Precision oncology across two continents

What we'd propose

• Enterprise AI

  Global Cloud Proteomics Data Lakehouse

  We deploy a cloud-native Data Lakehouse architecture spanning Lund and Watertown that eliminates the trans-Atlantic data gravity bottleneck — automated data ingestion from mass spectrometers, standardised phosphoproteomic data formats, and controlled access for Watertown scientists to process and view results within minutes of experiment completion in Sweden.

  • Automated mass spectrometer data ingestion pipeline

    Proteomics data available within minutes of acquisition

    Build an automated data ingestion pipeline from Lund's mass spectrometry instruments to the cloud data lakehouse — handling the large raw files, running the standard proteomics processing workflow automatically, and making processed results available to Watertown scientists through a secured browser interface without manual file transfer steps.

  • Cross-Atlantic data access governance

    Watertown scientists access Lund data within regulatory boundaries

    Implement data access governance that satisfies both Swedish and US data protection requirements — ensuring that patient-level proteomics data from clinical trials is accessible for analysis without violating jurisdictional data residency requirements.

  • Real-time phosphoproteomic heatmap visualisation

    Protein phosphorylation patterns visible to clinical teams

    Deploy a cloud-native visualisation layer that renders phosphoproteomic heatmaps and pathway activation scores in real time for Watertown scientists — the same views that Lund researchers see, accessible from the clinical team's workflow without installing specialised software.

  • Clinical decisions — patient stratification, trial inclusion/exclusion — are based on the latest phosphoproteomic characterisation rather than data that is days or weeks old because it had to be manually transferred.
  • The data lakehouse becomes the single source of truth for AP3 data, enabling Lund and Watertown to collaborate on the same dataset without the version control and file transfer overhead that currently creates confusion about which team has which version of which dataset.
  • The infrastructure investment is reusable — when Acrivon expands to additional therapeutic areas or imaging modalities, the data lakehouse ingestion framework provides the foundation without starting from scratch.

  We've done this before

  • Creating a Digital Lab for a Global Biotech Player
  • Ontology-Driven Ecosystem for Biomanufacturing Data

  See how we do Enterprise AI
• Digital Lab

  Digital Pathology Hub-and-Spoke Network for OncoSignature

  We design and implement a Hub-and-Spoke digital pathology infrastructure integrating regional PhenoImager instruments with centralised OncoSignature scoring algorithms — enabling algorithmic diagnostic scoring at distributed imaging sites with results available to the clinical team within hours of tissue preparation rather than days of physical sample transit.

  • Regional PhenoImager instrument connectivity

    Every regional PhenoImager connected to the OncoSignature cloud

    Deploy Akoya Proxima integration at regional imaging sites — connecting each PhenoImager to Acrivon's cloud scoring engine so that slide images are uploaded digitally as soon as staining is complete, eliminating the physical sample logistics that currently determine turnaround time.

  • Cloud OncoSignature algorithmic scoring engine

    AI scoring returns results without physical slide transit

    Deploy the OncoSignature scoring algorithm in a cloud-hosted environment that receives digitised slide images from regional instruments and returns AI-scored results to the clinical team — with the algorithm version and confidence score recorded for each result as part of the FDA PMA audit trail.

  • Turnaround time monitoring dashboard

    Every sample tracked from collection to result

    Build a clinical operations dashboard that tracks each patient sample from collection at the clinical site through tissue preparation, imaging, and algorithmic scoring — surfacing bottlenecks in the logistics chain that are causing turnaround delays and enabling proactive intervention before a patient is lost to a competing trial.

  • Turnaround time drops from seven to ten days to within 24 to 48 hours of slide preparation — the clinical team receives OncoSignature results before the patient's condition has deteriorated or before competing trial enrolment windows have closed.
  • The digital pathology network generates the imaging volume data required to continuously improve the OncoSignature algorithm — every additional case increases the model's accuracy and the evidence base for the FDA PMA submission.
  • Regional PhenoImager placement expands the geographic footprint of clinical trial sites — patients who live far from the central laboratory can participate in ACR-368 trials, increasing enrolment velocity.

  We've done this before

  • Digital Transformation of QC Labs: From Paper-Based Legacy to Integrated LES
  • Creating a Digital Lab for a Global Biotech Player

  See how we do Digital Lab
• Enterprise AI

  MLOps Pipeline for KaiSR Generative Proteomics Model

  We build an automated machine learning operations infrastructure for the KaiSR model — automated data labelling pipelines from clinical outcomes, HPC compute orchestration for burst model retraining, and model version management with regulatory-grade audit trails that satisfy FDA expectations for AI/ML-based medical device software.

  • Clinical outcome data integration pipeline

    Every outcome event triggers model retraining signal

    Build a data pipeline that automatically links clinical outcome data — patient response, progression-free survival, adverse events — to the phosphoproteomic profiles from the same patients, creating the labelled training dataset that triggers KaiSR model retraining with the latest clinical evidence.

  • HPC-on-demand compute for model training

    Burst training cycles without dedicated infrastructure investment

    Configure cloud-based HPC clusters that spin up on demand when a model retraining cycle is triggered — providing the compute capacity for intensive training sessions without requiring Acrivon to maintain permanent HPC infrastructure that would sit idle between training cycles.

