Q: Can Orbion really predict which mutations stabilize a protein?

Orbion is strongest at the honest and useful version of this problem: removing the mutations that will hurt you. The models identify destabilizing mutations with about 91% accuracy, and rank mutations by direction of effect with about 83% directional accuracy on GPCR targets. This lets you cut destabilizers early and focus bench time on the smaller set of promising candidates. We do not claim to design guaranteed stabilizers from a screen.

Q: Has this been validated with real wet-lab data?

Yes, in part, and more is in progress. Our first public wet-lab study is on bioRxiv: a prospective, head-to-head test of prediction-guided construct design on FGF21, run at an independent lab, with every platform score locked in before any protein was made. The platform's construct held more usable protein at higher concentration and stayed more monodisperse than the textbook design, and it predicted this in advance. We report the honest envelope with it: it is a single scoped case study (n=1), not a population benchmark, and the textbook construct actually won on raw yield, which we state first because it is true. Separately, a prospective panel with a validation partner confirmed our yield triage ranks constructs better than the prior scoring approach. Further wet-lab studies with partners are running now and will publish as they complete. We show data rather than claims, and we are candid about where the evidence is still thin.

Question 1

What is Orbion?

Accepted Answer

Orbion is a cloud software platform that tells protein scientists what is worth making before they spend months at the bench. You submit a protein sequence and get a full computational characterization in minutes, then design and score the construct most likely to express and behave, and stabilize hard targets. The platform works across protein classes for construct and yield decisions, with deep specialization in membrane proteins and GPCRs where stability is the wall.

Question 2

Is Orbion a CRO or a wet lab?

Accepted Answer

Orbion is a fully computational company. It runs no wet lab and no experiments. The product is software that you access through a web application and run on your own protein targets. This means no sample shipping, no queue, and no experimental turnaround time.

Question 3

Who is Orbion for?

Accepted Answer

Orbion is built for structural-biology and protein-engineering teams at CROs and biotech and pharma companies. CROs use it to characterize a client target before the next meeting. Biotech and pharma teams use it to shortlist variants and constructs before committing bench time. The platform fits teams that work on hard-to-express or hard-to-stabilize proteins.

Question 4

What problem does Orbion solve?

Accepted Answer

Orbion addresses two expensive bottlenecks that sit upstream of the biology. First, for protein production: choosing the construct that actually expresses well and stays well-behaved (the boundary, tag, signal peptide, and host) is hard, and getting it wrong costs weeks of make-test-redesign for low yield or poor developability. Second, for hard targets such as membrane proteins and GPCRs: low stability blocks the campaign before it starts. Orbion moves both decisions before the bench. You score and rank constructs for yield and developability, and screen mutations to remove the ones that would destabilize a target, so you take only the strongest candidates forward and waste fewer cycles.

Question 5

Who is behind Orbion?

Accepted Answer

Orbion pairs AI and product operators with deep protein-science expertise. The founding team combines software and machine-learning experience from companies such as Blue Yonder and GitHub with biology PhDs from Oxford, the MDC, and Charité. The company also works with academic partners including Columbia and Charité. This structure is deliberate: the core breakthrough is a data and machine-learning problem, supported by strong domain depth.

Question 6

Is the whole platform available now, or are parts still in development?

Accepted Answer

The core is live and in use by customers today: characterization, construct design and scoring, and protocol generation all run in production on your own proteins. Some capabilities, such as the productized Mutation Engine and customer calibration, are on a published roadmap. In an evaluation we show you what is live on your targets, not slideware.

Question 7

Where is Orbion based, and is my data kept in the EU?

Accepted Answer

Orbion is a German company (Orbion GmbH). Customer data is processed and stored in EU cloud regions (West Europe and Germany West Central). Proprietary model files may be processed in other regions under trade-secret controls, but customer data stays in the EU.

Question 8

How accurate are Orbion's predictions?

Accepted Answer

Orbion's models are benchmarked on membrane-protein and GPCR stability, where general tools perform poorly. On this class of target, Orbion's stability predictions reach a correlation of r = 0.655 with measured outcomes, against r < 0.4 for FoldX and Rosetta on the same public benchmark. Full benchmark reports by protein family are available, and more are on request.

Question 9

Can Orbion really predict which mutations stabilize a protein?

