Catalophore Tox Station

Off-target safety screening, on your desk.

An NVIDIA DGX Spark prepared for in silico toxicology. Compare the binding site of your target with the human proteome or with focused panels by target class, from kinases and GPCRs to nuclear receptors and ion channels, and let Catalophore Intelligence interpret every cavity. All computation runs on the system itself. Your structures and lead series stay inside your building, with or without an internet connection.

Launch offer: USD 7,500 including the hardware and a six-month full license.

Scientist analysing a protein binding site as a point cloud on the Catalophore Tox Station, an NVIDIA DGX Spark desktop system
849 kCavity descriptors of the human proteome, stored on the system
41,630Human proteins represented as searchable binding sites
467Confirmed drug-target pairs in the published validation
OfflineNo internet connection required for daily work
Confidentiality

We know how sensitive this information is. The strongest protection for your data is a system that works without any connection to the outside world.

01 · The gap

Between optimising a molecule and trusting it.

Roughly 30% of drugs fail clinical trials over safety concerns, and off-target toxicity is among the leading causes. Standard in vitro safety panels typically cover fewer than 60 protein targets, a small fraction of the possible off-target space.

Discovery

On-target optimisation

Ligands are designed and scored against the intended pocket. Docking and free energy methods focus on one known target.

The blind spot

Off-target binding

Unrelated proteins can evolve pockets of similar shape and chemistry. A molecule may bind targets that share neither sequence nor fold with the intended one, which sequence-based screening cannot detect.

Safety

Toxicity panels and DMPK

Panels and exposure models quantify risk for the targets they are given. The Tox Station supplies the candidates that deserve to be on that list.

02 · Applications

Three tasks, one system.

Catalophore describes every binding site as a three-dimensional point cloud of physico-chemical properties. Sites are compared by these properties, independent of sequence, fold or protein family.

Drug safety

Off-target search

Match the binding site of your target against the whole human proteome or against panels by target class. The result is a ranked list of potential off-targets with the structural evidence for every match.

Protein kinasesG protein-coupled receptorsNuclear receptorsIon channelsTransportersCytochrome P450EnzymesOther familiesWhole proteome
Early triage

In silico toxicology

The search is ligand-agnostic. It needs the binding site, so a program can be assessed before lead nomination and before a compound is fully characterised.

Structure analysis

Cavity analysis

Every pocket is annotated with more than 30 property fields. Catalophore Intelligence reads this point cloud and returns functional zones, probable mechanism and a druggability assessment, each with a stated confidence.

03 · The software

An expert read of the binding site, grounded in the numbers.

Catalophore Intelligence first distils thousands of grid points into statistics. The reasoning model then returns a structured assessment that cites this evidence.

  • Ab initio. The model receives cavity geometry and properties only, without the name of the protein, so the read is unbiased by prior annotation.
  • Structured. Executive summary, geometry, electrostatics, functional zones, druggability and recommendations in separate sections.
  • Transparent. Every section states its confidence level and its limitations.
  • Portable. The complete analysis exports as a PDF report.
catalophore studio · ai reasoning
Catalophore Studio with the electric field of a binding pocket as a point cloud and the AI reasoning panel showing the executive summary of the cavity

Cavity analysis. Electric field of a binding pocket with the executive summary written by Catalophore Intelligence.

Structures and cavities in one library.

The reference data for the off-target search is stored on the system. Your own structures join the same library and stay searchable.

  • Human cavitome. The binding sites of the human proteome as point clouds, ready for matching.
  • Your structures. Uploads and their detected cavities are indexed alongside the reference data.
  • One browser. List, card and image views with filters and saved presets.
  • Open formats. Results export as PDB, CIF and report files for your downstream tools.
catalophore platform pro · library
Library overview of Catalophore Platform Pro with counts for protein cavities, PDB structures, protein sequences, custom structures and homology models

Library overview. The counts depend on the libraries installed on a system.

04 · Workflow

From a structure to a mechanistic safety signal.

One continuous path in the browser. Each stage hands the next exactly what it needs.

Load

Structure

Open a PDB file of your target, with or without the bound ligand, and prepare it.

→
Detect

Cavity

Point-cloud detection describes the binding site by geometry and property fields.

→
Match

Off-target search

The site is compared with a panel or with the human cavitome stored on the system.

→
Interpret

AI reasoning

Catalophore Intelligence explains each match and relates it to known off-target effects.

→
Decide

Report

A ranked list with structural rationale, ready for the safety assessment and the assay plan.

