Xiangyu Wen

A standardized benchmarking framework for single-cell perturbation prediction, harmonizing datasets, model interfaces, and evaluation protocols. It covers three out-of-distribution scenarios — unseen cell contexts, unseen perturbations, and cross-dataset generalization — revealing that task design strongly influences model ranking and that cross-dataset generalization remains the key challenge.

RNA inverse folding designs sequences that fold into a target 3D backbone. We address the "Long RNA Dilemma": GNNs capture local geometry but miss long-range couplings, while global attention incurs O(N²) cost. GALS-Fold combines an SE(3)-equivariant short-range GNN with linear-time anchor attention and length-aware gating, achieving 52.2% recovery (SOTA) and 62.4% on long sequences (>200 nt).

A CNN-BiGRU-Attention network for suppressing ocular and myogenic artifacts in EEG signals. CNN extracts spatial features, BiGRU models temporal dependencies, and attention adaptively localizes artifacts, outperforming conventional baselines across multiple metrics.