Mingzhi Yuan

Virtual screening is a key technique for prioritizing active compounds from massive chemical libraries. While recent learning-based methods have achieved remarkable success, most existing approaches generate screening scores solely based on the query target–ligand pair, overlooking valuable interaction knowledge from evolutionarily related proteins. Since homologous proteins often exhibit similar ligand recognition behaviors and conserved binding preferences, leveraging such cross-target information may provide complementary evidence for compound prioritization. Retrieval augmentation offers a natural framework for incorporating this external knowledge during inference. However, experimentally validated protein-ligand interactions are extremely sparse, and retrieved homologous proteins often lack transferable binding annotations for the query target, making effective retrieval augmentation challenging. To address this limitation, we propose BURA, an uncertainty-aware retrieval augmentation framework for virtual screening. Specifically, BURA retrieves evolutionarily related proteins and leverages their screening behaviors over the same candidate library as transferable biochemical evidence. To further enhance the reliability of retrieved evidence, we develop a task-specific uncertainty estimation strategy to assess screening reliability and adaptively weight retrieved biochemical evidence during virtual screening. Extensive experiments on three widely used virtual screening benchmarks, including LIT-PCBA, DUD-E, and DEKOIS, demonstrate that BURA consistently improves diverse virtual screening backbones without requiring additional annotations or model retraining. Further ablation studies and case analyses verify the effectiveness of both the retrieval augmentation strategy and the proposed uncertainty. Code will be publicly available.

Inducing cuproptosis for cancer therapy currently relies on supraphysiological copper or copper ionophores. Although serum copper is elevated in patients with prostate cancer, whether this is sufficient to trigger physiological cuproptosis remains unclear. Here, we show that tumor copper levels positively correlate with androgen receptor (AR) activity, and castration‐resistant prostate cancer (CRPC) with hyperactivated AR exhibits pathological copper accumulation. AR activation enhances copper uptake while simultaneously conferring tolerance to copper toxicity, creating a buffered copper state. This adaptive response is mediated by metal‐responsive transcription factor 1 (MTF1), which is transcriptionally upregulated by AR and undergoes EP300‐dependent lactylation of lysine 218, promoting copper‐induced nuclear translocation. Nuclear MTF1 activates metallothioneins (MT1E, MT1F, and MT1M) that sequester cytosolic copper and restrict mitochondrial copper accumulation. Disrupting MTF1 collapses this buffering system, enabling endogenous copper to trigger cuproptosis and suppress CRPC growth. These findings identify cuproptosis as a therapeutically exploitable vulnerability in CRPC.

In recent years, deep learning on protein structures has attracted widespread attention, as structures determine proteins' function. A series of structure-based protein property prediction methods have been proposed, achieving remarkable performance. However, these methods often neglect the importance of the protein size and fail to fully leverage it, leading to biases toward certain sizes and suboptimal overall performance. To address this issue, we propose a protein size-guided conditional mixture-of-experts for improving deep learning on protein structures. It can adaptively activate the sub-networks with the guidance of protein sizes and network features. Its flexible combinations of sub-networks help mitigate biases toward certain protein sizes, while the deliberate incorporation of protein size guidance enables the network to effectively capture both universal and size-specific characteristics, resulting in more accurate predictive performance. Based on it, we propose a framework for protein property prediction and benchmark it on eight tasks with two representation forms of proteins and three different dataset splits, a total of forty-eight tests. Experiments show that our method can be seamlessly integrated into numerous existing models and achieve performance improvement across tasks under almost all settings. More importantly, our experiments reveal that although often overlooked, protein size serves as an important prior knowledge in deep learning on protein structures.

Protein-ligand blind docking refers to predicting the docking conformation of a ligand with a complete protein, which is crucial for drug discovery. However, the large size of complete proteins, coupled with the complexity of their structures, makes this task highly challenging. To address these issues, we propose a novel blind docking framework, IAP-CFDock, which consists of three stages: Iterative Anchor Prediction, Coarse Docking, and Fine-Grained Docking. To enable high-accuracy anchor prediction, we design an efficient protein-ligand interaction network, termed the AP Layer. Experimental results show that the anchor prediction mechanism exhibits strong robustness, high accuracy, and flexibility. Comprehensive comparisons across multiple data sets and experimental settings demonstrate that IAP-CFDock achieves superior docking accuracy and generalization capability while maintaining high inference efficiency. Furthermore, we investigate the docking performance of various blind docking methods across different ligand sizes. Finally, extensive ablation studies validate the effectiveness of each component in IAP-CFDock and highlight its strong potential for further improvement.

