03 Nov 2026 · 5 min read

Precision Medicine and AI: The Gap Between Promise and Practice

Precision medicine has transformed cancer treatment in specific contexts. In most common diseases, the promised era of molecularly tailored therapy has not arrived - and the reasons are structural, not just technical.

Precision Medicine and AI: The Gap Between Promise and Practice

In 2001, imatinib was approved for chronic myeloid leukaemia. It was the first drug designed to hit a specific molecular target - the BCR-ABL fusion protein produced by the Philadelphia chromosome - and its effect was dramatic: five-year survival rates in CML went from approximately 30% to above 90% in treated patients. This was the proof of concept for precision oncology. It was also, in retrospect, an unusually tractable case: a single causative mutation, a tractable pharmacological target, a disease defined by that mutation. Most diseases are not like this.

Where Precision Medicine Has Worked

Oncology is where the precision medicine model has delivered most convincingly. The identification of EGFR mutations in non-small cell lung cancer, HER2 amplification in breast cancer, BRAF V600E in melanoma, and ALK rearrangements in lung cancer has enabled the development of targeted therapies with substantial response rates in biomarker-selected populations. Companion diagnostics - tests that identify which patients should receive a given therapy - are now a regulatory requirement for certain drug approvals, and the FDA's biomarker qualification programme provides a pathway for formalising the evidentiary standard for new biomarkers.

The common thread in success cases is a high-penetrance variant with a clear causal role in the disease, a druggable target, and a clinical population where the biomarker is enriched enough to make a biomarker-selected trial feasible. These conditions are met in a relatively small proportion of cancer subtypes and almost never in common complex diseases.

Why It Has Not Worked in Common Complex Diseases

Common complex diseases - type 2 diabetes, coronary artery disease, hypertension, most psychiatric conditions - are driven by the combined effects of hundreds or thousands of genetic variants, each contributing a tiny fraction of disease risk, interacting with each other and with environmental exposures in ways that are not well understood. The polygenic architecture of these diseases means that precision medicine in the oncology sense - find the mutation, target the mutation - does not translate.

Genome-wide association studies have identified thousands of variants associated with common diseases, but the individual effect sizes are typically so small that knowing someone's genotype for these variants adds only modest predictive power beyond conventional clinical risk factors. The polygenic risk score (PRS) for coronary artery disease, one of the most developed in the field, improves risk stratification somewhat beyond traditional risk calculators - but the incremental benefit in clinical decision-making has been modest and its implementation in clinical practice has been limited.

Pharmacogenomics: Implementation Failures

Pharmacogenomics - using genetic variants to predict drug response - has a longer history than polygenic risk scores and a similarly underwhelming record of clinical implementation. The CYP2C19 genotype predicts clopidogrel metabolism, and carriers of loss-of-function alleles have worse cardiovascular outcomes on standard clopidogrel dosing than on alternative antiplatelet strategies. The FDA added a boxed warning to clopidogrel based on this pharmacogenomic data. Yet clinical uptake of CYP2C19 testing before prescribing clopidogrel remains low, despite a decade of guidelines supporting it.

The barriers to pharmacogenomics implementation are not primarily technical - the tests exist, the data exists, the guidelines exist. They are organisational, economic, and workflow-related. Genetic testing requires integration into clinical workflows, interpretation tools embedded in the electronic health record, and clinical decision support that presents actionable recommendations at the point of prescribing. These are implementation science problems, not genomics problems, and they have proven harder to solve than the science itself.

The Validation Gap in Clinical AI

AI-based tools for patient stratification, treatment selection, and outcome prediction have proliferated rapidly. The validation gap - the distance between a model that performs well in a retrospective development dataset and a model that improves outcomes in a prospective clinical deployment - is large in most cases. Models trained on single-centre data frequently degrade when deployed in different health systems with different patient populations, documentation practices, and clinical workflows. A model that predicts sepsis onset with high AUC in a hospital's own historical data may perform substantially worse in a different hospital.

The standard for clinical AI validation is undergoing rapid evolution, but the field is still catching up with the deployment reality - that models built from observational data, however sophisticated, reflect the biases and limitations of the data they were trained on.

A Realistic Assessment of the Next Decade

The next decade in precision medicine will likely see continued progress in oncology, particularly in early detection through liquid biopsy and in expanding the range of biomarker-drug pairings in established tumour types. Multi-omics approaches - integrating genomics, transcriptomics, proteomics, and metabolomics - will provide richer characterisation of disease subtypes in some areas. In common complex diseases, pharmacogenomics implementation will probably advance faster than novel biomarker discovery, because the implementation barriers are at least clearly defined.

What the next decade will not bring is the broadly personalised medicine vision of the early 2000s - a world where each patient's molecular profile determines their treatment across all disease areas. The biology is more complex, the implementation barriers more fundamental, and the evidence base for many precision medicine strategies less robust than the promotional narrative of the field has suggested. At NousLab, honest evidence synthesis is how we contribute to a more calibrated picture of what is known and what remains to be established. Get in touch to discuss evidence reviews in precision medicine.

Jesus Arias
Jesus Arias
Founder & CEO at NousLab

Join the Research Revolution

See how NousLab can accelerate your organization's research

AI