BioGeometry
BioGeometry builds generative AI models for protein design and uses them in partner projects and its own drug pipeline. GeoFlow, its all atom model for biomolecules, unifies sequence, structure and function in one diffusion framework: version 1 in June 2024 matched AlphaFold 3 on protein complex structure prediction, version 2 in April 2025 added de novo design, and version 3 in October 2025 added a generate, evaluate and optimize loop that mimics antibody affinity maturation. The company reports a hit rate near 20% across more than 20 targets with version 3 and lead discovery in under three weeks, with a fourth version in development. GeoCell, a second model, works at the cell level, modeling cell states and their response to perturbation, and high throughput automated laboratories test the designs.
The work spans epitope specific antibody generation and optimization in biomedicine and enzyme engineering in synthetic biology. Vazyme used GeoFlow V2 to optimize an enzyme for an in vitro diagnostic reagent and reached more than 21 times the original yield in 60 days, Sino Biological pairs its protein expression labs with the platform, and Nanning Harworld Biology signed an industrial biotechnology collaboration. The internal pipeline covers autoimmune disease, neurology and oncology, with bispecific antibody programs expected to reach investigational filing in 2027.
Jian Tang founded BioGeometry in 2022, with Yoshua Bengio as chief scientific advisor. It is headquartered in Beijing with a research center in Shanghai. A Series A+ round led by China Life Capital in October 2026 took its Series A financing past $100 million. GeoFlow Server offers free research access under terms that bar patent claims on outputs without written permission, and commercial projects are arranged through the business team.
Capability Axes
An AI Health Index grade measures what a buyer can verify from public sources on the date shown. It is not a rating of how good the product is. A vendor can build an excellent system and grade low on an axis because it publishes nothing an outsider can check. How grades read
GeoFlow is a generative model that proposes candidate proteins and evaluates properties of interest before lab testing, and for an antibody program its proposals provide the starting set of candidates scientists then test. The company applies multiple model generations in commercial research and in its own pipeline, and the October 2026 financing is aimed at GeoFlow, GeoCell and the automated laboratories that test their output. Its Vazyme work relied on a named model version. Laboratory testing evaluates the output, but generation and computational assessment, the steps before any experiment, come from the AI, so the product does not exist without the models.
A research sponsor sets program objectives, and the company uses GeoFlow to generate and assess candidates, then tests their properties in its laboratories under the project's conditions. The research server returns model confidence with its predictions and generations, giving users a signal for deciding what to examine, and the terms ask for discretion before relying on outputs. V3's loop of generating, evaluating and optimizing runs inside the model without a person between rounds.
No one is named as approving a prediction, no confidence threshold changes the next step, and nothing says what happens when model and bench results differ. Decisions to move a candidate toward the clinic stay with the development team.
The GeoFlow family unifies protein sequence, structure and function in an all atom diffusion model, with versions dated and described: V1 in June 2024 for structure prediction, V2 in April 2025 adding de novo design, V3 in October 2025 adding a multistep generate, evaluate and optimize loop, and V4 scaling both parameters and training data. A technical report accompanies V3.
V2 is named in the Vazyme work and V3 in the antibody design results, so a customer can tell which generation produced each example. Training data is not characterized, and purchasing terms give no model pinning rights, notification schedule or change management during a long research program.
The GeoFlow family is built in house and used alongside the company's own laboratory work, with GeoCell as a separate cell level model. Team members have also worked on outside research models, which the company keeps distinct from its own platform, and the research server terms identify the company's proprietary technology as the engine. That is a structurally short chain at the model layer.
There is no inventory of suppliers for the commercial service, no named infrastructure operator and no partner choice of model provider. The first party model and version can be tied to a named engagement, as GeoFlow V2 is to Vazyme.
BioGeometry reports that its Vazyme project reached more than 21 times the original yield for a diagnostic reagent enzyme under the tested conditions, in 60 days with GeoFlow V2. For antibodies, the company reports that GeoFlow V3 reached an average hit rate near 20% across more than 20 targets, close to a hundredfold gain over earlier AI methods, and cut lead discovery to under three weeks, with a technical report behind the model. Those are research and manufacturing results with a named partner and named model versions. No candidate from the platform has yet entered a clinical study, so measured patient outcomes do not exist.
The GeoFlow Server terms prohibit specified harmful uses, including help synthesizing illegal substances and attempts to override safety filters, and they forbid using outputs to train competing structure prediction or design models. Users keep ownership of their content and original outputs, and the company will not publish either without approval, while the server may give several users the same or similar output. These commitments apply to the research service.
Retention and deletion periods, processing locations, whether sponsor submissions train GeoFlow and incident response for negotiated programs are not stated, and the terms carry no workflow for patient records because the product works on molecules.
The buyer brings a drug or reagent research program, and the commercial offer centers on molecule design and laboratory results. The GeoFlow Server terms govern research access and set rights in submissions and outputs. There is no statement that the service accepts protected health information, acts as a business associate or offers a business associate agreement.
Research programs may concern a disease and use experimental data without identifying any patient, and the terms forbid using outputs for healthcare decisions. What would govern an engagement involving identifiable patient data, and the scope of any healthcare privacy agreement, are not described.
GeoFlow combines a hosted research server with work inside the company's own research programs, and the server terms cover use restrictions and the company's rights in its models. No security certification, audit period, certifying organization, report request process or trust center is named.
