FDA–EMA Good AI Practice for Pharma and Biologics

Executive summary

AI in pharma and biologics earns trust when a scientific team can explain what a model is intended to do, why its evidence is relevant, where it may fail and who is responsible for acting on its output. These questions connect investment in artificial intelligence to the practical responsibilities of process development, manufacturing science and technology, quality and regulatory leadership.

The FDA–EMA Guiding Principles of Good AI Practice in Drug Development, published in January 2026, provide a common foundation for this discussion. They cover evidence generation across the drug lifecycle, including manufacturing, and include biological products. They are guiding principles rather than a product certification scheme. [1]

BioMedAna’s approach combines physics, chemistry, biology, process data and machine learning. We develop mechanistic, deterministic and hybrid models for specific modalities, and use explainable AI to support scientific review. Agents help retrieve evidence, coordinate analyses and prepare findings. Human reviewers retain responsibility throughout the workflow, and agents do not execute operational actions. [2–3]

This paper maps that approach to the ten principles and shows how it can support a reviewable quality-prediction workflow. The central proposition is that scientific modeling, evidence management and human oversight should be evaluated together. A useful model must fit the intended question; a useful explanation must help a reviewer assess the result; and a useful agent must remain within the authority granted to it.

For executives, the deployment decision is therefore specific: which workflow, for which process, under which controls, supported by which evidence? BioMedAna’s reported capabilities provide a foundation for answering these questions. Demonstrating suitability remains a task for each deployment, involving both BioMedAna and the customer’s accountable teams.

Scope and evidence. Product descriptions reflect BioMedAna management statements and public product materials. The paper does not report an independent audit, regulatory endorsement or customer performance study. The worked example is illustrative and contains no client information or measured outcome claims.

What FDA–EMA Good AI Practice means for pharmaceutical leaders

The joint principles address human values, proportional risk controls, applicable standards, defined purpose, interdisciplinary expertise, traceable data, sound development, system evaluation, ongoing management and understandable communication. [1] We use these themes as an organizing framework for the implementation analysis below; the detailed recommendations are BioMedAna’s proposed practices, not additional text issued by the agencies.

For a leadership team, the practical benefit is a common vocabulary. Scientific leaders can describe the question and its evidence. Data scientists can explain the model and its uncertainty. Quality teams can assess how outputs enter controlled processes. Technology leaders can examine access, configuration and system boundaries. These perspectives meet at the actual decision being supported.

An AI program can fail at any of these interfaces. A technically strong model may be applied to an unfamiliar process. A scientifically plausible finding may rest on an incorrect source record. A reviewer may accept an output without seeing a material limitation. Evaluating only predictive accuracy leaves these workflow risks unresolved.

How the joint principles differ from FDA AI guidance and Annex 22

The FDA’s January 2025 draft AI guidance focuses on model credibility for a specified context of use when AI produces information supporting regulatory decisions about drug or biological-product safety, effectiveness or quality. It remains labeled draft and not for implementation on the FDA page reviewed. [4]

Draft EU GMP Annex 22 addresses AI in critical manufacturing applications and must be read with its own scope and restrictions. The June 2026 EMA workshop report discusses possible future approaches; it does not itself replace the draft. The joint principles should not be interpreted as overriding either framework or authorizing a specific generative-AI application. [5–6]

For BioMedAna, these documents have complementary roles: the joint principles organize a broad implementation approach, the FDA draft informs a credibility discussion where its scope applies, and Annex 22 provides a focused European GMP reference. The evidence needed for a deployment follows its actual role and applicable obligations.

BioMedAna’s scientific approach to AI in biologics

Biologics, also described as biological medicines or biologicals, require analyses that preserve the meaning of the process and product being studied. BioMedAna’s modality-specific approach is intended to retain that context while allowing teams to work through a shared platform. [3]

Our portfolio comprises eight groups: monoclonal antibodies; recombinant proteins; microbial and precision fermentation; chemical and pharmaceutical APIs; cell-culture viral vaccines; cell therapies; gene therapies and viral vectors; and RNA/oligonucleotides. These are portfolio groupings, not a claim that the underlying products or regulatory pathways are interchangeable. [3]

Mechanistic and hybrid models make assumptions reviewable

Mechanistic modeling expresses selected scientific relationships. A hybrid approach combines scientific structure with patterns learned from data. Deterministic execution concerns whether a fixed configuration produces repeatable outputs for the same inputs. These properties can coexist, but they answer different questions.