Accepted Answer

Orbion is strongest at the honest and useful version of this problem: removing the mutations that will hurt you. The models identify destabilizing mutations with about 91% accuracy, and rank mutations by direction of effect with about 83% directional accuracy on GPCR targets. This lets you cut destabilizers early and focus bench time on the smaller set of promising candidates. We do not claim to design guaranteed stabilizers from a screen.

Question 10

Has this been validated with real wet-lab data?

Accepted Answer

Yes, in part, and more is in progress. Our first public wet-lab study is on bioRxiv: a prospective, head-to-head test of prediction-guided construct design on FGF21, run at an independent lab, with every platform score locked in before any protein was made. The platform's construct held more usable protein at higher concentration and stayed more monodisperse than the textbook design, and it predicted this in advance. We report the honest envelope with it: it is a single scoped case study (n=1), not a population benchmark, and the textbook construct actually won on raw yield, which we state first because it is true. Separately, a prospective panel with a validation partner confirmed our yield triage ranks constructs better than the prior scoring approach. Further wet-lab studies with partners are running now and will publish as they complete. We show data rather than claims, and we are candid about where the evidence is still thin.

Question 11

Do you have results on a protein like mine, or on novel targets with little literature?

Accepted Answer

Orbion predicts from sequence, so it works on novel targets that have little or no published literature, and this is where it is most useful. We have benchmarked across thousands of GPCR and membrane-protein mutations and publish results by protein family, with more on request. During an evaluation we run your specific proteins in advance, so you judge the platform on your own targets rather than on a generic demo.

Question 12

Does Orbion just repeat what is already in the literature?

Accepted Answer

No. The models make their own predictions from sequence, rather than looking up known answers, so they produce results on proteins with no prior data. Where literature evidence exists, we use it to benchmark and sanity-check. The predictions that scientists find most valuable are the non-obvious ones a review of the literature would not have surfaced.

Question 13

How should I read the confidence scores?

Accepted Answer

Every prediction carries a confidence signal, and low-confidence predictions are flagged so you know which results to trust and which to verify. The intent is honesty: the platform tells you where it is sure and where it is not, rather than presenting one number as certain. If a prediction is low-confidence on a hard target, we say so.

Question 14

How do you stop the model from making things up?

Accepted Answer

Orbion runs purpose-built scientific models that return quantified predictions with confidence, not a free-text generator that can invent facts. Numeric outputs such as stability and suitability come from models benchmarked against measured data. Where the platform writes a plain-language summary, it is built from those underlying predictions and their sources, and uncertain calls are flagged rather than stated as fact.

Question 15

A protein can be stable but still low-yield or inactive. Does Orbion account for that?

Accepted Answer

Yes, and we are careful not to overstate what one score means. Stability, expression suitability, and function are predicted as separate signals, because a stable protein is not automatically high-yield or active. Orbion is a triage layer that narrows your options across these axes. It does not promise that a top-ranked variant will express well and stay fully active, which is what the bench step confirms.

Question 16

Can I see why the model made a prediction?

Accepted Answer

The platform surfaces residue-level and region-level detail behind key predictions, so you can see where a stability or modification call comes from and where to start. Explainability is a feature scientists ask for repeatedly, and improving it is an ongoing priority.

Question 17

Do you only work on membrane proteins and GPCRs?

Accepted Answer

No. Construct design, yield, and developability work across protein classes, including soluble proteins; our first wet-lab study was on a soluble metabolic hormone (FGF21). Where we specialize deepest is stability for membrane proteins and GPCRs, because that is where prediction is hardest, failure is most expensive, and general tools trained mostly on soluble proteins lose accuracy. So you get a protein-general production toolset plus a specialized edge on the hardest targets.

Question 18

What can Orbion not do yet?

Accepted Answer

Orbion focuses on characterization, developability triage, construct design, and stabilization for soluble and membrane and GPCR proteins. It does not do de novo binder design, antibody-specific design, or activity or catalytic-rate optimization. If your core need is one of these, we will tell you plainly rather than sell you a poor fit.

Question 19

Do you use AlphaFold? How is Orbion different from it?

Accepted Answer

Orbion uses predicted 3D structure as one input, and structure prediction answers a different question than Orbion does. A structure predictor tells you the likely shape of a protein. Orbion tells you whether that protein will express, stay stable, and be worth making, and which construct or variant to choose. It is the decision layer after structure.

Question 20

What do I give Orbion, and what do I get back?

