☰

Anyo Labs AB

Drug discovery as engineering, not alchemy

Industry
AI-driven drug discovery (TechBio)
Headquarters
Gothenburg, Sweden
Public information as of
February 2026

A4BEE prepared this analysis from publicly available sources. It reflects our own reading of Anyo Labs AB's published strategy and is not endorsed by, or produced in cooperation with, Anyo Labs AB. Company website

Strategic priorities

Anyo Labs is a five-person TechBio company spun out of the University of Gothenburg that screens drug-like compounds against protein targets using its iScore scoring function instead of traditional molecular docking. The published iScore framework was developed by the company's University of Gothenburg founders and is described in a 2024 paper in the Journal of Chemical Information and Modeling. The company is now packaging this screening capacity in the NOVA nano-cluster, a 6-litre device that combines four NVIDIA Jetson modules with an x86 system-on-chip and screens more than 10,000 molecules per second at under 200 W.

Anyo addresses fast gaps in the early-stage drug discovery workflow. The team states publicly that it can score around one billion compounds in a single day, an order-of-magnitude cost reduction versus traditional screening, and that the same machinery is already serving AI-driven lead optimization programs in oncology, inflammatory disease and tropical disease (including Dengue). The NOVA launch in May 2025 is the move that turns the company from a software service into a hardware-integrated product, which is also the move that introduces IT/OT (information technology/operational technology) convergence, fleet management, and supply-chain concerns that the founding team has not had to worry about before.

Anyo does not run its own wet lab. Every experimentally validated hit is synthesised and tested by external contract research organisations, which means the pace of the company is capped by the pace of its partners. The recurring 10x speed mismatch between in-silico scoring and physical validation is the operational friction the leadership cites most often, and the work of turning partner assay data into a form iScore can refine against is the operational priority the v3 page can engage with.

Funding to date is Seed- to Accelerator-stage: Chalmers Ventures in 2023, Vinnova Innovation Agency in 2024, and acceptance into the Swedish-American Chamber of Commerce accelerator in 2025. Total disclosed funding is in the single-digit million-dollar range, which is consistent with a company that has to choose between hiring wet-lab scientists and building the digital infrastructure that lets five scientists do the work of fifty.

Challenges we see

  • Operations Integration

    Closing the gap between in-silico scoring and partner wet-lab validation

    Anyo scores up to one billion compounds per day, but every hit must be synthesised and tested by external contract research organisations. The cycle time of the partners sets the pace of the company, and the gap between the speed of prediction and the speed of physical validation is the operational friction the founders describe in public statements.

    Where the digital discovery step is faster than the validation step by orders of magnitude, the question becomes how the partner's data arrives back to the in-silico side in a form iScore can refine against, rather than how to run more screen.

  • Operations Manufacturing

    Sourcing and shipping NOVA hardware at small-batch scale

    The NOVA launch in May 2025 introduces NVIDIA Jetson modules, x86 system-on-chips and 6-litre enclosures into a company with five founders and no manufacturing facility. Hardware is currently assembled in small batches through contract manufacturers for deployment to pharmaceutical client sites.

    For a hardware product assembled in small batches, the supply chain for the components that make up the unit sets the size of the operational promise, and the way the team documents and tests each unit before shipment determines what the next twenty units will look like.

  • Digital Integration

    Operating a fleet of NOVA units across client sites

    As NOVA units are deployed to pharmaceutical sites, Anyo has to maintain them remotely, push updated MolGen and MolAI models over the air, and monitor thermal and computational performance across a hardware fleet that sits inside client environments.

    Where each NOVA unit ships with proprietary models that evolve over time, the practical question for the customer becomes how the updates and the unit's health are seen from outside, rather than whether the algorithms inside the unit are correct.

  • Compliance Regulatory

    Showing how a neural-network score reaches its answer

    iScore bypasses explicit protein-ligand interaction modelling in favour of neural-network descriptors. Regulators at the FDA and EMA are described in the 2025 literature as increasingly demanding explainable AI and documented proof that models are not amplifying biases in training data, especially as AI-derived molecules enter pre-clinical and clinical review.

    Where the scoring function is a neural network, the question for the regulator becomes what the model looked at and what it ignored, and the practical work is to make the path from protein-ligand input to predicted affinity readable and reproducible.

