Selected work

Representative Papers

FACET: Factorized Asymmetric Conditioning for Efficient Transport in High-Fidelity Fluorescence Microscopy Synthesis

Oct 04, 2026

This study addresses the asymmetry and coupling of sequential and morphological conditions in fluorescence microscopy image generation by proposing FACET, a framework for high-fidelity localization synthesis of unimaged proteins. FACET introduces explicit inductive biases to disentangle conditional explanatory power through a factorized asymmetric encoding architecture. It leverages cross-protein semantic memory to share coarse-grained patterns while modeling fine-grained bounded residuals, combined with variance-preserving state projection for efficient diffusion transport. Experimental results demonstrate that, with minimal parameter overhead, FACET improves spatial overlap by 34.3%, reduces FID by 27.2%, and decreases network evaluation calls by 75%, significantly enhancing the recovery of biologically relevant structures and predictive calibration.

0 citationsRead paper

PRISM: Enhancing Protein Inverse Folding through Fine-Grained Retrieval on Structure-Sequence Multimodal Representations

Oct 11, 2025

Protein inverse folding—generating sequences compatible with a target 3D structure—is highly challenging due to the vast sequence space and complex local structural constraints. This paper proposes a fine-grained multimodal retrieval-augmented generation framework: first, retrieving evolutionarily conserved local structure–sequence motifs from a natural protein database; second, integrating retrieved motifs with the target structure via a hybrid self-cross-attention decoder; and third, explicitly modeling and reusing these motifs through a latent-variable probabilistic model. To our knowledge, this is the first approach to incorporate fine-grained, naturally occurring structure–sequence co-occurrence patterns into inverse folding. Evaluated on five standard benchmarks, our method achieves significant improvements in amino acid recovery rate and perplexity, while simultaneously enhancing foldability metrics—including RMSD, TM-score, and pLDDT—demonstrating both effectiveness and strong generalization capability.

0 citationsRead paper

Optimization over Sparse Support-Preserving Sets: Two-Step Projection with Global Optimality Guarantees

Jun 10, 2025

We address sparse optimization problems with a support-set preservation constraint—a mixed nonconvex–convex constraint rendering Euclidean projection infeasible or only locally convergent. To circumvent reliance on closed-form mixed projections, we propose a novel two-step successive projection iterative hard-thresholding algorithm. First, we establish a generalized three-point lemma for nonconvex two-step projections. Second, under deterministic, stochastic, and zeroth-order optimization settings, we provide global suboptimality guarantees on the objective value, with explicit error bounds. Notably, in the zeroth-order setting, our method achieves a faster convergence rate than existing approaches, without requiring unbiased gradient estimates—thereby eliminating systematic bias entirely. This work significantly advances the theoretical foundations and practical applicability of sparse optimization.

0 citationsRead paper

Convergence of Spectral Principal Paths: How Deep Networks Distill Linear Representations from Noisy Inputs

Jun 10, 2025

This study addresses the lack of structured theoretical foundations for representation mechanisms in deep neural networks (DNNs), particularly regarding how DNNs learn interpretable and robust linear representations from noisy inputs. To this end, we propose the **Input-Space Linearity Hypothesis (ISLH)** and introduce the **Spectral Principal Path (SPP) framework**, the first to characterize the dynamical process of layer-wise representation distillation and convergence via spectral analysis. By modeling principal paths, analyzing interpretability along linear directions, and conducting empirical validation on multimodal large language models, we demonstrate that deep networks progressively converge—along a small set of dominant spectral directions—to human-interpretable concept subspaces. This mechanism substantially enhances representation transparency, cross-domain robustness, and fairness. Extensive experiments on vision-language models confirm its generalizability and effectiveness.

0 citationsRead paper

Few-shot Species Range Estimation

Feb 20, 2025

Addressing the challenge of estimating species distribution ranges under sparse geographical observations (only 5–10 locations), this paper introduces a novel multimodal prompting-based species encoding paradigm, enabling, for the first time, text- and image-guided zero-shot cross-species generalization. Our method integrates geographical embeddings, contrastive learning, and a lightweight multimodal encoder within a meta-learning framework to model species-specific spatial priors, supporting efficient feed-forward inference. Evaluated on two standard benchmarks, it achieves state-of-the-art performance—improving average AUC by 2.1%, accelerating inference by 3.2×, and reducing parameter count by 47%. The core contribution lies in incorporating multimodal semantic priors into species distribution modeling, substantially mitigating spatial extrapolation bias under few-shot conditions. This yields a highly efficient and scalable prediction tool for endangered species conservation.