  • ML model governance and audit trail

    Every model version documented for FDA PMA requirements

    Implement a model governance layer that tracks every version of the KaiSR model — training data, hyperparameters, performance metrics, and deployment status — creating the documentation that FDA reviewers expect for AI/ML-based software as a medical device and enabling rollback to any previous version if a new release underperforms.

  • The model learning cycle accelerates from months to weeks — every new clinical outcome contributes to model improvement within days rather than being added to a backlog of manual retraining requests.
  • The HPC-on-demand approach eliminates the capital cost of permanent HPC infrastructure while providing burst compute when needed — the model training cost scales with training frequency rather than with maximum theoretical compute demand.
  • The model governance framework satisfies FDA expectations for AI/ML device software — the PMA submission includes the model version management documentation that regulators increasingly require for algorithmic diagnostics.

  We've done this before

  • Creating a Digital Twin to Simulate Biopharma Processes
  • Microalgae Cell Detection & Classification with an ML Model

  See how we do Enterprise AI
• Digital CDMO

  CMC Digital Twin and CDMO Supply Chain Control Tower

  We implement a real-time manufacturing visibility platform connecting ACR-368 CDMO partners to a centralised CMC Digital Twin — enabling Continuous Process Verification, predictive batch failure detection, and proactive supply chain management for the clinical trial material that ACR-368 trials depend on.

  • CDMO sensor data integration and normalisation

    Manufacturing data from CDMO sites normalised to common model

    Deploy OPC UA or equivalent data connectors to the CDMO manufacturing sites, normalising production data — temperature profiles, pH curves, yield metrics, deviation logs — into a common data model that the CMC Digital Twin can ingest and analyse regardless of which CDMO or which SCADA system is used at each site.

  • Continuous Process Verification dashboard

    Every batch visible from Gdynia in real time

    Build a Continuous Process Verification dashboard that presents each active manufacturing batch with its real-time sensor profile against the validated process recipe — flagging any parameter that approaches an out-of-specification boundary before the batch is complete and must be dispositioned as failed.

  • Predictive batch outcome model

    Batch quality predicted from sensor data before completion

    Train a predictive model on historical batch data that takes the real-time sensor profile from an in-progress batch and predicts the probability of meeting final quality specifications — enabling the manufacturing team to make go/no-go decisions on a batch before the quality outcome is measured rather than after.

  • The supply chain for ACR-368 clinical trial material is no longer a Black Box — real-time visibility into CDMO manufacturing means that a developing deviation is identified and addressed while the batch is still in progress rather than after it has already failed.
  • Predictive batch outcome modelling reduces the number of failed batches that Acrivon discovers only through the Certificate of Analysis — each prevented failed batch avoids the patient impact of a supply disruption and the cost of replacing clinical trial material.
  • The CDMO relationship is transformed from a passive reporting model to an active partnership — the Control Tower data enables Acrivon to have informed conversations with CDMO partners about process optimisation rather than accepting whatever the CoA reports.

  We've done this before

  • Real-Time Process Intelligence & KPI Visualization
  • Lifecycle Management for Adaptive Biotech Assets

  See how we do Digital CDMO
• Digital Lab

  Unified Scientific Data Fabric Across Lund and Watertown

  We deploy an enterprise scientific data management platform that integrates phosphoproteomic spectra, spatial imaging data, and clinical metadata into a single searchable source of truth — connecting the Lund discovery engine and the Watertown clinical operation through a unified data architecture that enables multi-modal analysis, ELN synchronisation, and the Common Operating Picture that the trans-Atlantic structure currently prevents.

  • Multi-modal data ontology and harmonisation layer

    Proteomics and imaging data in a unified data model

    Build a data ontology that harmonises phosphoproteomic spectra and spatial imaging data into a common patient-centric data model — enabling queries that span both data types for the same patient without requiring manual cross-referencing between separate systems.

  • Mass spec instrument IoT connectors

    Lund mass spec data flows automatically into the data fabric

    Deploy IoT data connectors for the Lund mass spectrometry instruments that automatically capture raw spectra, run the standard processing pipelines, and store processed results in the harmonised data model — eliminating the manual export and import steps that currently create data versioning problems.

  • ELN synchronisation with clinical trial管理系统

    Lab notebook and trial management always in sync

    Implement bi-directional synchronisation between the Lund ELN and the Watertown clinical trial management system — ensuring that any protocol amendment or data clarification is reflected in both sites' record-keeping systems without manual re-entry.

  • The KaiSR model trains on multi-modal data — phosphoproteomic and imaging signals combined — rather than on each data type in isolation, which improves the model's predictive accuracy for patient response to ACR-368.
  • Cross-site collaboration shifts from email and file sharing to structured data queries — scientists in Lund and Watertown work from the same data platform rather than from different versions of the same dataset.
  • The unified data fabric provides the foundation for any future expansion — whether into new therapeutic areas, new imaging modalities, or new partnership arrangements — without requiring a new data architecture.

  We've done this before

  • Transforming Raw Bioprocess Data into Actionable Insights
  • Creating a Digital Lab for a Global Biotech Player

  See how we do Digital Lab