  • Digital Integration

    Stitching partner laboratory data into the iScore refinement loop

    External partners return experimental results in fragmented formats with inconsistent metadata. Anyo scientists currently clean and re-contextualise this data manually before they can use it to refine iScore, which is the work the company refers to as building a 'broken data chain' back into a 'clean data chain'.

    Where partner-assay data has to be cleaned by hand before it can refine the model, the delay between physical experiment and model update is set by the cleaning step, and the work of standardising the data at the partner boundary shortens that loop.

Opportunities, by urgency and business impact

Each bubble is one opportunity, numbered to match the list below. Further right means it bites sooner; higher means a bigger effect on the business. A bigger bubble means a bigger implementation effort.

Source: A4BEE analysis of public sources

  1. Standardising partner laboratory data at the submission boundary

    Anyo reports that partner-supplied assay data is fragmented, inconsistent in metadata and arrives in formats that the iScore pipeline cannot load directly. The current practice is to clean and re-contextualise the data by hand before it can be used to refine the model.

    An agreed data schema and a submission portal would let validation partners upload assay results with the metadata Anyo needs, and ingestion tools would clean and contextualise the data before the result reaches the iScore refinement pipeline.

    • Anyo Labs, About — Discovery, 2025
    • iScore: A ML-Based Scoring Function for De Novo Drug Discovery, Journal of Chemical Information and Modeling, 2024

  2. Remote health monitoring and model updates for the NOVA fleet

    The NOVA nano-cluster ships with proprietary MolGen and MolAI models that need to be updated over time. Each deployed unit operates inside a customer environment with its own IT and OT estates, and there is no fleet-level view of thermal performance, GPU utilisation, or update status.

    A fleet management layer over the deployed NOVA units would surface thermal and performance metrics, push model updates with rollback capability, and give Anyo's CTO a single view of the fleet while leaving the underlying algorithms proprietary.

    • Anyo Labs introduces NOVA, May 2025
    • Anyo Labs, About — The next frontier in medicine, 2025

  3. An ontology-based data backbone for compound and assay records

    Anyo screens hundreds of millions of compounds against protein targets and retrains iScore as new PDBbind and CASF data becomes available. The training pipelines and the live scoring data are currently maintained through bespoke scripts and manual handoffs.

    An ontology-based data lakehouse would define compound, target, assay, descriptor and binding-pocket entities once, then load the screening output and the training data against the same model so the model and the prediction pipeline can share what they know.

    • iScore: A ML-Based Scoring Function for De Novo Drug Discovery, Journal of Chemical Information and Modeling, 2024
    • Anyo Labs, About — The next frontier in medicine, 2025

  4. Explainability and submission infrastructure for AI-derived molecules

    Anyo is preparing for FDA and EMA submission of AI-derived lead optimisation candidates. The submission requires an eCTD (electronic Common Technical Document) package, PDF version control, an XML backbone, and evidence that the iScore score is reproducible and explainable to a reviewer.

    A submission infrastructure with an audit-trail of every model version, every descriptor set, and every result going into a submission would replace the manual preparation of an eCTD package for every lead optimisation candidate.

    • The future of AI regulation in drug development: a comparative analysis, PMC, 2025
    • eCTD Validation Requirements: A Comprehensive Technical Guide, IntuitionLabs

  5. Continuous ML-Ops for the iScore model family

    iScore is published as a family of models (iScore-DNN, iScore-RF, iScore-XGB, and a hybrid base learner). Each model is retrained as new data appears, and the current practice is to manage the retraining and validation steps by hand.

    An ML-Ops (machine learning operations) pipeline that watches for new PDBbind and CASF releases, retrains the iScore model family against the new data, and validates the result against held-out benchmarks would keep the model's accuracy lead current without the team having to coordinate the steps by hand.

    • iScore: A ML-Based Scoring Function for De Novo Drug Discovery, Journal of Chemical Information and Modeling, 2024

What we'd propose

  • Enterprise AI

    Partner laboratory data submission hub

    A secure, automated data port through which Anyo's validation partners submit assay results with the metadata iScore needs, the data is cleaned and validated at the boundary, and the cleaned result is loaded into the iScore refinement pipeline.