Consider the difference between predicting a process trajectory and recommending a change to that process. A model may predict well within familiar operating conditions while having weak support for an intervention outside them. Scientific structure can help expose that boundary. Its presence does not establish that every proposed change is safe or that the model transfers to another modality.

For each use, we recommend identifying which relationships are represented, which parameters are estimated, which assumptions are imposed and which effects remain outside the model. Where a hybrid model is claimed to improve performance, the comparison should use a relevant baseline and evidence appropriate to the intended workflow. The improvement should be demonstrated rather than inferred from the model’s architecture.

Explainable AI adds a review step with scientific substance

BioMedAna supports scientific review of model behavior. The reviewed Twin implementation exposes global feature importance; it should not be described as per-prediction SHAP attribution. Where an appropriately evaluated SHAP implementation is available for a particular model, it can help identify features contributing to a prediction. [3] It does not by itself establish that changing a feature causes a corresponding change in the real process; confounding and relationships among inputs can affect interpretation. [7]

A scientist should therefore examine an attribution alongside the underlying measurements and scientific assumptions. If a prediction appears to be driven by an implausible feature, that is a reason to investigate the model or data. An explanation is useful when it helps the reviewer challenge a result and decide what additional evidence is needed.

Mapping the ten FDA–EMA principles to BioMedAna

The numbered themes below are paraphrased navigation labels corresponding to the ten principles in the source document. [1] The mapping separates company-reported capabilities from evidence that should be examined for a particular deployment. A feature being implemented does not establish that every use of it is adequate.

The third and fifth rows illustrate why software capabilities alone cannot complete this mapping. Applicable obligations depend on the deployment, and the customer must assign competent people to accountable roles. BioMedAna supplies capabilities and relevant product evidence; the regulated organization retains responsibilities for its processes and decisions.

Illustrative workflow for AI quality prediction in biologics

Illustrative example only. The following describes a proposed application pattern for a monoclonal-antibody process. It is not a statement that BioMedAna has validated the particular target described, and it contains no customer data, sample counts or performance results.

1 Define the scientific question and decision

A process-development team wants to explore whether available process measurements can estimate a selected product-quality attribute before the corresponding analytical result is available. For illustration, the candidate target is a specified glycan proportion measured by the team’s reference analytical method.

The intended role is to help scientists prioritize development investigations. The team explicitly excludes batch release, replacement of required analytical testing and autonomous changes to manufacturing conditions. If it later proposes a different role, it must assess that new use rather than extend the initial conclusion by assumption.

The target also defines a timing question: which inputs genuinely exist when the prediction is to be made? Measurements available only later must not silently enter an early-prediction model. Otherwise, a convincing retrospective result may not represent the proposed workflow.

2 Establish the evidence and preserve its context

The team identifies relevant batches, process measurements and reference results, retaining their product and process context. It examines missing values, corrected measurements and differences in sampling or analytical methods. A data preparation record explains how these issues were handled.

The evaluation design should address related observations. Measurements from the same batch or closely related campaign can create misleading evidence if they are divided across development and testing without regard to dependence. The split should reflect the future setting the team intends to assess. A separate question is whether any site or scale transition requires additional evaluation.

These are proposed design considerations. The paper does not claim that a particular dataset, analytical method or split has already been evaluated for this example.

3 Compare models against a meaningful baseline

The team considers a simple baseline and candidate scientific or hybrid models suitable for the available evidence. It documents why the selected inputs and model structure are reasonable for the question. A hybrid model is useful only if its behavior and evidence justify its role in this workflow.

For a continuous prediction, illustrative evaluation measures could include absolute error, systematic bias and the coverage of a prediction interval. The team should choose tolerances based on what errors would mean for the intended investigation decision. This paper deliberately supplies no universal threshold: the appropriate value depends on the attribute, measurement method and use.

4 Review the prediction and its limitations together

A review package identifies the model version and inputs, displays the estimate and uncertainty, and flags any limitation relevant to the current case. The scientist inspects the source data and available explanation. An unexpected feature contribution prompts a question, rather than becoming a causal explanation by default.

If the input context is poorly represented in the supporting evidence, the appropriate response may be to withhold reliance on the estimate and wait for additional information. The interface should make that path understandable and practical.

5 Let the agent prepare work for a human decision

An agent can organize the evidence, coordinate an authorized analysis and draft a comparison for review. It can also prepare questions for follow-up work. The scientist determines whether the finding merits further investigation or a proposed experiment. Any approved activity proceeds through the organization’s authorized process.

