mAb/ATMP bioreactor control — anticipate drift minute by minute and hold per-patient traceability.

The problem

A bioreactor does not forgive — and the critical data sits scattered across four systems.

Producing mAb or ATMP in a bioreactor (single-use or stainless steel, allogeneic or autologous) means living with three pressures at once:

1. Minute-by-minute process control. The bioreactor holds a living culture — CHO cells producing mAb, T cells in autologous expansion, iPSCs in a master bank. A pH or DO drift that goes unnoticed for 4 hours degrades the culture and compromises the batch. The DCS has alarms, but alarms are reactive — they fire once the damage has started.
2. Critical data in islands. The CPPs are in the bioreactor DCS; the offline CQAs (HPLC, MALS, glycan analysis) are in the LIMS; the feed batch parameters are in the process engineer's spreadsheet; the shift observations are in the operator's notebook; the apheresis acknowledgements for the CAR-T patient are in the donor hospital's system. Cross-referencing all of that by hand is what the QA person does after the batch — not before, when there was still time to intervene.
3. Non-negotiable per-patient traceability. In autologous production, breaking the vein-to-vein custody chain means losing the patient — not just the batch. Correctly identifying which cell bag belongs to which donor is an absolute human responsibility, and today it rests on a label plus a checklist plus a second signature at every point. It works — and every error makes headlines.

The result: a typical lost-batch rate of 5-15% caused by drift not detected in time, and a custody-chain error rate that trends toward zero but never gets there. Every mAb batch costs tens to hundreds of thousands of euros; every autologous ATMP batch, even more. The classic setup (operator + DCS + checklist) works — and every failure is either hard cost or irreparable harm to a patient.

How it fits the IRIS system

iLEAN does not replace the DCS — it anticipates drift and holds per-patient traceability without touching it.

Bioreactor control suffers because the data lives in islands (DCS, LIMS, feed spreadsheet, notebook) and because alarms are reactive. iLEAN acts as the putty between the bioreactor DCS, the LIMS, the ERP, the notebook and the hospital's apheresis system — without touching the validated DCS, without replacing the LIMS. It captures continuously, cross-references CPPs and CQAs, anticipates drift using the batch's real data and holds per-patient traceability with a Part 11 audit trail.

Edge sees and captures at the bioreactor. Brain cross-references CPPs and CQAs and anticipates drift. Tracer holds per-patient traceability. The person signs any setpoint change — the AI proposes with evidence.

The three iLEAN pieces applied to mAb/ATMP bioreactor control:

• iLEAN Edge — a local terminal in the culture suite. It reads the bioreactor DCS over OPC-UA, ProfiNet or Modbus, integrates with online probes (pH, DO, glucose, lactate, NIR biomass, Raman where available), and applies vision to the bioreactor panel or to the SUB bag when integration does not reach. It works without a network. Inference happens in the box — the data never leaves the OT ring.
• iLEAN Brain (Central + Agents) — cross-references the CPPs from Edge with the CQAs from the LIMS (HPLC, glycan, host cell protein), the feed batch parameters from the spreadsheet and the observations from the notebook. The agent trained on your culture (CHO mAb, autologous T cell, iPSC, gene-modified ATMP) anticipates drift before the DCS alarm fires: it spots the Crabtree pattern, the lactate rise that precedes the viability drop, the osmolality drift that is about to compromise titer. It proposes the adjustment to the supervisor; QA signs.
• iLEAN Tracer — holds per-patient traceability for autologous products (vein-to-vein: donor apheresis → cryogenic transport → bioreactor → expansion → harvest → cryopreservation → shipment → infusion) and MCB batch traceability for allogeneic ones, with a Part 11 audit trail and validation at every transfer point. Patient identification is verified by two factors at every step; the system does not move forward without the human signature that belongs there.

See the full IRIS architecture →

Before and after

Classic bioreactor control vs. control with iLEAN

Aspect	Classic control (DCS + LIMS + spreadsheet)	With iLEAN Edge + Brain + Tracer
Process alarms	Reactive — they fire once damage has started	Drift anticipation with CPPs and CQAs cross-referenced
CPP + CQA cross-reference	Manual, after the batch	Continuous, during the culture
Feed batch spreadsheet	Shared folder, the "good" version	Captured at second zero, linked to the batch
Autologous traceability	Label + checklist + second signature	Two-factor at every point + Part 11 audit trail
Allogeneic traceability (MCB)	Manual link back to the bank	Automatic chain batch → MCB → WCB → product
Batches lost to drift	5-15% (typical industry range)	≥30% fewer in the first 6-12 months

Impact estimate

Impact estimate for your plant — to be validated with your numbers.

The block below is an estimate to be validated with the specific data of your plant. We put it forward so the committee has an order of magnitude; we refine it during the diagnostic.

• Biotech facility with one or several bioreactors (200L-2000L SUB for mAb, autologous ATMP at patient scale, cell-bank allogeneic ATMP), standard DCS, validated LIMS.
• First value expected within a few weeks: the first bioreactor instrumented with Edge and the drift-anticipation agent live on the Pareto culture.
• Expected reduction in batches lost to drift not detected in time: ≥30% in the first 6-12 months.
• Indicative payback between 4 and 9 months, dominated by the cost of a single recovered batch (tens to hundreds of thousands of euros, depending on the product and the clinical phase).
• Industry reference figure: +40% productivity in the more digitalized sectors compared with the least digitalized ones ^[1]. The gap is widening; whoever does not anticipate loses ground in biotech.

And the process manager's reasonable doubt

“What if the AI proposes a setpoint change and compromises the culture or product integrity?” — the rule is strict: the agent proposes with evidence (which CPPs/CQAs trigger the drift, which setpoint to adjust, within which validated range); the culture supervisor decides; QA signs the setpoint change in the DCS. The three rings exist for exactly this: the bioreactor DCS lives in ring 1 (critical OT) and never accepts inbound connections; the AI observes from ring 3, validates in ring 2, and the person executes. And in anchored tasks (reading a culture pattern and comparing it with the product's golden batch) the best models brought error below 1.5% ^[2].

^[1] +40% productivity in the more digitalized sectors (2000-2021). Source: Fundación BBVA/Ivie.

^[2] OpenAI paper “Why Language Models Hallucinate”, 2025 — on the reliability of AI in anchored tasks.