Precision medicine requires computational methods that can integrate genomic, transcriptomic, proteomic, metabolomic, epigenomic, single-cell and spatial measurements into decisions about individual patients, and artificial intelligence has become the enabling technology for doing so. This review argues that the binding constraint is no longer modelling capability but validation, calibration and governance. We compare seventeen multi-omics integration algorithms on the sample sizes they actually require and on whether independent groups have reproduced them; we set classical machine learning against deep learning by omics task and sample-size regime, and find that penalised regression and tree ensembles remain competitive wherever the number of samples is small relative to the number of features. We then examine seven documented failures of deployed clinical artificial intelligence, trace each to its root cause, and derive an eighteen-item appraisal checklist adapting existing reporting and risk-of-bias instruments to the failure modes of molecular data. Calibration, uncertainty quantification, batch effects and information leakage are treated as first-class problems rather than caveats. To show what leakage costs, we analysed 696 breast tumours with matched transcriptomic and copy-number profiles under randomly permuted labels, where the only honest result is chance. A pipeline that selects features before splitting the data reports an area under the receiver operating characteristic curve of 0.95 in cohorts of forty and 0.65 on the full cohort; the corresponding leak-free pipeline returns 0.50 at every size. What limits clinical adoption is the evidence a model can be held to, not the sophistication of the model.

Integration of artificial intelligence and multi-omics for precision medicine

Carlo Cenciarelli
Secondo
Writing – Review & Editing
;
David Vagni
Ultimo
Data Curation
2026

Abstract

Precision medicine requires computational methods that can integrate genomic, transcriptomic, proteomic, metabolomic, epigenomic, single-cell and spatial measurements into decisions about individual patients, and artificial intelligence has become the enabling technology for doing so. This review argues that the binding constraint is no longer modelling capability but validation, calibration and governance. We compare seventeen multi-omics integration algorithms on the sample sizes they actually require and on whether independent groups have reproduced them; we set classical machine learning against deep learning by omics task and sample-size regime, and find that penalised regression and tree ensembles remain competitive wherever the number of samples is small relative to the number of features. We then examine seven documented failures of deployed clinical artificial intelligence, trace each to its root cause, and derive an eighteen-item appraisal checklist adapting existing reporting and risk-of-bias instruments to the failure modes of molecular data. Calibration, uncertainty quantification, batch effects and information leakage are treated as first-class problems rather than caveats. To show what leakage costs, we analysed 696 breast tumours with matched transcriptomic and copy-number profiles under randomly permuted labels, where the only honest result is chance. A pipeline that selects features before splitting the data reports an area under the receiver operating characteristic curve of 0.95 in cohorts of forty and 0.65 on the full cohort; the corresponding leak-free pipeline returns 0.50 at every size. What limits clinical adoption is the evidence a model can be held to, not the sophistication of the model.
2026
Istituto per la Ricerca e l'Innovazione Biomedica - IRIB - Sede Secondaria Messina
Istituto di Fisiologia Clinica - IFC - Sede Secondaria di Roma (soppressa)
Artificial intelligence
Clinical validation
Data leakage
Model calibration
Multi-omics
Precision medicine
File in questo prodotto:
Non ci sono file associati a questo prodotto.

I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.

Utilizza questo identificativo per citare o creare un link a questo documento: https://hdl.handle.net/20.500.14243/608641
 Attenzione

Attenzione! I dati visualizzati non sono stati sottoposti a validazione da parte dell'ente

Citazioni
  • ???jsp.display-item.citation.pmc??? ND
  • Scopus 0
  • ???jsp.display-item.citation.isi??? 0
social impact