BioGeometry Unveils GeoFlow V2: World's First All-Atom General-Purpose Protein Foundation Model Unifying Structure Prediction and De Novo Design
BioGeometry
Large Molecule Drug Developer
'%20fill='%23FFFFFF'%3E%3Crect%20x='249'%20y='126'%20width='1'%20height='1'%3E%3C/rect%3E%3C/g%3E%3C/g%3E%3C/svg%3E)
On April 16, 2025, BioGeometry was officially released.GlobalThe First Full-ScenarioAtomic-level EggWhite Matter Large Model——GeoFlow V2。The ModelFirst to Unify Protein Structure Prediction and De Novo Design with Atomic-Level Precision, providing a one-stop solution for macromolecules of life from understanding to creation.
From AlphaFold initiating the protein structure prediction revolution, to RFDiffusion becoming the pinnacle of de novo protein design, and further toIntegrating Understanding Prediction with De Novo Design, the More Accurate, More Universal, and More Application-Oriented GeoFlow V2 is ReleasedScientists can now precisely generate proteins from scratch on specified epitopes across modal substrates such as proteins, small molecules, and DNA/RNA, bringing more possibilities to industrial-level protein design applications.From Zero to OneThe breakthrough.
“
The All-Purpose Engine for Protein Design:The Fourfold Breakthrough of GeoFlow V2
01
Unified Base
The First Universal Atomic-Level Large Model for Proteins
Integrated Architecture for Structure Prediction and De Novo Design
Compared with the unidirectional nature of mainstream models in the industry today, which "can only predict or only generate," GeoFlow V2 achieves the integration of protein structure prediction and de novo design for the first time: it can not only perform precise structural predictions of existing proteins but also build entirely new protein structures and sequences from scratch. It has achieved comprehensive advancements in performance metrics, prediction accuracy, and design rationality.
GeoFlow V2 Model Process Diagram
02
Precision Modeling
Intelligent Structure Prediction for Complex Systems
GeoFlow V2 has powerful capabilities for predicting and modeling protein complex structures, enabling precise docking structure predictions under various constraints. It is widely applicable to interaction systems such as antigen-antibody, protein-protein, and protein-small molecule. The Top-1 prediction success rate (the proportion of the most confident model with a DockQ > 0.23 compared to the real model, as shown in the left figure below) of GeoFlow V2 on low-homology antigen-antibody datasets reaches45%, significantly outperforming Chai-1 (33%), Protenix (31%), AFM2.3 (25%), and other models with the highest recognition in the current industry. In the protein-small molecule structure prediction task (right figure below), GeoFlow V2 also demonstrates top-tier performance.
GeoFlow V2 Achieves Leading Performance in Antigen-Antibody Complex and Protein-Small Molecule Complex Structure Prediction
Given the provision of one or more epitope amino acids (i.e., the amino acids on the antigen that bind to the antibody), the structural prediction success rate of GeoFlow V2 can be further improved to50% to 75%`, which reflects the model's precise modeling of atomic-level interactions and lays a solid foundation for epitope-based antibody design from scratch.`
Evaluation of Antigen-Antibody Complex Structure Prediction Performance Under Specified Epitope Conditions
In the antibody structure prediction task, the distilled small model GeoFlow V2-ab achieves a prediction speed comparable to AFM2.3 while maintaining higher accuracy.More than a hundred times, Predicted time compressed toSecond-level, contributing to the rapid inference of antibody structures at the million-scale.
GeoFlow V2-ab Achieves Leading Accuracy in Antibody Structure Prediction
GeoFlow V2-ab shows a 150 to 250 times speed advantage over AFM2.3 in antibody prediction runtime.
03
Innovative Design
A More Obedient and Efficient General Protein Design Paradigm
In protein design tasks represented by antibody design,GeoFlow V2 Demonstrates Industry-Leading Design Structural Rationality and Combination Success Rate, achieving a dual breakthrough in both generation quality and hit efficiency. The epitope compliance, CDR binding rate, and virtual screening pass rate of the structures generated by GeoFlow V2 are significantly superior to those of the SOTA model RFAntibody (fine-tuned from RFDiffusion on an antibody dataset). This means that, given the same number of designs, GeoFlow V2 can generate more functional candidate molecules, greatly improving the success rate of protein design projects.
Antibody Design and Validation Workflow Based on GeoFlow V2
On the antibody design test set composed of seven targets, 1000 antibodies were generated for the specified epitope amino acids, with the antibody framework region unified asUnder the framework of Trastuzumab, GeoFlow V2 successfully reproduced the binding mode of the reference molecule on five antigens (minimum FR RMSD < 5Å); whereas RFAntibody failed to accurately reproduce the binding mode of the reference molecule on all seven targets, only achieving antibody structures with < 10Å on two targets.
04
Cross-Modal Empowerment
Adapt to Multi-source Inputs and Diverse Tasks
GeoFlow V2 has full-stack multi-modal protein design capabilities, enabling highly flexible matching with input targets across different types of tasks. Whether based onProtein Sequence or StructureThe optimized design, orSmall molecules, nucleic acidsDe novo design or scaffolding of binding proteins for input, or direct generation of high stabilityMorePeptidePrefaceColumnGeoFlow V2 can consistently generate structurally sound and functionally clear candidate designs, demonstrating its broad adaptability in drug discovery, molecular engineering, and synthetic biology.
“
Actual application scenario: EnableThe Freedom Gateway of Life Molecular Design
Previous protein discovery and design either relied on the complex systems evolved by nature (such as animal immunity for antibody discovery), or on high-throughput experiments for large-scale trial and error (such as saturation mutagenesis of existing molecules), or on rational design based on the experience of domain experts.
