Skip to content

Services · Bioprocess modelling

From a process question to a model your QA can sign off

We build mechanistic, statistical, machine-learning and hybrid models of bioprocesses, and take them through a staged qualification path — from a first feasibility check to a monitored model in production.

Book a technical call How we work

When teams call us

Design space in R&D

Parameter relations are too complex for simple regression, and covering the space by experiment alone is too slow.

Scale-up & tech transfer

You need a model built at small scale that carries to the next vessel or site.

Recurring deviations

The same deviation keeps coming back and CAPA hasn’t fixed it for good.

Continuous & real-time release

Moving to continuous manufacturing (ICH Q13) or real-time release needs tighter, model-based control.

Not every process problem is a modelling problem. If a sensor calibration or a procedure fix solves it, we’ll tell you in the first call.

Our framework

Four stages, each closed by a signed gate

The lifecycle GMP applies to equipment and processes, applied to the model itself. A working model is never, on its own, a reason to move to the next stage.

  1. Stage 0 · 8–12 weeks

    Feasibility study

    Is this a modelling problem, and is it worth a full cycle? Four qualifying questions, an initial ICH Q9 risk check, a data audit and a feasibility model built from the simplest baseline.

    Gate output

    Feasibility report, updated URS, go / no-go

  2. Stage 1 · Process design

    Understanding before data

    Mechanistic understanding and DoE, an ALCOA+ data audit, a documented architecture choice, and acceptance thresholds set with your SME and QA before any training.

    Gate output

    Frozen Model Design Specification (MDS)

  3. Stage 2 · DQ → IQ → OQ → PQ

    Qualification

    Design review, reproducible training, validation on data never seen in training, uncertainty quantification and prospective runs in shadow or advisory mode.

    Gate output

    Signed PQ report and documentation package

  4. Stage 3 · Whole model life

    Continued verification

    Production deployment with a rollback path, a monitoring panel, drift detection, planned retraining under change control and a business-value review.

    Gate output

    Monitoring panel, retraining plan, Model Card

Start at any stage. If you already have a model, we pick up where it is and close the gaps.

In every project

Six rules that keep a model auditable

Criteria before results

Acceptance thresholds are fixed before training — metric, threshold, dataset, method — and don’t move in Stage 2.

Thresholds from four sources

Process specification width, reference-method uncertainty, the existing baseline and the cost of an error, agreed with your SME and QA.

Batch-level data splits

Training, validation and test sets are split by batch or run, never by time point, so results aren’t inflated by leakage.

Uncertainty in numbers

Every prediction carries an interval and an out-of-distribution flag. We report “RMSE vs. threshold”, not “works well”.

Your SME as co-author

Operator heuristics become explicit model inputs, credited in the documentation. QA is asked on day one what would block qualification later.

Static models, planned retraining

For GMP use the model is frozen between retrainings, and each retraining is a documented change, as the EU GMP Annex 22 draft favours.

Model families

The simplest model that meets the criteria

The architecture follows what you know about the mechanism and the data you have. Every choice is documented with the alternatives we rejected.

Mechanistic

Mass balances, Monod kinetics, kLa correlations. Best when the mechanism is known and data is scarce, including cold starts.

Statistical / DoE

DSD, CCD, PLS. The family regulators already know from QbD and design-space work.

Data-driven / ML

Gradient boosting, GRU and other sequence models, where data is dense — online sensors or in-line spectroscopy.

Hybrid

A physical backbone plus a learning component on the residual. Interpretable and workable with few batches, which is why it most often pays off.

Ways to work with us

Pick the engagement that fits where you are

2–3 weeks

Model & data audit

An independent review of an existing model or dataset against the framework: data integrity, splits, criteria, documentation gaps.

You get: Audit checklist and a gap-closing plan

8–12 weeks

Feasibility study (Stage 0)

Scoping, URS, data audit, a feasibility model and an evaluation against criteria set before seeing any results.

You get: Feasibility report, updated URS, go / no-go

Scoped per project

Model design (Stage 1)

Process understanding, DoE design, architecture choice and acceptance criteria, up to a frozen MDS.

You get: MDS and a Stage 2 plan

Scoped per project

Build & qualification (Stage 2)

Reproducible training pipeline, OQ validation with uncertainty quantification and prospective PQ runs on your process.

You get: PQ report, Model Risk Assessment, documentation package

Ongoing

Model lifecycle care (Stage 3)

Deployment, monitoring, drift response and planned retraining under your change control.

You get: Monitoring panel, retraining schedule, Model Card

Scoped per project

Models on the MODICA twin

Run your process models live in MODICA ADT, next to your bioreactors.

You get: Your model running live, with simulation and replay

Documents you keep

Written to survive an inspection two years from now

  • User Requirements Specification (URS)
  • Model Design Specification (MDS)
  • Model Risk Assessment
  • PQ protocol and report
  • Audit checklist
  • Model Card
  • A project plan per stage, with gates and roles

Regulatory frame

Aligned with the guidance your QA works to

  • ICH Q8, Q9, Q10, Q13: QbD, quality risk management, PQS, continuous manufacturing
  • GAMP 5: software category and proportionate validation rigour
  • 21 CFR Part 11 / EU Annex 11: audit trail, access control, e-signatures
  • EU GMP Annex 22 (draft): AI in GMP manufacturing
  • EMA Reflection Paper on AI: and the FDA/EMA Good AI Practice principles

R&D and production use the same framework with different formal rigour. We never let an informally validated R&D model drive production decisions.

Service areas

Bioprocess modelling

Mechanistic, statistical, ML and hybrid models of your bioprocess, taken from a feasibility study to a qualified, monitored model through our staged framework.

Learn more

Connectivity & integration

Connecting your bioreactors and plant data to the digital twin, and building the adapter your equipment needs.

Learn more

Decision support

Turning what the twin knows into guidance operators and process engineers can act on.

Learn more

Book a 30-minute technical call

Bring your process question. We’ll tell you whether it’s a modelling problem, which stage to start at, and what data you’d need.

Book a technical call

Opens your e-mail app with a short checklist. Or write to hello@modica.bio.