Hits, repurposed drugs, off-targets. A binding-site-first way to find them.
Innophore runs computational programs for global pharma, biotech and CDMOs across discovery, safety and manufacturing. Hit identification by virtual screening against curated pockets. Drug repurposing by cavity-based matching to approved-drug databases. Off-target screening across the human proteome. De novo ligand design with paralog-level selectivity. API and biologics biocatalysis for the chiral steps and the production enzymes behind them. Every program runs on a dedicated Catalophore™ platform.
Four areas of pharma discovery, one platform underneath.
Most programs land in one of four buckets. Each one starts from a customer target, ligand series or process question, and ends with a deliverable that fits the customer’s next experiment, whether that is a binding assay, an animal model or a panel screen.
Hit identification
Virtual screening at scale against orthosteric and allosteric pockets. Up to 414k ChEMBL-derived ligands docked, rescored with semi-empirical free-energy methods, and validated with positive and decoy controls. Used recently for kinases, GPCRs and ATP-binding enzymes.
Drug repurposing
Cavity-based matching of disease-causing protein structures against approved-drug databases. The DrugSolver CavitomiX workflow has surfaced repurposing candidates for GPCRs, anti-infective targets and rare disease proteins.
De novo design and selectivity
Generative ligand design for hard targets, including those with closely related paralogs. MD-detected dynamic pockets, paralog-disagreement objectives and LigForge generation with MD plus semi-empirical QM rescoring.
Off-target safety screening
Proteome-wide cavity matching to flag structural off-targets before lead nomination. Ligand-agnostic, validated against 467 experimentally confirmed drug-target pairs in a joint study with a global pharmaceutical company. Top-10 recall of 24% for strongly modulated targets, a conservative lower bound.
Other modalities
TCR alpha and beta chain design for cancer immunotherapy, MHC-peptide epitope modeling, AI-based CDR redesign with ProteinMPNN, and homology modeling of MHC class I complexes. Where the chemistry leaves the small-molecule space, the platform follows.
API and manufacturing biocatalysis
Enzymes for the chiral and difficult steps in API manufacturing, plus engineering of biologics-production enzymes. Imine reductases, ketoreductases, reductive aminases, nucleoside transferases, RNA polymerases, DNA glycosylases. The IRED Copilot alone covers more than 11,000 curated sequences.
The work is run by senior computational scientists against a defined target question. The customer keeps the IP on every hit, ligand and TCR variant we deliver, with a nine-month exclusivity option on shared analysis results.
How we run pharma engagementsValidation set first. Screening second. Honest scoring throughout.
A typical campaign is six to twelve weeks. We start by building a validation set with positive controls and DUDE-Z decoys so the docking pipeline is benchmarked against the customer’s actual chemistry before any new ligand is scored.
Validation set
Positive controls, literature inhibitors, DUDE-Z decoys. Re-dock the natural substrate. Confirm the scoring function actually separates binders from non-binders for this target. Two to four weeks.
Screening or design
Catalophore™ cavity detection, virtual screening against curated libraries, MD on dynamic pockets, generative design with selectivity objectives, or proteome-wide off-target search. Three to eleven weeks.
Post-analysis and hand-off
Top hits rescored, binding poses inspected, drug-likeness filters (PAINS, QED, Lipinski-style ranges) applied. Final ranked list with PDB files, PyMOL sessions and a written report. Roughly two weeks.
Pharma work is held to a higher standard. AI-generated structures are useful but they get the chemistry wrong often enough to matter. Stereochemistry errors occur in 4.4% of predicted ligand poses by one published benchmark, and one study found that a single mutation can simultaneously attenuate transcription termination and RNA-dependent activity in a key biopharma enzyme. Models that look right are not always right.
Every model that enters a Catalophore™ campaign is energy minimized, refined and reviewed by a senior scientist before it informs a hit list, a repurposing call or an off-target flag.
Anonymized examples from recent pharma programs.
Nine representative engagements covering oncology, anti-infectives, metabolic disease, rare disease, cardiovascular and pharma safety. Customer names, target identities and ligand structures are generalized. Project scope, methods and deliverable shape are reported as run.
Virtual screening of 414k ChEMBL compounds against an ATPase
An oncology-focused diagnostics company needed inhibitors of a multifunctional ATPase target, both at the ATP-binding pocket and at a potential allosteric site. We filtered ChEMBL to roughly 414,000 drug-like ligands, generated low-energy 3D conformers with RDKit, ran ten AutoDock-Vina iterations per ligand-target pair on the Catalophore™ Pro cluster, then rescored top binders with semi-empirical QM. The pipeline was validated against ATP, a partial inhibitor and DUDE-Z decoys before any production screen.