Off-target search with the metribolone binding site of the androgen receptor: query point cloud and three hits in other proteins with matching scores and volumetric overlap
Example: metribolone. The binding site in the androgen receptor (PDB 1XOW) is searched against the human cavitome. All five known targets of metribolone appear within the top 100 results. Hits 1, 39 and 65 are shown, among them a hormone-binding globulin and an aldo-keto reductase without structural or sequence similarity to the query. Figure from Parigger et al., Briefings in Bioinformatics 27, bbag398 (2026), CC BY-NC 4.0.
Every predicted off-target is a testable hypothesis with a structural cause. The system prioritises which proteins belong on the experimental panel.
05 · Validation

Published and validated with a global pharmaceutical company.

The method was evaluated in a joint study on experimentally confirmed drug-target pairs. It operates without ligand information and without using sequence or overall protein structure.

24%
Top-10 recall for validated, strongly modulated targets, a conservative lower bound
13.9%
Of known off-targets with less than 30% sequence identity recovered at Top-100 and missed by BLASTP and Foldseek
467
Experimentally confirmed drug-target pairs in the validation set
15
Jurisdictions in which the point-cloud cavity method is patent-protected
Percentage of known targets recovered within the top-N ranked results for eight activity classes, compared with random expectation
Recovery by activity class. Percentage of experimentally verified targets found within the top-N results, aggregated across 467 binding sites. Strongly modulated targets are recovered first. The dashed line marks random expectation among 41,630 human proteins. Figure from Parigger et al., Briefings in Bioinformatics 27, bbag398 (2026), CC BY-NC 4.0.

Parigger, L. et al. Ligand-agnostic off-target site prediction for early toxicity screening by leveraging point cloud-based protein cavity analysis. Briefings in Bioinformatics 27, bbag398 (2026). doi.org/10.1093/bib/bbag398

Structures entering the cavitome are energy-minimised, refined and curated. On the limits of AI stereochemistry: Steinkellner, G. et al. Nature 637, 548 (2025). doi.org/10.1038/d41586-025-00111-5

06 · Data sovereignty

Designed for air-gapped operation.

Safety questions arise on unpublished structures and confidential lead series. On the Tox Station they are stored, processed and analysed in one place: the device in front of you.

Computation

Nothing leaves the system

Cavity detection, off-target search and AI analysis run on the system. The language model behind Catalophore Intelligence is served locally on its GPU.

Updates

Signed updates by USB stick

Innophore delivers each update as a single file. The system verifies the cryptographic signature before anything is unpacked, installs the update and writes a report back to the stick.

Support

Remote support under your control

The encrypted support connection is switched off by default. You open it in the dashboard when you want assistance and close it when the session is finished.

07 · Setup

From the box to the first off-target list in three steps.

The system arrives installed, configured and tested. There is no software to deploy, no cluster to book and no cloud account to open.

1

Connect power

All services start by themselves. The workbench is available after about three minutes, the AI features a few minutes later.

2

Join the system

Connect your laptop to the built-in WiFi access point, plug the system into your office network, or attach a screen, keyboard and mouse.

3

Open your browser

A welcome page links to the workbench and the system dashboard. One user account gives access to everything.

08 · The hardware

Built on NVIDIA DGX Spark.

A compact desktop system with the NVIDIA GB10 Grace Blackwell chip, a 20-core Arm CPU and 128 GB of unified memory. It runs from a standard power outlet and needs no server room.

  • Dedicated compute. The system belongs to your team. Searches run without a queue and without competing for shared resources.
  • Local language model. An OpenAI-compatible interface on the system is available for your own scripts and pipelines.
  • System dashboard. Live system metrics, network status, WiFi control, updates and support on one page.
09 · Pricing

One price, hardware and software.

The price includes the NVIDIA DGX Spark hardware and the complete Tox Station software for the first six months. Afterwards you license the modules your projects need.

Launch offer · Catalophore Tox Station
USD 7,500

Hardware included, with a six-month full license for all Tox Station modules.

Request a quote →

Included

  • NVIDIA DGX Spark hardware, installed and configured
  • Off-target search against the human cavitome
  • Panels for all major target classes
  • Cavity analysis with Catalophore Intelligence
  • Signed software updates by USB stick
  • Remote support on request

After six months

Licensing continues per module. The hardware is yours and your data remains on your system.

Get started

Bring off-target screening to your site.

Tell us about your targets and your IT requirements. We will configure the system for your programs.

Order a system

Receive a quote with delivery time and the module overview for the period after the first six months.

Request a quote →

See it first

Send us a PDB ID or a structure of your target. We will show you the off-target analysis on a live system.

Book a demo →

Questions? Write to customer@innophore.com · Built and maintained by Innophore, Graz, Austria.