GeoFlow V2 Pushes Protein Design Forward with Reliable and Versatile Model CapabilitiesDefinable, Optimizable, Scalable for Mass ProductionThe intelligent era is bringing substantial restructuring to research and development efficiency and the boundaries of innovation.
In a variety of real-world projects, the GeoFlow series of models have already demonstrated their capabilities and are gradually becomingThe Key Infrastructure for Unlocking "Design Freedom" in Proteins。
01
Antigen Design
Solubilization Modification of Transmembrane Proteins
In collaboration with a leading pharmaceutical company, GeoFlow V2 successfully achieved the modeling of a multi-transmembrane protein.Solubilization TransformationThe team used GeoFlow V2 to redesign the transmembrane hydrophobic region, successfully constructing a transmembrane protein sequence that can stably exist in aqueous solution while maintaining the structure and functional characteristics of the extracellular region.90% can achieve stable soluble expression, where multiple sequences have the same binding specificity as the extracellular region of the original protein, but only the transmembrane region has22% sequence similarity, demonstrating strong design innovation and engineering adaptability.
02
De Novo Antibody Design
Targeting "Undruggable" Epitopes
In the de novo antibody design tasks targeting HER2 (ECD2/ECD4 epitopes) and multiple internal/collaborative targets, several antibodies generated by GeoFlow V2 were identified after high-throughput library screening.Low nanomolar affinity, meets epitope binding requirements, and has good developability.The full anti/nanobody molecule,Some molecules have completed functional verification in animal models.
Compared with traditional animal immunization and library screening methods, GeoFlow has opened up a completely new path for antibody discovery, allowing scientists toNo longer rely on the power of the natural life immune system for antibody discovery, but by using generative AI for precise modeling of epitope targets to generate binding antibodies, it is expected to bring revolutionary changes to the discovery of difficult-to-drug targets, special epitopes, and cross-species antibodies.
03
Functional Protein Design
Quickly Obtain High-Performance Patentable Molecules
In aIn the task of redesigning functional proteins, GeoFlow V2 achieves a patentable redesign of the target binding region (loop area), generating 100 candidate molecules in one round, with80% can be expressed normally and have good thermal stability., where50% of the samples showed even better affinity for the target peptide than the patented molecule., reaching up to tens of pM levels.
Compared with the original patented molecule requiring 5 rounds of library construction and experimental iteration, GeoFlow V2 significantly shortened the R&D cycle, demonstrating its capability inIntellectual Property Avoidance, Protein Performance Optimization, and Development Efficiency ImprovementThree powerful capabilities.
04
Industrial Enzyme Optimization
A 55-Day Achievement of Multi-Objective Performance Leap in Transaminase
In the optimization of a key pharmaceutical intermediate synthesis process, GeoFlow V2 was usedHigh-Performance Directed Evolution of Transaminases, to solve the catalytic challenges of key chiral precursor amines. Through in-depth modeling of enzyme-substrate relationships and variant design, combined with high-throughput experimental platforms, only55 Days, BioGeometry has completed the comprehensive optimization of enzyme performance - catalytic activity enhanced.52.3 times, Left-handed selectivityFrom 0% to 99.7%, all key indicators meet the requirements of the pilot-scale process.
05
Novel Enzyme Mining for Industrial Enzymes
Design a brand-new synthetic pathway
A certain non-natural amino acid has significant commercial value, but the conversion rate of traditional synthesis methods cannot meet industrial requirements.
BioGeometry used GeoFlow to develop an enzyme capable of catalyzing novel reactions with commercial cost advantages through new enzyme mining and two rounds of enzyme evolution. They innovatively designed a one-step enzymatic in vitro biosynthesis route. This route uses low-cost raw materials and has a high conversion rate, offering superior cost benefits.
The success of this project highlights GeoFlow's strong advantages from new enzyme discovery to enzyme optimization — GeoFlow not only successfully discovered this enzyme capable of catalyzing novel reactions (an enzyme able to catalyze such reactionsNo literature reports have been seen before.), also enhanced the catalytic activity of the enzyme through two rounds of enzyme evolution.Increased by 65 timesThanks to the excellent capabilities of GeoFlow and the high-throughput wet lab capabilities built by BioGeometry, all of this was completed in just 2 months.
“
From Tools to Paradigms: Building the Underlying Capabilities for Next-Generation Protein Design, Inviting You to Co-Create
The significance of GeoFlow V2 goes beyond the improvement in model accuracy and generation capabilities. It offers a new solution for the entire industry — structural prediction and de novo design have, for the first time, evolved collaboratively within a unified architecture, integrating into a system framework that spans from fundamental research to application development, making proteins intelligent components that can be understood, created, and validated.
Design freedom, prediction accuracy, optimization efficiency — The mission of GeoFlow V2 is not to solve a local problem, but to redefine rules and unleash productivity across the entire system engineering of protein design.
The Next Generation of Protein, Inviting You to Co-create.
Starting from todayGeoFlow V2 Officially Released Worldwide to Biomedical, Synthetic Biology, and Protein Engineering Research Institutions, Industry Experts, and Innovation TeamsOpen Limited-Time Free Trial. You can log in at any time[https://prot.design]Start the experience. For in-depth cooperation or customized needs, please feel free to contact us.[bd@biogeom.com]。
We sincerely hope to collaborate with researchers and engineers at the forefront of breakthroughs in protein structure and functional innovation to advance the future progress of protein design.