GPCR allosteric modulators by cavity matching
A specialty pharma company needed allosteric small-molecule modulators of a class-B GPCR involved in metabolic disease, with the goal of breaking the dependence on injectable peptide drugs. We mapped the orthosteric and previously unexplored allosteric cavities with the DrugSolver CavitomiX workflow, matched them against ligand-containing cavities in approved-drug databases, then filtered for scaffold diversity and physico-chemical compatibility.
Proteome-wide off-target call before lead nomination
A global pharmaceutical partner wanted to test whether structural cavity matching could replace ligand-only off-target screening for early triage. We ran a ligand-agnostic match of the binding site across the human cavitome (849,355 descriptors covering 41,630 proteins), validated against 467 experimentally confirmed drug-target pairs. The method recovered off-targets across structurally unrelated families using no ligand, sequence or overall fold information.
Proteome-scale target discovery for antibiotics
An antibiotic-discovery biotech needed to find multi-target binding pockets in a Gram-negative pathogen, because drug-resistance is harder to evolve when multiple essential enzymes are hit at once. We modeled the entire pathogen proteome with AlphaFold 2, OpenFold and ESMFold, ran cavity analysis with the CavitOmiX co-pilot, then cross-matched 300 essential-gene pockets against each other and against the full proteome to surface convergent binding-site clusters.
TCR design against an MHC-presented tumor antigen
An immunotherapy company needed novel T-cell receptors that recognize a specific intracellular tumor antigen peptide presented by MHC class I (HLA-A*02:01). We modeled the MHC-peptide complex from existing TCR3d templates, optimized side chains and ran MD relaxation, then redesigned TCR alpha and beta CDR loops with a mix of homology modeling, AI folding and ProteinMPNN-based CDR generation, with iterative energy minimization and binding-affinity ranking.
Drug repurposing for rare disease proteins
A clinical research group needed candidate molecules to repurpose against disease-causing proteins in two rare diseases driven by gain-of-function or gene-duplication mechanisms. We modeled the variant protein structures, detected cavities and encoded their physico-chemical properties as 3D point clouds, then matched against public ligand and approved-drug databases. Promising candidates were progressed through docking, binding-affinity calculation and MD validation before hand-over to the customer’s wet-lab pipeline.
Paralog-selective inhibitors of a metalloprotease
A top-tier global pharma needed selective small-molecule inhibitors of a Zn-dependent metalloprotease while sparing a paralog with an almost identical active-site geometry. We mapped cavities on cryo-EM structures, completed missing regions with Boltz-2, ran MD to find dynamic exosite pockets, and used a paralog-disagreement objective to drive de novo ligand generation through LigForge, with MD and semi-empirical QM rescoring and glycosylation-shield filtering.
Target deconvolution for an unknown-mechanism ligand
In a deconvolution program, a customer brought a ligand whose intended target was known but whose broader binding profile was open. We built homology models for thousands of candidate proteins, identified pockets across them and ran docking to assess binding affinity across the set. 128 candidates were prioritized on predicted free energy, ligand planarity and key residue distances, and the top 50 were recommended for wet-lab follow-up. Wet-lab testing confirmed exceptional binding affinity at three of those targets.
CavitOmiX screening of a producer strain
A global generics manufacturer wanted to know whether the genome of its production strain harbored enzymes that could degrade the final product. We ran six-frame SeqScan of all open reading frames, modeled candidates with AlphaFold 2, ESMFold and RoseTTaFold, generated cavities and point clouds, matched them against query cavities from the relevant degrading family, and added an AI classifier trained on family-specific data.
Enzymes for the chemistry your API and biologics need.
A second body of work, often running in parallel to discovery programs. Discovery and engineering of biocatalysts for the chiral steps in API synthesis, the production enzymes behind biologics workflows, and specialty enzymes for diagnostic chemistry. All anonymized.
Imine reductase for a chiral amine API
A global top-10 pharma needed an imine reductase capable of reductively aminating a bulky ketone with ammonia at greater than 10 mM substrate concentration. We ran a Catalophore™ in-silico search starting from the customer sequence, modeled representative IRED structures across the family, matched 3D point clouds against the substrate pocket, and proposed a focused variant library.
Ketoreductase engineering for a chiral alcohol API
A major pharma had a ketoreductase producing a chiral alcohol API intermediate but needed higher than 98% ee toward the target enantiomer. We did in-silico protein engineering with substrate positioning in the active-site cavity, sequence-space and hot-spot analysis, RosettaDesign plus structure-guided rational mutagenesis, MD simulations of selected designs, and ML ranking trained on the customer’s prior mutagenesis data.
