Large Language Models Could Help Detect Pharmacogenomic Phenoconversion

A 2026 perspective in Frontiers in Genetics argues that large language models could help clinicians identify pharmacogenomic phenoconversion, the mismatch between a patient’s genotype-predicted metabolic capacity and the metabolic function present at the time of treatment. The proposed role is not to replace established precision-dosing models or make autonomous prescribing decisions, but to retrieve, organize and temporally contextualize information distributed across electronic health records. By connecting genetic results with medication use, laboratory findings, diagnoses, clinical notes and environmental exposures, the authors say, LLM-supported systems may flag clinically important possibilities that are difficult to identify through static genotype interpretation alone.

Pharmacogenomics has demonstrated clinical relevance for selected drug–gene pairs involving genes such as CYP2D6, CYP2C19, CYP2C9, DPYD, SLCO1B1 and NAT2. Genetic variation can influence drug disposition, metabolism, efficacy and the risk of adverse reactions, leading organizations including the Clinical Pharmacogenetics Implementation Consortium and the Dutch Pharmacogenetics Working Group to develop prescribing recommendations. Resources such as PharmGKB, PharmVar, ClinVar and ClinGen support evidence curation, variant interpretation and standardized nomenclature. Implementation nevertheless remains uneven, and even when a guideline exists, inherited genotype does not represent every influence on a patient’s response to a medicine.

The central problem is that enzyme activity is not necessarily fixed by genotype. Concomitant medicines, inflammation, liver dysfunction, aging, pregnancy, smoking and other physiological or environmental conditions can modify metabolic pathways. Strong inhibitors of CYP2D6, including paroxetine, fluoxetine, bupropion and quinidine, can reduce activity to levels comparable to those of poor metabolizers. CYP2C19 and CYP3A activity may also change in response to medications and clinical conditions. Inflammation-mediated suppression of CYP expression is particularly relevant in infectious disease, autoimmune disorders, cancer and critical illness. As a result, a genetically classified normal metabolizer may currently resemble an intermediate or poor metabolizer, while induction can produce the opposite change.

Phenoconversion is especially consequential for medicines with narrow therapeutic indices or those requiring metabolic activation, including tamoxifen, codeine, tramadol and numerous psychotropic agents. Inhibition can increase exposure and produce unexpected toxicity, whereas induction can reduce exposure and contribute to treatment failure. Large cohort evidence cited by the review indicates that concomitant medication use and other non-genetic factors substantially increase the proportion of patients with clinically actionable phenotypes compared with genotype-only prediction. These findings support a dynamic view of metabolic capacity, but they do not mean that every apparent genotype–phenotype discrepancy qualifies as clinically meaningful phenoconversion.

The review locates the potential value of LLMs in a clinical knowledge-integration problem. Medication records, laboratory measurements, diagnostic codes, imaging reports and unstructured notes can all contain modifiers of drug metabolism. Information such as over-the-counter drug use, supplements, smoking cessation, an inflammatory flare or a temporary physiological change may appear only in narrative text. Traditional rule-based systems can struggle when evidence is embedded in free text or must be aligned with a particular treatment period. An LLM might extract these signals, connect them with the patient’s pharmacogenomic report and present them in a form that allows a clinician to evaluate whether the predicted phenotype may no longer reflect current function.

Temporal reasoning would be central to that task. Finding an inhibitor anywhere in a longitudinal record does not establish that it affected metabolism when a medicine was prescribed. Similarly, an infectious or inflammatory condition may coincide with a change in exposure, adverse effects or clinical response without proving that it caused the change. An LLM-assisted workflow could reconstruct when exposures began and ended, distinguish current from historical factors, and relate them to treatment events. The resulting output would identify potential phenoconversion for review rather than confirm it, which is a critical distinction because functional status can be checked through independent measures such as probe-drug testing, therapeutic drug monitoring or a validated biomarker.

The authors propose that LLM outputs could also become inputs to dynamic phenotype models that update metabolic estimates as a patient’s circumstances change. Conventional pharmacogenomic reports generally assign discrete categories such as poor, intermediate, normal or ultrarapid metabolizer status. Dynamic approaches instead combine genotype with current medicines, inflammatory status, organ function, lifestyle and other covariates to estimate the probability that a patient presently falls within a particular metabolic range. These proposals overlap with model-informed precision dosing, population pharmacokinetics, Bayesian forecasting and physiologically based pharmacokinetic modeling. The defensible division of labor is for LLMs to retrieve and structure clinical inputs, while validated pharmacokinetic and pharmacodynamic models estimate exposure, dose requirements and response.

Several scientific and technical barriers remain. LLMs can misinterpret records, generate unsupported statements or make errors in temporal reasoning, and isolated errors may carry clinical consequences. Missing or outdated medication lists, incomplete lifestyle histories and contradictory records can reduce reliability regardless of model sophistication. The magnitude and duration of many non-genetic modifiers are also insufficiently characterized, and routine care lacks a universally accepted reference standard for functional metabolic activity. A system can identify a suspected modifier but cannot derive its effect size, explain an unknown interaction or create a validated dose adjustment when the underlying pharmacological relationship has not been established.

Safe deployment would require validation with real-world clinical data, external testing across institutions and populations, explicit uncertainty estimates, documentation of contributing factors and continuous human oversight. Generalization is a particular concern because electronic record structures, prescribing patterns, language and coding practices vary between health systems, while allele frequencies and genotype–phenotype relationships may differ among ancestry groups. Privacy, cybersecurity, computational cost, model updating and regulatory accountability also complicate implementation. Any system would need to expose supporting quotations or citations and distinguish established evidence from hypotheses, especially when it surfaces unfamiliar combinations of drug–drug–gene or drug–gene–environment interactions.

Future research would need to quantify when changes in metabolic activity meaningfully alter drug exposure, efficacy or toxicity, then test whether LLM-assisted workflows improve medication safety, therapeutic effectiveness, adverse-event detection, healthcare use or patient outcomes. Such systems could integrate with established pharmacogenomic and precision-dosing infrastructure and, over time, analyze large longitudinal datasets to generate hypotheses about previously unrecognized interactions. The perspective therefore presents LLM-enabled phenoconversion detection as a promising but unproven extension of precision pharmacotherapy. Its eventual usefulness will depend not on language-model capability alone, but on pharmacological evidence, complete clinical data, transparent quantitative models and continued oversight by clinicians.

Subject of Research: Potential use of large language models for detecting pharmacogenomic phenoconversion in clinical records

Article Title: The potential role of large language models in addressing phenoconversion in pharmacogenomics

Article References: García-Agúndez, A., & Agúndez, J. A. G. (2026). The potential role of large language models in addressing phenoconversion in pharmacogenomics. Frontiers in Genetics, 17, Article 1987469. https://doi.org/10.3389/fgene.2026.1987469

Image Credits: AI Generated

DOI: 10.3389/fgene.2026.1987469

Keywords: pharmacogenomics, phenoconversion, large language models, precision dosing, drug metabolism, electronic health records, drug interactions, cytochrome P450, clinical decision support, precision medicine, potential, role