0 citationsRead paper

Recent publications

Latest Papers

FACET: Factorized Asymmetric Conditioning for Efficient Transport in High-Fidelity Fluorescence Microscopy Synthesis

Oct 04, 2026

This study addresses the asymmetry and coupling of sequential and morphological conditions in fluorescence microscopy image generation by proposing FACET, a framework for high-fidelity localization synthesis of unimaged proteins. FACET introduces explicit inductive biases to disentangle conditional explanatory power through a factorized asymmetric encoding architecture. It leverages cross-protein semantic memory to share coarse-grained patterns while modeling fine-grained bounded residuals, combined with variance-preserving state projection for efficient diffusion transport. Experimental results demonstrate that, with minimal parameter overhead, FACET improves spatial overlap by 34.3%, reduces FID by 27.2%, and decreases network evaluation calls by 75%, significantly enhancing the recovery of biologically relevant structures and predictive calibration.

0 citationsRead paper

PRISM: Enhancing Protein Inverse Folding through Fine-Grained Retrieval on Structure-Sequence Multimodal Representations

Oct 11, 2025

Protein inverse folding—generating sequences compatible with a target 3D structure—is highly challenging due to the vast sequence space and complex local structural constraints. This paper proposes a fine-grained multimodal retrieval-augmented generation framework: first, retrieving evolutionarily conserved local structure–sequence motifs from a natural protein database; second, integrating retrieved motifs with the target structure via a hybrid self-cross-attention decoder; and third, explicitly modeling and reusing these motifs through a latent-variable probabilistic model. To our knowledge, this is the first approach to incorporate fine-grained, naturally occurring structure–sequence co-occurrence patterns into inverse folding. Evaluated on five standard benchmarks, our method achieves significant improvements in amino acid recovery rate and perplexity, while simultaneously enhancing foldability metrics—including RMSD, TM-score, and pLDDT—demonstrating both effectiveness and strong generalization capability.

0 citationsRead paper

Optimization over Sparse Support-Preserving Sets: Two-Step Projection with Global Optimality Guarantees

Jun 10, 2025

We address sparse optimization problems with a support-set preservation constraint—a mixed nonconvex–convex constraint rendering Euclidean projection infeasible or only locally convergent. To circumvent reliance on closed-form mixed projections, we propose a novel two-step successive projection iterative hard-thresholding algorithm. First, we establish a generalized three-point lemma for nonconvex two-step projections. Second, under deterministic, stochastic, and zeroth-order optimization settings, we provide global suboptimality guarantees on the objective value, with explicit error bounds. Notably, in the zeroth-order setting, our method achieves a faster convergence rate than existing approaches, without requiring unbiased gradient estimates—thereby eliminating systematic bias entirely. This work significantly advances the theoretical foundations and practical applicability of sparse optimization.

0 citationsRead paper

Convergence of Spectral Principal Paths: How Deep Networks Distill Linear Representations from Noisy Inputs

Jun 10, 2025

This study addresses the lack of structured theoretical foundations for representation mechanisms in deep neural networks (DNNs), particularly regarding how DNNs learn interpretable and robust linear representations from noisy inputs. To this end, we propose the **Input-Space Linearity Hypothesis (ISLH)** and introduce the **Spectral Principal Path (SPP) framework**, the first to characterize the dynamical process of layer-wise representation distillation and convergence via spectral analysis. By modeling principal paths, analyzing interpretability along linear directions, and conducting empirical validation on multimodal large language models, we demonstrate that deep networks progressively converge—along a small set of dominant spectral directions—to human-interpretable concept subspaces. This mechanism substantially enhances representation transparency, cross-domain robustness, and fairness. Extensive experiments on vision-language models confirm its generalizability and effectiveness.

0 citationsRead paper

Few-shot Species Range Estimation

Feb 20, 2025

Addressing the challenge of estimating species distribution ranges under sparse geographical observations (only 5–10 locations), this paper introduces a novel multimodal prompting-based species encoding paradigm, enabling, for the first time, text- and image-guided zero-shot cross-species generalization. Our method integrates geographical embeddings, contrastive learning, and a lightweight multimodal encoder within a meta-learning framework to model species-specific spatial priors, supporting efficient feed-forward inference. Evaluated on two standard benchmarks, it achieves state-of-the-art performance—improving average AUC by 2.1%, accelerating inference by 3.2×, and reducing parameter count by 47%. The core contribution lies in incorporating multimodal semantic priors into species distribution modeling, substantially mitigating spatial extrapolation bias under few-shot conditions. This yields a highly efficient and scalable prediction tool for endangered species conservation.

0 citationsRead paper