The Beijing and Shanghai locations show where the company has facilities, not where research accounts are processed, and the site footer carries the internet content filing and public security registration a Beijing website needs. Security controls and attestations for a partner program are left to its agreement, which matters for sponsors sending unpublished sequences.
The software and research partnerships support discovery and development of candidates in preclinical settings, and the pipeline lists stages from hits and leads through preclinical candidates to investigational filing preparation, with bispecific antibodies expected to file in 2027. Those are development milestones for candidate medicines. The research server terms state that outputs are for theoretical modeling only and are not intended, validated or approved for clinical use.
Vazyme's case concerns a component of an in vitro diagnostic reagent and does not bring FDA clearance to GeoFlow or to a finished test. Regulatory status for any drug or test rests with that product and its sponsor.
GeoFlow is a named model series whose versions are distinguished by date and capability, and the company pairs computational work with laboratory assessment. Research server terms define allowed uses, forbid attempts to circumvent safety filters and bar outputs from use in decisions that affect individual rights or wellbeing, such as healthcare. Those are controls on research access.
There is no production model review authority for sponsor projects, no breakdown of how performance varies across target classes or protein families, and no stated point at which a model update would be stopped or rolled back, so oversight for a given research program sits in its agreement.
A sponsor receives lab results that test whether a proposed candidate has the properties the program seeks, so the scientific check sits inside the research workflow, separate from any remedy for an incorrect output. The research server terms tell users to interpret predictions carefully, warn that outputs may already be patented and place the burden of avoiding infringement on the user. Claims under the terms can be brought within two years of a party becoming aware of the facts. There is no promise to correct research output within a set period or to compensate a sponsor for a failed hypothesis, and commitments for commissioned work rest on negotiated terms.
No clinical record integration exists, which fits a protein design service working upstream of any patient, so the question is how results reach a sponsor's research systems. Researchers use GeoFlow through the hosted server, which accepts PDB and UniProt identifiers or an uploaded structure, or through a direct commercial collaboration.
The Vazyme case describes delivering project results to the partner but names no transfer software, record format or connector to the sponsor's laboratory systems. A sponsor receives research data or results under its project agreement, with no standard interface into its laboratory information or research data platforms.
The research server is hosted, and commercial projects run on experimental facilities the company operates, with headquarters in Beijing and a research center in Shanghai. The Vazyme agreement describes contracted project work rather than a deployment into Vazyme's computing environment. A lab address does not show where submissions, outputs or model processing reside.
No storage or computing region is committed, no option to run the model in a partner's own environment is offered, and how customer material is kept separate across hosted programs is not described. For a sponsor outside China, where its sequences are processed is left for each project agreement to settle.
GeoFlow Server is free for research under terms that bar patent, copyright or trademark claims on outputs without written permission from the company, which marks a defined research route and a clear line where commercial rights begin. Commercial design and laboratory work is negotiated through the business development team. The Vazyme and Sino Biological collaborations describe projects and results without prices.
There is no public rate per project, candidate or laboratory service, and no stated cost for the commercial rights the research terms withhold. Project prices are set around scientific scope and laboratory deliverables, separately from access to the hosted research service.
Drug developers use the protein design platform in preclinical research, and the company runs its own pipeline in oncology, autoimmune disease and neurology, with bispecific antibodies the most advanced. Biomedicine work centers on epitope specific antibody generation and optimization, including single domain antibody design on the research server.
Synthetic biology is the second setting: the Vazyme project optimized a diagnostic reagent enzyme, Nanning Harworld Biology signed an industrial biotechnology collaboration, and the company is working toward commercial scale natural borneol production. That gives the model a practical range across research and manufacturing. None of it reaches clinical populations or use by care teams.
Alternatives to BioGeometry
Other products that do the same kind of work as BioGeometry, and what each one does.
immunitoAI designs antibody drugs with AI and tests the designs in its own laboratory.
Earendil Labs uses AI to design antibody drugs and develops them through its own pipeline, licensing individual programs to large pharmaceutical companies.
Medvolt, based in Pune, India, sells MedGraph, an AI and physics based in silico platform for early drug discovery, to pharmaceutical and biotechnology companies as software modules and as delivered projects.
Antiverse designs antibodies with generative AI for G protein coupled receptors, ion channels and other membrane proteins that conventional antibody discovery struggles to reach, and sells that work to drug companies through research and licensing partnerships.
Isomorphic Labs is a London based AI drug design company founded in 2021 by Sir Demis Hassabis, who leads it as chief executive.
San Francisco company that licenses foundation models for molecular design to pharmaceutical and life sciences research organizations, rather than using them to build its own pipeline.
Pricing
Vendor-published figures are labeled as such. Figures labeled “Estimated” are derived from third-party sources and have not been confirmed by the vendor.
| Entry Price | Pricing Basis | BAA Tier | Implementation | Source |
|---|---|---|---|---|
|
Free research access. Contact for commercial pricing
|
Free restricted research service and separately negotiated research collaborations | No business associate agreement is stated. The product works on molecules, not patient records | No setup or project fee is listed. | Vendor Published |
GeoFlow Server is free for research under terms that bar patent claims on outputs without written permission. BioGeometry prices commissioned research through its business team. Project rates and commercial licensing fees are not listed.