Accepted Answer

You submit a protein sequence, or a multi-chain complex, in the platform. In minutes you get a characterization report: literature context, physicochemical properties, predicted function and binding sites, post-translational modification sites, transmembrane topology, disorder and aggregation risk, a predicted structure with confidence metrics, and a plain-language summary. From there you can design and score constructs and generate experiment-ready protocols.

Question 21

How long does a characterization take?

Accepted Answer

Minutes, not days. This is the reaction we hear most often in demos: results appear on a scientist's own protein while they watch, in place of days of manual literature and database work.

Question 22

Does Orbion support protein complexes, not just single chains?

Accepted Answer

Yes. Orbion works on single proteins and on multi-chain complexes. Characterization runs on both.

Question 23

Can Orbion recommend an expression system?

Accepted Answer

Yes. The platform predicts experimental suitability, including likely expression host and system, and can filter and compare construct options by expression fit. Scientists ask for this feature often, and it is built in.

Question 24

Can I run my own proteins, or only demo examples?

Accepted Answer

Your own proteins. Orbion is most useful on the exact targets your team is stuck on. During an evaluation we prepare predictions on your sequences in advance, and you can submit more live.

Question 25

Does Orbion integrate with our internal systems?

Accepted Answer

Orbion is a web platform and supports data import and export (for example CSV upload of variants and constructs). During a paid pilot we also take on custom development such as workflow or LIMS integration. Tell us your stack and we will scope it.

Question 26

Can I upload my own experimental structure and run on it?

Accepted Answer

This is a frequently requested capability and is on our near-term roadmap. Today Orbion generates a predicted structure with confidence metrics as one input. Support for uploading your own solved or in-house structure, so predictions run on the model you trust, is a priority we are building toward. Ask us for the current status.

Question 27

Does Orbion learn from our lab results over time?

Accepted Answer

Not automatically, and never by training on your data. The base models do not use customer data for training. Separately, we are building calibration that tunes predictions to your lab's experimental context, isolated to your organization with no cross-customer learning. This is a roadmap capability, not something the base platform does silently today.

Question 28

Can I export constructs, reports, and sequence files into our tools?

Accepted Answer

Yes. You can export constructs and analysis, and import variants and constructs by CSV. Broader export, such as construct and plasmid files for ordering and for use in standard sequence-editing tools, is part of how we fit into existing workflows. Tell us the formats your team needs.

Question 29

How current is the underlying data?

Accepted Answer

Orbion's homology and context searches run against large, regularly refreshed public protein databases (700M+ proteins across UniProt and PDB). The literature context is generated at analysis time. If data freshness for a specific source matters to your work, ask us and we will confirm the current cadence.

Question 30

Do you train your models on our data?

Accepted Answer

No. Orbion does not train its models on customer data, ever. The base models are trained on public datasets and Orbion's own proprietary data. Your sequences are used only to run predictions for you, and results are shared inside your organization but never across organizations.

Question 31

Do our sequences leave Orbion's systems?

Accepted Answer

No. Predictions and searches run entirely inside Orbion's own infrastructure. Homology search across 700M+ public proteins runs internally rather than through third-party services, so your sequences are not sent to outside APIs.

Question 32

Do you offer an on-premise or private installation?

Accepted Answer

Orbion is delivered as an EU-hosted cloud service. On-premise or private deployment is possible for enterprise engagements as part of a commercial agreement. If cloud is a blocker for your team, tell us and we will scope it.

Question 33

Is Orbion secure enough for enterprise and pharma review?

Accepted Answer

Yes. Orbion is ISO 27001:2022 certified and runs a full information-security management system. Data is encrypted in transit and at rest, access is least-privilege with mandatory MFA, and Orbion maintains documented incident-response, backup, and business-continuity processes. We are happy to complete security questionnaires and speak with your IT and legal teams directly.

Question 34

Who owns the results and IP?

Accepted Answer

You do. Any results, constructs, and analyses you generate on your targets are yours. Orbion retains ownership of its platform and models. We can sign an NDA or mutual confidentiality agreement before you share any sequence, and a GDPR-compliant data processing agreement where personal data is involved.

Question 35

Do you have references or certifications you can share?

Accepted Answer

Yes. We can share our ISO 27001:2022 certificate, plus policy extracts and audit summaries under a signed confidentiality agreement. Benchmark reports and validation summaries are available on request.

