Xiangzhe Kong

Xiangzhe Kong

Auto-generated profile, based on publicly available information

Xiangzhe Kong

Department of Computer Science and Technology, Tsinghua University, Beijing, China

Research Focus & Metrics

Academic Career Length

2Years (2025-Present)

(Calculated once Scilit publication data is synced: current year − year of first publication + 1)

Representative Publications

View All Publications
1 October 2025 in Nature Biomedical Engineering
Verified
Journal Article
Closed access
PDF file
XKXiangzhe Kong
Xiangzhe Kong
RJRui Jiao
Rui Jiao
HLHaowei Lin
Haowei Lin
RGRuihan Guo
Ruihan Guo
WHWenbing Huang
Wenbing Huang
AMand Ma
and Ma

10 February 2026 in Nature Communications
Verified
Journal Article
Open access
PDF file
RJRui Jiao
Rui Jiao
XKXiangzhe Kong
Xiangzhe Kong
LZLi Zhang
Li Zhang
ZYZiyang Yu
Ziyang Yu
FRFangyuan Ren
Fangyuan Ren
YLYang Liu
Yang Liu
Pretraining on a large number of unlabeled 3D molecules has showcased superiority in various scientific applications. However, prior efforts typically focus on pretraining models in a specific domain, missing the opportunity to leverage cross-domain knowledge. To mitigate this gap, we introduce Equivariant Pretrained Transformer, an all-atom foundation model that can be pretrained from multiple domain 3D molecules. Built upon an E(3)-equivariant transformer, the model learns both atom-level interactions and graph-level structural features ( e.g . residuals in proteins), allowing it to generalize across diverse tasks. The model achieves strong gains in ligand binding affinity prediction, while also performing competitively in predicting properties of proteins and small molecules. We further show that the model can help identify potential antiviral compounds against the main protease of the COVID-19 virus, and validate promising candidates through computational and experimental studies.

15 March 2026 in bioRxiv
Verified
Preprint
Open access
PDF file
XKXiangzhe Kong
Xiangzhe Kong
JCJunwei Chen
Junwei Chen
ZZZiting Zhang
Ziting Zhang
GLGaodeng Li
Gaodeng Li
QZQingyuan Zhu
Qingyuan Zhu
AMand Ma
and Ma
Biomolecular interactions lie at the core of cellular life, spanning diverse molecular modalities from small molecules to nucleic acids and proteins. Nevertheless, design strategies remain separated despite shared physicochemical principles of molecular recognition. Here we present AnewOmni, a unified generative framework trained on more than 5 million biomolecular complexes, that enables transferable molecular design across molecular scales by assembling chemically meaningful building blocks at atomic resolution. We further introduce programmable graph prompts to support user-defined chemical, topological, and geometric steering during generation, exploring hybrid and unconventional chemistries beyond canonical structures. We demonstrate that transferable learning of interaction patterns and physical constraints across molecular modalities is possible, via an atom-to-block latent space capturing both atomic details and structural priors. The framework successfully designed small molecules, peptides, and nanobodies targeting the challenging KRAS G12D switch II pocket, as well as orthosteric peptides and allosteric small-molecule inhibitors for PCSK9 in the absence of known binding site, achieving 23%-75% success with only low-throughput validation, bypassing modality-specific high-throughput screening. AnewOmni is the first to succeed in functional molecular design across all scales, from small chemical entities to large biologics, and represents a stepstone towards general molecular reasoning engines, advocating a generative foundation model for biomolecular interactions to enter regimes where data and human intuition remain limited.

21 January 2025 in bioRxiv
Verified
Preprint
Open access
PDF file
XKXiangzhe Kong
Xiangzhe Kong
RJRui Jiao
Rui Jiao
HLHaowei Lin
Haowei Lin
RGRuihan Guo
Ruihan Guo
WHWenbin Huang
Wenbin Huang
AMand Ma
and Ma
Peptides offer distinct advantages for targeted therapy, including oral bioavailability, cellular permeability, and high specificity, which set them apart from conventional small molecules and biologics. In this work, we developed an AI algorithm, named PepMimic, to transform a known protein receptor or an existing antibody of a target into a short peptide drug by mimicking the binding interfaces between targets and known binders. The structural root mean square deviation and interface DockQ with reference binders were 61% and 75% better than the best existing methods on the PepBench datasets. We then applied this novel peptide-design methodology to five drug targets: PD-L1, CD38, BCMA, HER2, and CD4. SPRi results show that 8% of the peptides exhibited dissociation constant (K D ) values at the 10 -8 M level, and 26 peptides achieving K D values as low as 10 -9 M. This success rate was 20,000 times higher than that observed in a random library screening conducted under identical conditions. PepMimic was applied to target proteins lacking available binders by first utilizing AI algorithms to design protein binders, followed by the generation of peptides through simulation of these artificial interfaces. The top-ranked peptides underwent extensive cellular validation and in vivo testing through tail vein injections in breast, myeloma, and lung tumor mouse models. Experimental results demonstrated effective membrane binding and highlighted the strong potential of these peptides for clinical diagnostic imaging and targeted therapeutic applications.

27 October 2025 in Nature Biomedical Engineering
Verified
Journal Article
Open access
PDF file
XKXiangzhe Kong
Xiangzhe Kong
RJRui Jiao
Rui Jiao
HLHaowei Lin
Haowei Lin
RGRuihan Guo
Ruihan Guo
WHWenbing Huang
Wenbing Huang
AMand Ma
and Ma

8 November 2025 in bioRxiv
Verified
Preprint
Open access
PZPeidong Zhang
Peidong Zhang
RHRong Han
Rong Han
XKXiangzhe Kong
Xiangzhe Kong
TCTing Chen
Ting Chen
JMJianzhu Ma
Jianzhu Ma
Structure-based generative models often optimize single-target affinity with ignorance of specificity, resulting in the generation of high-affinity candidates that exhibit promiscuous binding across unrelated targets. This decoupling of affinity and specificity not only compromises therapeutic efficacy but also elevates off-target risks that constrain translational potential. Therefore, we introduce SpecLig, a unified structure-based framework that jointly generates small molecules and peptides with improved target affinity and specificity. SpecLig represents a complex as a block-based graph, combining a hierarchical SE(3)-equivariant variational autoencoder with an energy-guided geometric latent-diffusion model. Chemical priors derived from block–block contact statistics are explicitly incorporated, biasing generation towards pocket-complementary fragment combinations. We benchmark SpecLig on peptide and small-molecule tasks using standard public datasets and propose precision/breadth testing paradigms to quantify specificity. Across multiple evaluations, ligand candidates generated by SpecLig usually bind to the target pocket with high specificity and affinity while maintaining competitive advantages in other attributes. Ablations indicate that both hierarchical representation and energy guidance contribute to success. Finally, we present multiple real applications that demonstrate how SpecLig improves ligands in natural complexes to mitigate potential off-target risks. SpecLig, therefore, provides a practical route to prioritize higher-specificity designs for downstream experimental validation. The codes are available at: https://github.com/CQ-zhang-2016/SpecLig .