SeqMaestro: From nucleotide sequences to biological hypotheses through interpretable machine learning
SeqMaestro通过可解释的机器学习模型从核苷酸序列提出生物学假设，解决了传统方法灵活性不足和深度学习模型难以解释的问题。
0 citationsRead paper
Selected work
Representative Papers
Advanced Brain Tissue Imaging with Data-Consistent Diffusion Priors in Laminographic X-Ray Nanoimaging
为解决X射线层析成像中因数据缺失导致的脑组织成像失真问题，提出LUCID框架，结合多视角扩散先验与投影域数据一致性方法恢复未测量信息。
0 citationsRead paper
BIOMERO 2.0: end-to-end FAIR infrastructure for bioimaging data import, analysis, and provenance
Current bioimaging analysis lacks an end-to-end platform compliant with FAIR principles and equipped with fine-grained provenance tracking, resulting in fragmented workflows and poor reproducibility. To address this, we propose a two-layer provenance architecture that deeply integrates OMERO throughout the entire data ingestion and analysis pipeline. Leveraging our custom BIOMERO library, OMERO.web plugins, and containerized analysis components, the platform enables full-chain traceability—from image acquisition and preprocessing to analysis and sharing. It supports automated workflow orchestration, rich metadata annotation, and high-throughput computing integration, ensuring real-time logging of parameters, software versions, and analytical results. Unlike existing solutions, this work is the first to embed native provenance capture directly into the analysis stage within the OMERO ecosystem. It significantly enhances data findability, interoperability, and reusability, providing foundational infrastructure for standardized, verifiable bioimaging research.
0 citationsRead paper
Inference with correlated priors using sisters cells
In neural perception, conventional probabilistic inference requiring direct connections between latent units becomes biologically implausible and computationally prohibitive when latent variables exhibit genuine statistical dependencies. To address this, we propose a novel neural circuit mechanism inspired by “sister cells”: pairs of neurons sharing common input but exhibiting divergent local connectivity, thereby indirectly encoding prior correlations among latent variables without direct inter-latent connections. We introduce a geometric construction method for synaptic connectivity that formally establishes the invertibility of prior structure from circuit architecture. Integrating geometric modeling, dynamical circuit simulation, and probabilistic inference theory, we validate the mechanism within an olfaction-inspired architecture. Results demonstrate significantly improved inference accuracy under noise, dynamical behaviors consistent with experimental observations, and—under biologically reasonable assumptions—the unique reconstruction of the latent prior structure from sister-cell activity alone.
0 citationsRead paper
Recent publications
Latest Papers
SeqMaestro: From nucleotide sequences to biological hypotheses through interpretable machine learning
SeqMaestro通过可解释的机器学习模型从核苷酸序列提出生物学假设，解决了传统方法灵活性不足和深度学习模型难以解释的问题。
0 citationsRead paper
Advanced Brain Tissue Imaging with Data-Consistent Diffusion Priors in Laminographic X-Ray Nanoimaging
为解决X射线层析成像中因数据缺失导致的脑组织成像失真问题，提出LUCID框架，结合多视角扩散先验与投影域数据一致性方法恢复未测量信息。
0 citationsRead paper
BIOMERO 2.0: end-to-end FAIR infrastructure for bioimaging data import, analysis, and provenance
Current bioimaging analysis lacks an end-to-end platform compliant with FAIR principles and equipped with fine-grained provenance tracking, resulting in fragmented workflows and poor reproducibility. To address this, we propose a two-layer provenance architecture that deeply integrates OMERO throughout the entire data ingestion and analysis pipeline. Leveraging our custom BIOMERO library, OMERO.web plugins, and containerized analysis components, the platform enables full-chain traceability—from image acquisition and preprocessing to analysis and sharing. It supports automated workflow orchestration, rich metadata annotation, and high-throughput computing integration, ensuring real-time logging of parameters, software versions, and analytical results. Unlike existing solutions, this work is the first to embed native provenance capture directly into the analysis stage within the OMERO ecosystem. It significantly enhances data findability, interoperability, and reusability, providing foundational infrastructure for standardized, verifiable bioimaging research.
0 citationsRead paper
Inference with correlated priors using sisters cells
In neural perception, conventional probabilistic inference requiring direct connections between latent units becomes biologically implausible and computationally prohibitive when latent variables exhibit genuine statistical dependencies. To address this, we propose a novel neural circuit mechanism inspired by “sister cells”: pairs of neurons sharing common input but exhibiting divergent local connectivity, thereby indirectly encoding prior correlations among latent variables without direct inter-latent connections. We introduce a geometric construction method for synaptic connectivity that formally establishes the invertibility of prior structure from circuit architecture. Integrating geometric modeling, dynamical circuit simulation, and probabilistic inference theory, we validate the mechanism within an olfaction-inspired architecture. Results demonstrate significantly improved inference accuracy under noise, dynamical behaviors consistent with experimental observations, and—under biologically reasonable assumptions—the unique reconstruction of the latent prior structure from sister-cell activity alone.
0 citationsRead paper