The control pattern is shown below. It is a proposed representation of BioMedAna’s stated human-review boundaries, not a verified software architecture diagram.

6 Evaluate continued usefulness

The team compares predictions with reference results as those results become available and reviews whether the use remains within its established boundaries. A change in process conditions, measurement method or model configuration should prompt a documented assessment of relevance.

The retained record connects the scientific question, data, model, findings, reviewer and next step. It enables a later reviewer to understand what was known when the decision was made. The objective is a defensible chain of reasoning, with visible uncertainty and accountability.

Applying the approach to other pharmaceutical AI workflows

Process optimization

For process optimization, the central distinction is between predicting an outcome and demonstrating the effect of a proposed intervention. We recommend presenting the objective alongside product-quality constraints, the model’s supported operating domain and the evidence needed to test a candidate scenario. A scientist can then decide whether to propose an authorized experiment.

BioMedAna’s modality-specific models provide a basis for considering relevant scientific relationships. The experiment or approved change process supplies evidence about the proposed intervention. A simulated improvement should remain clearly labeled as a prediction until demonstrated.

Deviation investigation

For deviation investigation, the agent can organize a chronology and prepare candidate explanations from authorized records. The reviewer must be able to distinguish direct observations, analytical inferences and unresolved hypotheses. A plausible narrative is insufficient if the source evidence does not support it.

We recommend requiring traceable support for material findings and retaining competing explanations where the evidence is inconclusive. The investigator and quality organization own the conclusion and any corrective action. Agent assistance can improve preparation of the review package while leaving that accountability intact.

A decision record for the quality-prediction review

The principle mapping becomes useful when a reviewer can inspect one complete decision record. The following is a proposed worksheet for the development example, not a completed validation record or an additional agency requirement. Its purpose is to expose the evidence that could change the next scientific action.

A useful review does not end with a signature. If an input belongs to the wrong process version, the record returns for correction. If the estimate lies outside the evaluated region, the team withholds reliance and chooses the agreed fallback. If the evidence supports a development study, the authorized owner carries that proposal into the established study route. These branches make the controls observable in a demonstration.

For an executive sponsor, the first assessment should compare the current and proposed workflow on time to a reviewable decision, reviewer corrections, unresolved evidence and the consequence for the next study. A shorter preparation step is not sufficient if it increases the effort or uncertainty at review. Quantified benefits need a measured baseline and follow-up evidence.

Making human oversight effective in agentic AI

BioMedAna reports that agents operate with guardrails, context boundaries, auditability and human review throughout their work. [3] The practical strength of those controls should be assessed in realistic conditions, including situations in which the output appears persuasive but is wrong.

An evaluation should test whether the reviewer has the time, competence and source access needed to detect an error. It should also examine whether the system exposes its limitations clearly enough for the reviewer to act. Recording an approval does not, by itself, show that the underlying work received adequate scrutiny.

Useful proposed challenge scenarios include an input from the wrong product context, an absent source record, conflicting measurements and a request beyond the agent’s authorized tools. The assessment should show whether boundaries hold and whether the user understands what prevented completion. These are recommended exercises, not assertions about tests already performed.

The no-operational-action boundary is valuable, but recommendations can still influence consequential decisions. BioMedAna and the customer should therefore evaluate the significance of the output, not only whether the agent can execute a command.

What executives should request before deployment

A focused assessment can begin with six questions:

These questions can guide a demonstration without requiring a prospective customer to disclose proprietary information. BioMedAna can use an illustrative process to show the workflow, then agree the evidence needed for a customer-specific evaluation under appropriate confidentiality arrangements.

From principles to evidence for a particular use

BioMedAna’s contribution is the combination of modality-specific scientific models, explainable analytics, governed information and agents that assist accountable people. The joint FDA–EMA principles provide a useful structure for explaining that approach to scientific, quality and executive audiences.

The strongest demonstration is a complete workflow: a well-framed question, relevant evidence, an appropriate model, an understandable result and a documented human decision. That demonstration should make the limits of the analysis as visible as its capabilities.

For an organization evaluating AI in pharma or biologics, the next step is to select one priority workflow and review the scientific and governance evidence together. BioMedAna’s process optimization, quality prediction and deviation investigation capabilities provide starting points for that discussion.

Discuss a BioMedAna workflow evaluation: Visit BioMedAna.

References and basis of assessment