Reductive aminase at neutral pH and 1:1 stoichiometry
A major pharma needed reductive aminases acting on an acetonide-protected ketone with a primary alkyl amine at equimolar ratio and neutral pH, conditions outside the comfort zone for most RedAms. We combined the IRED Copilot platform (more than 11,000 curated IRED and RedAm sequences and structures with an ML reactivity model) with an adapted Boltz-2 deployment for protein-ligand binding and affinity prediction.
T7 RNA polymerase variant for mRNA production
A pharma supplier of in-vitro transcription reagents wanted a T7 RNA polymerase variant with reduced double-stranded RNA byproduct formation during mRNA manufacturing. We modeled up to 2,000 representative sequences, identified cavities including cryptic pockets near the C-terminal motif, redesigned the C-terminus along alternative paths to avoid existing patents, and scored stability with Rosetta ΔΔG on both the initiation and elongation complexes.
Transferases for non-natural nucleoside bases
A US top-tier pharma needed nucleoside (deoxy)ribosyl transferases that accept a non-natural deaza purine base for nucleoside analogue synthesis. We ran Catalophore™ cavity volume and chemistry characterization, substrate-walking docking, BLAST plus structural modeling of representative sequences, and metagenomic mining via MGnify to expand the candidate pool.
DNA glycosylase for 5MeC and 5hMeC excision
A diagnostics division of a global pharma needed a DNA glycosylase that excises 5-methylcytosine or 5-hydroxymethylcytosine, binds DNA sequence-agnostically, has no AP-lyase activity, and lacks the Fe-S cluster. We ran constraint-driven sequence mining (public plus metagenomic plus proprietary), AlphaFold and homology modeling, Catalophore™ cavity and Halo analysis, base-flipping geometry checks, in-silico mutagenesis to remove lyase chemistry and swap recognition loops, MD validation, and up to three iterative design cycles with customer wet-lab feedback.
Patented method, GPU acceleration, validated with global pharma.
The core technology is protected, benchmarked on production hardware and validated through a joint study with a global pharmaceutical company. The service is built on top of that foundation.
Protected core method
The point-cloud cavity method is a granted patent (US 2015/0302142 A1, also granted as US 10,825,547 B2) and protected across 15 jurisdictions.
100x on production GPUs
Cavity matching throughput accelerated by a factor exceeding 100 on enterprise GPU systems. A proteome match that took roughly 625 seconds now completes in about 5.
Official NVIDIA partner
Listed NVIDIA healthcare and life sciences partner. The proteome-wide cavitome was built jointly with NVIDIA on BioNeMo in two weeks, against an earlier baseline of over a year.
467 drug-target pairs
Validated in a joint study with a global pharmaceutical company across 467 experimentally confirmed drug-target pairs. Cavity-based matching, no ligand information used.
The same cavity search that surfaces off-targets across the human proteome also finds repurposing opportunities and antibiotic multi-target pockets. One method, validated across modalities, from small molecules to TCRs.
Cavity-based discovery, generalized“NVIDIA boosted our performance so that we can run five million off-target predictions per second. We had been at a few hundred before.”
Christian C. Gruber · CEO, Innophore
One platform, several ways to work with us.
Most pharma customers start with a fixed-scope pilot work package and grow from there. Every engagement includes a customer-branded Catalophore™ instance for the duration of the project. IP on enzymes is transferred to the customer. IP on shared discovery results comes with a nine-month exclusivity option.
- WP
Work package projects
Fixed scope, defined deliverables, six to twelve weeks. Most common entry point for new partners.
- API
Platform integration
Standard REST endpoints for structure submission and ranked match retrieval. Plugs into existing discovery and safety stacks.
- IP
IP on results
Customer keeps the IP on identified, built or modified enzymes. Shared analysis results come with a nine-month exclusivity option for the customer.
Typical engagement shapes
- Pilot Single work package One target or one ligand series. Validation set plus screening or design. Six to twelve weeks. Ranked candidates with structural rationale and a customer Catalophore™ instance.
- Program Multi-WP campaign Hit identification followed by lead optimization or off-target follow-up, defined milestones, monthly review. Optional wet-lab hand-off through customer’s existing CROs.
- License Customer Catalophore™ instance Time-bounded platform license with a dedicated cavitome and Copilot modules, including options for on-premise deployment on dedicated hardware.
Bring us a brief.
Discovery teams, safety teams and platform teams come in through different doors. Both lead to the same scientists.
Start a project
Send us a target, a ligand series, a disease protein or an existing program where the off-target picture is unclear. We come back with a scoped pilot proposal and an honest call on feasibility.
Send a brief →Explore integration
Add proteome-wide cavity analysis to your existing safety or discovery stack through a standard API. We can demo the integration on your test target.
Talk integration →