Deep learning-based protein function analysis is rapidly advancing, but most methods rely heavily on labeled data, which is typically scarce due to the need for expensive wet lab experiments. Self-supervised learning (SSL) is a viable solution because it can improve the model performance on downstream tasks by pre-training on large-scale unlabeled data. However, existing SSL methods overlook the disparities in data distributions between pre-training and downstream tasks, which we show has a large influence on the effectiveness of SSL. To tackle this problem, we propose a new protein representation learning framework with a protein size prompt (ProSP). ProSP allows the pre-trained model to learn size-related information during pre-training and adapt to downstream tasks with different protein size distributions. Results of 48 experiments across eight downstream tasks, three data splits, and two data representations show that ProSP successfully improves the performance in all experiments with minimal modifications to the original model.

Deep learning has significantly advanced the development of point cloud registration. However, in recent years, some methods have relied on additional sensor information or complex network designs to improve registration performance, which incurs considerable computational overhead. These methods often struggle to strike a reasonable balance between computational cost and performance gains. To address this, we propose a plug-and-play orthogonal self-ensemble module designed to enhance registration performance with minimal additional overhead. Specifically, we design a novel ensemble learning strategy to mine the complementary information within the extracted features of previous methods. Unlike most ensemble learning methods, our method does not set multiple complex models for performance enhancement. Instead, it only cascades a lightweight dual-branch network after the features extracted by the original model to obtain two sets of features with more diversity. To further reduce redundancy between features and prevent the degradation of the dual-branch network, we introduce an orthogonal constraint that ensures the features output by the two branches are more complementary. Finally, by concatenating the two sets of complementary features, the final enhanced features are obtained. Compared to the original features, these enhanced features thoroughly exploit the internal information and exhibit greater distinctiveness, leading to improved registration performance. To validate the effectiveness of our method, we plug it into GeoTransformer, resulting in consistent performance improvements across 3DMatch, KITTI, and ModelNet40 datasets. Moreover, our method is compatible with other performance-enhancing methods. In conjunction with the overlap prior in PEAL, GeoTransformer achieves a new state-of-the-art performance.

Predicting the docking conformation of a ligand in the protein binding site (pocket), i.e., protein-ligand docking, is crucial for drug discovery. Traditional docking methods have a long inference time and low accuracy in blind docking (when the pocket is unknown). Recently, blind docking techniques based on deep learning have significantly improved inference efficiency and achieved good docking results. However, these methods often use the entire protein for docking, which makes it difficult to identify the correct pocket and results in poor generalization performance. In this study, we propose a two-stage docking paradigm, where pocket prediction is followed by pocket-based docking. Following this paradigm, we design a new blind docking method based on pocket prediction (PPDock). Through extensive experiments on benchmark data sets, our proposed PPDock outperforms existing methods in nearly all evaluation metrics, demonstrating strong docking accuracy, generalization ability, and efficiency.

Proteins can be represented in different data forms, including sequence, structure, and surface, each of which has unique advantages and certain limitations. It is promising to fuse the complementary information among them. In this work, we propose a framework called ProteinF3S for enzyme function prediction that fuses the complementary information across protein sequence, structure, and surface. To achieve more effective fusion, we propose a multi-scale bidirectional fusion strategy between protein structure and surface, in which the hierarchical features of a surface encoder and a structure encoder interact with each other bidirectionally. Based on these interactions, more distinctive features can be obtained. After that, we achieve further fusion by concatenating the sequence features with the features containing structure and surface information, so that better performance can be achieved. To validate our method, we conduct extensive experiments on tasks including enzyme reaction classification and enzyme commission number prediction. Our method achieves new state-of-the-art performance and shows that fusing different forms of data is effective in enzyme function prediction.

Self-supervised learning has achieved significant success in various fields such as point cloud detection and segmentation. However, self-supervised learning for point cloud registration is less explored. The recently proposed self-supervised learning framework MSC has paved the way for investigating self-supervised learning for registration. Based on this, we conduct a comprehensive exploration of applying self-supervised learning to point cloud registration in this study. Surprisingly, regardless of the proxy tasks used, self-supervised learning consistently results in negative performance. To uncover the underlying reasons, we conduct comprehensive experiments and analysis, including gradient-based analysis and proxy task adjustments. Moreover, we also attempt to use adapter-tuning to achieve further exploration both qualitatively and quantitatively. Upon these experiments and analysis, we find that the main obstacle stems from the inconsistent data distribution between pre-training and downstream task, which indicates that reducing the data distribution gap may be the core of future research. We hope that our work can offer valuable insights and inspiration for following studies on self-supervised learning and point cloud registration.