Selected work

Representative Papers

Atelier: Learning Local Self-Supervised Features for CryoEM Volumes via Hypernetworks

Sep 24, 2026

This study addresses the high computational cost of independently fitting implicit neural representations (INRs) to cryo-EM volumes and the resulting misalignment of representations across samples. To overcome these limitations, this work proposes Atelier, a self-supervised framework that pioneers the use of amortized INRs as geometry-aware analytical primitives. By leveraging a Transformer-based hypernetwork to amortize INR fitting, Atelier generates coordinate-conditioned continuous local feature fields that transcend the constraints of voxel grids and tokenizers, thereby facilitating 3D segmentation tasks. Pre-trained on 5,439 cryo-EM density maps, the proposed framework achieves high-fidelity reconstruction of multi-subunit assemblies and significantly outperforms existing baselines across eight voxel-level attribute prediction tasks.

0 citationsRead paper

Higher-Order Cell Tracking Transformer

Jul 13, 2026

This work addresses the challenge of lineage tracking in live-cell microscopy, where cell divisions lead to entangled trajectories that existing methods struggle to disambiguate due to insufficient graph-topological supervision. To overcome this, the authors propose an edge-centric Transformer architecture that, for the first time, incorporates high-order relational modeling into cell tracking. By leveraging an edge-centered attention mechanism fused with 3D geometric priors, the model enables end-to-end optimization without requiring a pretrained deep image encoder. The method achieves state-of-the-art performance on both the Cell Tracking Challenge and bacterial division benchmarks. Notably, with only 400 annotated samples, human-in-the-loop fine-tuning reduces tracking errors by 59%, substantially outperforming a LoRA-finetuned Transformer baseline by 6.75%.

0 citationsRead paper

4D Vessel Reconstruction for Benchtop Thrombectomy Analysis

Apr 08, 2026

This study addresses the lack of time-resolved, full-field three-dimensional measurement techniques for vascular deformation and procedural injury in mechanical thrombectomy experiments. The authors propose a low-cost, nine-camera multi-view system that, for the first time, employs 4D Gaussian splatting to reconstruct dynamic surface geometry in a silicone middle cerebral artery model. Regional displacements are tracked using a fixed-connectivity edge graph, and a Neo-Hookean constitutive model is leveraged to compute a proxy metric for relative surface stress. This approach enables standardized, time-resolved quantification of vascular kinematics and stress, facilitating quantitative comparisons across different procedural conditions. Synthetic validation demonstrates displacement reconstruction accuracy of 0.964–0.972 and Chamfer distances of 1.714–1.815 mm. Preliminary benchtop experiments indicate that catheter insertion during carotid aspiration thrombectomy induces greater displacement and stress responses.

0 citationsRead paper

How Well Do LLMs Understand Drug Mechanisms? A Knowledge + Reasoning Evaluation Dataset

Nov 09, 2025

This study evaluates large language models’ (LLMs) capabilities in understanding and reasoning about drug mechanisms, focusing on factual recall of established mechanisms and causal reasoning under counterfactual scenarios. To this end, we introduce the first open-world evaluation dataset specifically designed for drug mechanism pathways and propose a novel assessment framework that jointly integrates knowledge retrieval and chain-of-thought reasoning. Crucially, we innovate by applying fine-grained counterfactual perturbations to internal steps within mechanism pathways—substantially increasing reasoning complexity. Our experimental design comprises both open-world settings (requiring autonomous knowledge recall) and closed-world settings (where factual premises are provided). Results demonstrate that o4-mini achieves the highest performance, while Qwen3-4B-thinking attains comparable or locally superior performance, validating the framework’s effectiveness and generalizability across diverse LLMs.

0 citationsRead paper

Recent publications

Latest Papers

Atelier: Learning Local Self-Supervised Features for CryoEM Volumes via Hypernetworks

Sep 24, 2026

This study addresses the high computational cost of independently fitting implicit neural representations (INRs) to cryo-EM volumes and the resulting misalignment of representations across samples. To overcome these limitations, this work proposes Atelier, a self-supervised framework that pioneers the use of amortized INRs as geometry-aware analytical primitives. By leveraging a Transformer-based hypernetwork to amortize INR fitting, Atelier generates coordinate-conditioned continuous local feature fields that transcend the constraints of voxel grids and tokenizers, thereby facilitating 3D segmentation tasks. Pre-trained on 5,439 cryo-EM density maps, the proposed framework achieves high-fidelity reconstruction of multi-subunit assemblies and significantly outperforms existing baselines across eight voxel-level attribute prediction tasks.

0 citationsRead paper

Higher-Order Cell Tracking Transformer

Jul 13, 2026

This work addresses the challenge of lineage tracking in live-cell microscopy, where cell divisions lead to entangled trajectories that existing methods struggle to disambiguate due to insufficient graph-topological supervision. To overcome this, the authors propose an edge-centric Transformer architecture that, for the first time, incorporates high-order relational modeling into cell tracking. By leveraging an edge-centered attention mechanism fused with 3D geometric priors, the model enables end-to-end optimization without requiring a pretrained deep image encoder. The method achieves state-of-the-art performance on both the Cell Tracking Challenge and bacterial division benchmarks. Notably, with only 400 annotated samples, human-in-the-loop fine-tuning reduces tracking errors by 59%, substantially outperforming a LoRA-finetuned Transformer baseline by 6.75%.

0 citationsRead paper

4D Vessel Reconstruction for Benchtop Thrombectomy Analysis

Apr 08, 2026

This study addresses the lack of time-resolved, full-field three-dimensional measurement techniques for vascular deformation and procedural injury in mechanical thrombectomy experiments. The authors propose a low-cost, nine-camera multi-view system that, for the first time, employs 4D Gaussian splatting to reconstruct dynamic surface geometry in a silicone middle cerebral artery model. Regional displacements are tracked using a fixed-connectivity edge graph, and a Neo-Hookean constitutive model is leveraged to compute a proxy metric for relative surface stress. This approach enables standardized, time-resolved quantification of vascular kinematics and stress, facilitating quantitative comparisons across different procedural conditions. Synthetic validation demonstrates displacement reconstruction accuracy of 0.964–0.972 and Chamfer distances of 1.714–1.815 mm. Preliminary benchtop experiments indicate that catheter insertion during carotid aspiration thrombectomy induces greater displacement and stress responses.

0 citationsRead paper

How Well Do LLMs Understand Drug Mechanisms? A Knowledge + Reasoning Evaluation Dataset

Nov 09, 2025

This study evaluates large language models’ (LLMs) capabilities in understanding and reasoning about drug mechanisms, focusing on factual recall of established mechanisms and causal reasoning under counterfactual scenarios. To this end, we introduce the first open-world evaluation dataset specifically designed for drug mechanism pathways and propose a novel assessment framework that jointly integrates knowledge retrieval and chain-of-thought reasoning. Crucially, we innovate by applying fine-grained counterfactual perturbations to internal steps within mechanism pathways—substantially increasing reasoning complexity. Our experimental design comprises both open-world settings (requiring autonomous knowledge recall) and closed-world settings (where factual premises are provided). Results demonstrate that o4-mini achieves the highest performance, while Qwen3-4B-thinking attains comparable or locally superior performance, validating the framework’s effectiveness and generalizability across diverse LLMs.

0 citationsRead paper