Galixir Unveils TBind: A Groundbreaking AI Model for Simultaneous Prediction of Drug-Protein Binding Structures and Affinities, Paving the Way Post-AlphaFold Era

Following the groundbreaking releases of AlphaFold in 2018 and 2020, protein structure prediction has approached experimental accuracy, laying a solid foundation for computation-driven drug discovery. In the lengthy process of drug development, predicting molecular interactions with proteins is considered the next "holy grail" for computational advancements in this field. Recently, Galixir pre-released TBind: Trigonometry Aware Neural NetworK for Drug-Protein Binding Structure Prediction.This model is the world's first deep representation learning framework capable of simultaneously predicting the three-dimensional binding conformation and binding affinity of small molecules and target proteins, significantly surpassingExistingThe best result of the method.TBind adopts an end-to-end data-driven paradigm, combined with physics-inspired geometric graph neural networks, achieving dual predictions of complex three-dimensional binding modes and binding strength. It surpasses the accuracy and efficiency of international commercial molecular docking software, providing the first breakthrough solution in China for molecular-protein interaction prediction. Following AlphaFold, it ushers computational-driven drug discovery into a new era.

Drug discovery is an extremely challenging task. In the vast chemical space (approximately 10^60 drug-like molecules), only a small fraction can bind to specific biological targets and produce therapeutic effects. Currently, most drugs target proteins, treating diseases by designing small molecule compounds that interact with them. Therefore, discovering small molecule compounds that can interact with protein molecules and elucidating their binding modes with the target protein are crucial for new drug development.

In order to address the aforementioned pain points and empower new drug development, following AprilJointly released the structure-based deep affinity prediction model STAMP-DPI with AstraZeneca's global R&D center.Later, Galixir, in collaboration with researchers from Fudan University and Sun Yat-sen University, released the latest Trigonometry Aware Neural NetworK (TANK) based on a three-body deep neural network.TBind v1.0.1, Specializing in 3D Structure Prediction of Small Molecule Ligand-Protein Complexes.

Inspired by the "Triangle Multiplicative Update" architecture within amino acids in AlphaFold2, TBind organically extends this module to intermolecular interactions between small molecules and target proteins, with multiple upgrades. These improvements enable the model to break through the limitations of traditional intermolecular force fields, granting it the ability to directly fit many-body effects without significantly increasing the model's complexity. Building on the three-body neural network module for intermolecular interactions, TBind also independently developed a protein block-based technology rooted in contrastive learning and the divide-and-conquer concept. This approach focuses separately on structural and functional regions of proteins, extracting local information from conserved regions and achieving implicit data augmentation under structural data. The research team also proposed the max-margin contrastive affinity loss function to drive the model to make full use of affinity information and global 3D structural information.The trade-off between local and global information greatly enhances the accuracy and generalization performance of TBind, enabling it to make rapid and effective predictions for novel protein pockets and new binding modes.

TBind Model Schematic Diagram. The input of the model is a protein 3D structure and a molecule 3D structure, and the output is the binding mode and binding strength between the two.

On the industry-standard test set PDBBind, TBind's performance significantly surpasses the current state-of-the-art deep learning method (EQUIBIND, developed by MIT's Tommi Jaakkola group, ICML 2022[1]) as well as multiple international commercial and academic docking software programs (including GLIDE, VINA, SMINA, GINA, etc.). TBind was trained on 17,787 three-dimensional structures of small molecule complexes published before 2020. In the task of predicting three-dimensional complex structures formed by 142 new proteins released after 2020 that were not seen in the training set,TBind increases the proportion of LigandRMSD less than 5Å from approximately 30% to 56%.[2]; The proportion of predicted binding centers with a distance less than 5Å to the true center has increased from 48% to 76%.

In the PDBBind new protein test set, the proportion of predicted structures with RMSD less than 5Å compared to the real co-crystal structures. TBind significantly outperforms other models.

In the PDBBind new protein test set, the proportion of predicted ligand centroids with a distance to the true centroid of less than 5Å. TBind significantly outperforms other models.

Since the model abandons cumbersome traditional sampling methods and utilizes data-driven AI potential energy surfaces for structure generation, the efficiency of prediction and screening has been significantly improved.In the global docking task, each molecule only takes 0.5 seconds.，It is one-four-hundredth of the academic software VINA and one-two-thousandth of the commercial software GLIDE.

The time to complete a docking scoring. TBIND only needs 0.5 seconds to complete the prediction, significantly faster than traditional docking methods.

The research team has currently released on GitHubFreeOpen-source test version TBind v0.5.0 has been released, providingCase Showcase, scan the QR code to redirect to GitHub.

TBind Commercial Version v1.0.1 Has Been Deployed on Galixir's Next-Generation Intelligent Computing Platform M1, Capable of Completing Ultra-High-Throughput Virtual Drug Screening at the Hundred Million Level in a Short Time, Empowering Multiple Key Stages of Drug Development Such as Hit Compound Discovery and Lead Compound Optimization.Business partners please contact m1@galixir.com for more usage information.

The TBind method can not only be applied in the field of small molecule and protein binding but also generalized to intermolecular interaction problems such as protein complex binding and nucleic acid-protein binding. Galixir will continue to maintain its innovative spirit, deeply integrate AI technology with drug discovery practices, continuously improve prediction accuracy and speed, and support more application scenarios.

Release of TBind v1.0.1Marks that Galixir has reached the international top level in small molecule protein binding prediction capability.As a key link in small molecule drug design, TBind has been organically integrated with Galixir's protein structure modeling algorithms, molecular design algorithms, property prediction algorithms, retrosynthesis analysis algorithms, and the intelligent computing platform M1. Together with biological experimental platforms, it forms a new "AI-computation-experiment iteration"三位一体 drug R&D paradigm. Galixir will continue to focus on differentiated pipelines that bring more benefits to the pharmaceutical industry, especially projects targeting undruggable or hard-to-drug targets, efficiently and accurately predicting molecular interactions, expanding the imaginative space of traditional chemistry, and exploring more novel drug molecules. Let the world be free from pain, and let new drugs be within reach.

About Galixir

Galixir, founded in 2019, is a company that leverages cutting-edge artificial intelligence technology to empower drug discovery. By utilizing advanced AI algorithms and integrating tools and expertise from computational chemistry, medicinal chemistry, and biology, Galixir addresses complex challenges in the early stages of small-molecule drug discovery, rapidly identifying candidate molecules with high activity, good drug-like properties, and novel structures. Galixir is collaborating with pharmaceutical companies and research institutions both in China and internationally to advance multiple drug discovery pipelines, covering various disease areas such as central nervous system disorders, autoimmune diseases, oncology, and respiratory conditions. Galixir significantly reduces the cost and time required for preclinical drug discovery pipelines, making it possible to maintain multiple drug discovery programs simultaneously and optimize overall strategic planning.