    • Shared submission schema

      One agreed format for partner data

      Define and publish a submission schema for compound identifiers, assay conditions, target identifiers, replicate metadata and quality flags that every validation partner uploads against, so the incoming data arrives in a form the iScore pipeline can load directly.

    • Automated validation and normalisation

      Cleaning happens at the boundary

      Build validation and normalisation rules at the submission boundary so anomalous results, missing fields and unit inconsistencies are flagged before the data enters the iScore refinement pipeline, rather than being discovered by a scientist.

    • Closed-loop feedback to partner labs

      Two-way partner communication

      Provide a workspace where Anyo scientists can request follow-up experiments, share AI-generated insights, and resolve data ambiguities with the originating partner, so the data loop between prediction and validation is bidirectional and trackable.

    • Partner data arrives iScore-ready, so the hand-cleaning step is removed from the loop.
    • The pace of the refinement loop is set by partner throughput, not by Anyo's cleaning capacity.
    • Anomalous data is flagged at the source, so the iScore refinement loop is no longer the place errors are caught.

    We've done this before

    See how we do Enterprise AI
  • Digital CDMO

    NOVA fleet orchestration platform

    A fleet management layer over the deployed NOVA nano-clusters that monitors thermal and computational health, pushes MolGen and MolAI updates with rollback, and gives Anyo's CTO a single view of the fleet without exposing the proprietary model internals.

    • Thermal and utilisation monitoring

      Health of every unit on one screen

      Deploy IIoT (industrial internet of things) instrumentation across the NOVA fleet so each unit reports GPU utilisation, thermal behaviour, power draw and screening throughput, and the CTO has a single view of the fleet's health.

    • Edge-to-cloud model updates

      Updates with rollback

      Build a secure orchestration layer that pushes MolGen and MolAI updates to every deployed NOVA unit with version control, staged rollout, and automatic rollback when an update fails validation, so the model improves in the field without manual site visits.

    • Zero-trust security at the edge

      Proprietary models stay protected

      Apply a zero-trust security model (an architecture where every request is authenticated regardless of network location) to each deployed unit so the proprietary models and the customer data they process are protected against physical and network-level compromise, with hardware-rooted identity for each NOVA appliance.

    • The CTO sees the fleet's health without travelling to the customer site.
    • Model updates reach every unit with version control and rollback, not a manual site visit.
    • The proprietary models and the customer data stay protected under a zero-trust security model.

    We've done this before

    See how we do Digital CDMO
  • Enterprise AI

    Ontology-based data lakehouse for compound and assay records

    An ontology-based data lakehouse (a unified storage architecture that combines the flexibility of a data lake with the structure of a data warehouse) that defines compound, target, assay, descriptor and binding-pocket entities once, then loads the live screening output and the PDBbind and CASF training data against the same model, so the model and the production pipeline share what they know.

    • Shared ontology for chemistry and assay data

      One entity model for everything

      Define compound, target, assay, descriptor, binding pocket and result as explicit entities with explicit relationships, so the live screening output and the training data load against the same model and a query written once returns comparable answers.

    • High-throughput screening storage

      Built for billion-compound days

      Implement a storage architecture that can hold the output of a billion-compound screening day and can serve it back to the iScore refinement pipeline in minutes, rather than rebuilding the screen on every retraining.

    • Analytics and retrieval layer

      Questions answered without IT tickets

      Expose the model through a retrieval layer so the scientist and the model retraining pipeline can both ask questions of the data set without commissioning a new extract for each one.

    • The same data model serves the production pipeline and the retraining pipeline.
    • A new target or a new binding-pocket descriptor joins the model rather than triggering a new pipeline.
    • The retraining loop sees the same data the production screening sees, so the model and the prediction stay aligned.

    We've done this before

    See how we do Enterprise AI
  • Digital Lab

    Explainability and submission infrastructure for AI-derived molecules

    An explainability and submission infrastructure that produces the eCTD package, the iScore model audit trail, and the reviewer-facing explainability layer for each AI-derived lead optimisation candidate, so the regulatory submission is built from the system rather than assembled by hand.

