AI in drug development: a pharmacist's view on data integrity

Apharmacist learns early that a small discrepancy is never small. Two milligrams off, a label half-read, a batch number that looks right but isn't — these are the events that, traced back, turn out to be the start of something. The same instinct is what most AI in drug development is still missing.

The discussion is usually framed around models: which transformer, which generative architecture, which foundation model trained on which corpus. That is the visible half. The other half — the one that quietly decides whether the model is useful or merely confident — is the data layer underneath. In discovery, that layer is almost never as clean as the slide deck implies.

Where AI is actually moving discovery

Three places are doing real work, not theatre. First, target identification: graph and language models trained on the literature and on multi-omics datasets are surfacing target–disease links that human teams would take quarters to enumerate. Second, molecule generation and property prediction: diffusion and graph-based generators are producing candidate molecules with optimised ADMET profiles, which medicinal chemists then filter. Third, trial design and patient stratification: models trained on real-world evidence are sharpening eligibility criteria so smaller cohorts can produce cleaner signals.

None of this is hypothetical anymore. The economics are also hard to ignore: the industry has spent decades watching the cost per approved drug climb while approval rates stayed flat. Anything that compresses the discovery phase has compounding value downstream.

The data integrity problem nobody puts on the slide

Discovery datasets are stitched together from instruments, CROs, public repositories, decade-old internal projects, and licensed external sources. Each has its own conventions, units, controls, and quietly different definitions of the same word. "Active" in one assay is not "active" in another. A potency value entered in nM in one system and µM in another will not announce itself; it will just bias your model.

Drift is the second layer of the problem. An assay recalibrated after a reagent lot change. An instrument that ages. A protocol revised between Q2 and Q3. These are normal — they are good laboratory practice. The training set, however, treats them as one population. So the model learns a small lie and rewards it with low loss. Months later, the wet lab spends weeks chasing a ghost hit, and nobody connects the dots back to the lot change.

This is the part regulators care about. The FDA's ALCOA+ principles for data integrity (Attributable, Legible, Contemporaneous, Original, Accurate, plus Complete, Consistent, Enduring, Available) were written for paper records and then for electronic ones. They were not written with training pipelines in mind, but they translate cleanly: an AI prediction inherits the integrity of the data it learned from, and that lineage must be reconstructable.

What a pharmacist's eye catches

Pharmacy training rewards a specific kind of suspicion: the assumption that the label is wrong until the chain of custody says otherwise. Move that instinct into a drug-discovery ML pipeline and three checks become non-negotiable.

One — units, controls, and provenance are part of the schema, not metadata bolted on afterwards. If a record cannot answer "which assay, which protocol version, which operator, which instrument, which lot" in a query, it is not a training row yet. It is a candidate.

Two — drift detection runs on inputs, not just outputs. By the time prediction quality degrades enough to trip a downstream alert, the wet lab has already moved on bad guidance. The cheaper place to catch drift is at ingestion: distributions of assay readouts by batch, by week, by site. A pharmacist would never accept a stock bottle without looking at the seal.

Three — every prediction the model emits carries a confidence interval and a data-lineage hash. If a chemist cannot, in one click, see which experiments taught the model this answer, they will, correctly, distrust it. Trust is not a UI problem. It is a traceability problem with a UI on top.

A practical checklist before the model

• Map every data source to its system of record, its owner, and its last protocol revision date.
• Normalise units and controls at ingestion, not in the notebook. Reject silently-coerced values.
• Tag each record with assay version, instrument ID, operator, and date. Treat untagged records as quarantine.
• Hold out a temporal validation set — the last quarter, not a random split. Random splits hide drift.
• Run input-distribution monitors on the live feed. Alert on shift before alerting on prediction quality.
• Make data lineage queryable from any prediction. If the chemist cannot trace it, the model cannot defend it.

The compounding effect

None of this is glamorous. It does not feature on conference slides or in funding rounds. It is closer to the work of a QA pharmacist than to the work of an ML researcher — and that is exactly why it matters. The teams that compound advantage in AI-driven discovery over the next five years will not be the ones with the biggest models. They will be the ones whose training data still tells the truth in year three.

A small discrepancy, caught early, is a saved quarter. Caught late, it is a withdrawn paper, a failed Phase I, or a regulatory letter. The instinct to chase the small one is the most underrated edge in the field.

Efthymios Vogiatzis is a pharmacist turned product leader in AI and life sciences, based in Vienna. He has led data and AI work across Thermo Fisher Scientific, Takeda, Swiss Re, and Johnson Controls. Get in touch.