Question 36

How much does Orbion cost?

Accepted Answer

Orbion is an annual software subscription, invoiced monthly, priced by the size of the buying team rather than the size of your company. A single lab or small team starts at the entry tier; larger teams and org-wide rollouts move up. Request the current pricing sheet and we will match a tier to your team.

Question 37

Are there royalties, milestone payments, or hidden fees?

Accepted Answer

No. No royalties, no milestone payments, no hidden fees. You pay a flat subscription for your tier, with unlimited users and unlimited targets included. This is a software model, not a drug-discovery deal.

Question 38

We work on hundreds of proteins. Will you charge us per protein?

Accepted Answer

No. The subscription includes unlimited targets and unlimited users within your tier, so your cost does not grow every time you add a protein. We priced it this way on purpose, because per-target metering does not fit how CROs and engineering teams actually work. The only per-target work is an optional stabilization campaign on a hard target, which is separate project work you choose to run.

Question 39

Can we try it before committing to a year?

Accepted Answer

Yes. Orbion offers a short free evaluation and a two-month paid pilot at half the subscription rate. The pilot includes full platform access, custom development, and weekly check-ins, and ends with a conversion proposal. Most teams start with an evaluation on their own targets.

Question 40

Do you offer academic access?

Accepted Answer

Yes. Academics can sign up and request academic access through our researcher program. Access is free under a partnership agreement.

Question 41

How is Orbion different from FoldX, Rosetta, or general stability tools?

Accepted Answer

Those tools score single mutations for stability; Orbion covers the whole upstream decision and specializes where they are weakest. It designs and ranks whole constructs for yield and developability, not just point mutations, which is the part validated in our FGF21 wet-lab study. On membrane-protein and GPCR stability, where FoldX and Rosetta were trained mostly on soluble proteins and lose accuracy, Orbion correlates more closely with measured outcomes on the same public benchmark. We also frame stability honestly: rather than promise perfect stabilizer design, we reliably remove the destabilizers so you waste less bench time.

Question 42

Is Orbion just AlphaFold and FoldX wrapped in a nicer interface?

Accepted Answer

No. Orbion uses best-in-class structure prediction as one input, the way most modern tools do, but the characterization, suitability, and stability scoring come from Orbion's own models, trained and specialized for membrane proteins and GPCRs. That is where the accuracy difference comes from: on this class of target, Orbion tracks measured stability far more closely than FoldX or Rosetta on the same public benchmark. We are happy to share head-to-head benchmark results so your team can justify the difference internally.

Question 43

How is Orbion different from other AI protein-design platforms?

Accepted Answer

Most AI protein-design platforms are horizontal, spanning many protein types and modalities. Orbion is focused: a production and engineering decision layer that covers construct, yield, and developability across protein classes, with a specialized edge on stability for membrane proteins and GPCRs, the hardest and highest-value class. It is also designed to calibrate to each customer's own experimental context over time. Focus is the difference in accuracy on the hard targets.

Question 44

Is this just another large language model like ChatGPT for proteins?

Accepted Answer

No. Orbion runs purpose-built scientific models trained on protein data and public structural and biochemical datasets, not a general chatbot. A general model cannot rank a mutation's effect on a GPCR's stability. Orbion's models are benchmarked specifically for that.

Question 45

Why not just use a CRO for characterization?

Accepted Answer

A CRO gives you experimental answers, but on a timeline of weeks and at a per-target cost. Orbion gives you a computational answer in minutes, so you enter the lab with a shortlist instead of a full slate. Many customers use Orbion to decide what to send to the bench, not to replace the bench.

Question 46

How do we start an evaluation?

Accepted Answer

Contact us with one or two target sequences, or start straight away in the platform. We prepare predictions on your proteins in advance, walk through the results with your team, and set up an evaluation account so you can run your own targets. There is no sample to ship and no wait.

Question 47

What happens to our data if we stop using Orbion?

Accepted Answer

Your data remains yours and is never used to train our models. On request we delete customer data in line with our data-handling policies and any agreement in place. We can put the specifics in writing before you start.

Question 48

Who do we contact?

Accepted Answer

Reach out through the contact options on our site and we will match you with the right person for a technical walkthrough or a scoping conversation.

The Question We Get Asked Most

About Orbion

Science and validation

How it works

Data, security, and IP

Pricing and commercial model

How Orbion compares

Getting started

Still have a question?