    • Electronic Common Technical Document assembly

      Folder structure, XML backbone, PDF version control

      Implement the eCTD folder structure, the XML backbone and the PDF version-control pipeline required by FDA and EMA, so the submission package is generated from the source records rather than assembled by hand for each candidate.

    • Model audit trail

      Every model version, every result

      Build a complete audit trail of every iScore model version, every descriptor set, every binding-pocket feature used in a screening run, and every input compound that led to a lead optimisation candidate, so a reviewer can trace a predicted affinity back to the data and the model that produced it.

    • Reviewer-facing explainability

      What the model looked at

      Build an explainability layer that translates iScore's neural-network descriptors into a reviewer-facing view of which features drove the prediction, with confidence intervals and a comparison against a traditional docking run so the regulator can see what the model was and was not relying on.

    • The eCTD package is generated from the source records, with version control and audit trail built in.
    • A reviewer can trace a predicted affinity back to the data and the model that produced it.
    • An AI-derived molecule is presented to the regulator with the evidence path intact, rather than as a 'black box'.

    We've done this before

    See how we do Digital Lab
  • Agents

    AI agents for regulatory document work

    Narrow, reviewable agents that take the repetitive part of regulatory document work for AI-derived molecules: drafting the first eCTD section from source records, checking a draft against the FDA and EMA submission template before review, and finding every controlled document a standards change or a model update touches. A named reviewer signs off every output.

    • First-draft eCTD section drafting

      Drafts from system records

      Generate the first draft of an eCTD section (quality, non-clinical, or computational model description) directly from the underlying system records, so the regulatory author edits and judges rather than assembles.

    • Template and completeness check

      Gaps found before review

      Check a submitted draft against the FDA and EMA submission template and the company's own regulatory checklist, returning missing or inconsistent sections before the document enters the human review queue.

    • Impact search across the regulatory document set

      Where a model change lands

      When a new iScore model version is released, or a binding-pocket descriptor set is updated, retrieve every controlled regulatory document that references the previous version and rank them by how directly they are affected, so the scope of the update is known on day one.

    • Review queues move faster because drafts arrive complete and against the template.
    • The scope of a model or descriptor change is established by search rather than by recollection.
    • Every output is traceable to the source records it came from and signed off by a named reviewer.

    We've done this before

    See how we do Agents

Digital maturity: today and target

Scored out of 100 across six dimensions. The target is what Anyo Labs AB's own published ambition implies — not a perfect score.

Source: A4BEE analysis of public sources
Partner data integration 45 → 85
The iScore scoring function and the partner wet-lab data are still joined by hand. The submission hub proposal is the practical path to a single, machine-readable handshake between Anyo and its validation partners.
IT/OT convergence 35 → 80
The NOVA launch in May 2025 is the moment the company takes on operational technology. The fleet orchestration layer is the first time the operational side of the product has a continuous view rather than a per-site manual one.
Discovery pipeline automation 60 → 85
The iScore scoring and the MolGen generative platform are highly automated for the discovery side. The data plumbing around the model and the submission side are still hand-coordinated, which is the gap the ontology and explainability proposals close.
Regulatory readiness 30 → 75
AI-derived molecules into FDA and EMA review is a moving regulatory target. The submission infrastructure and the agents proposal are the practical work to make the regulatory path reproducible rather than one-shot per candidate.
Cybersecurity 40 → 80
The proprietary iScore model and the customer data that NOVA processes are the core intellectual property. The zero-trust edge security model in the NOVA fleet proposal is the layer that protects the IP at the customer site.
Data backbone scalability 50 → 85
The screening side already handles billion-compound days. The data backbone that serves the model and the retraining pipeline is what the ontology-based lakehouse proposal builds, so the daily screening and the model retraining share one data model.

Check this yourself

Our Service Portal has free self-assessments and market comparisons. These are the ones that line up with what we've read above — no sales call required.

Think we've read this right?

Talk to us

Related reading

This is an independent analysis prepared by A4BEE from publicly available information as of February 2026. It reflects A4BEE's own interpretation and opinion, is not affiliated with, endorsed by, or verified with Anyo Labs AB, and may be incomplete or inaccurate. All company names and trademarks are the property of their respective owners. To request a correction or removal, contact [email protected].