through Common Latent Epigenetics

DNA methylation is a crucial regulator of gene transcription and has been linked to various diseases, including autoimmune diseases and cancers. However, diagnostics based on DNA methylation face challenges due to large feature sets and small sample sizes, resulting in overfitting and suboptimal performance. To address these issues, we propose MIRACLE, a novel interpretable neural network that leverages autoencoder-based multi-task learning to integrate multiple datasets and jointly identify common patterns in DNA methylation. MIRACLE's architecture reflects the relationships between methylation sites, genes, and pathways, ensuring biological interpretability and meaningfulness. The network comprises an encoder and a decoder, with a bottleneck layer representing pathway information as the basic unit of heredity. Customized defined MaskedLinear Layer is constrained by site-gene-pathway graph adjacency matrix information, which provides explainability and expresses the site-gene-pathway hierarchical structure explicitly. And from the embedding, there are different multi-task classifiers to predict diseases. Tested on six datasets, including rheumatoid arthritis, systemic lupus erythematosus, multiple sclerosis, inflammatory bowel disease, psoriasis, and type 1 diabetes, MIRACLE demonstrates robust performance in identifying common functions of DNA methylation across different phenotypes, with higher accuracy in prediction dieseases than baseline methods. By incorporating biological prior knowledge, MIRACLE offers a meaningful and interpretable framework for DNA methylation data analysis in the context of autoimmune diseases.

Heterogeneous Value Evaluation for Large Language Models

The emergent capabilities of Large Language Models (LLMs) have made it crucial to align their values with those of humans. Current methodologies typically attempt alignment with a homogeneous human value and requires human verification, yet lack consensus on the desired aspect and depth of alignment and resulting human biases. In this paper, we propose A2EHV, an Automated Alignment Evaluation with a Heterogeneous Value system that (1) is automated to minimize individual human biases, and (2) allows assessments against various target values to foster heterogeneous agents. Our approach pivots on the concept of value rationality, which represents the ability for agents to execute behaviors that satisfy a target value the most. The quantification of value rationality is facilitated by the Social Value Orientation framework from social psychology, which partitions the value space into four categories to assess social preferences from agents' behaviors. We evaluate the value rationality of eight mainstream LLMs and observe that large models are more inclined to align neutral values compared to those with strong personal values. By examining the behavior of these LLMs, we contribute to a deeper understanding of value alignment within a heterogeneous value system.