A promising approach towards personalised care is AI-driven multiomics – the combined analysis of genomic, transcriptomic, proteomic, metabolomic, and other biological layers of information facilitated by AI. As each omic dataset provides a unique lens into patient biology, such combined analyses yield powerful insights that no single dataset can capture. They could be immensely valuable in clinical settings as they hold the potential to inform clinical decisions, for example, in disease risk prediction, disease subtyping, and predicting response to treatment.
At PHG we believe now is the time to prepare for wide scale translation of these important technologies into health systems. This means ensuring that
- AI-driven multiomic solutions in development are addressing real-world clinical need
- There is system-wide readiness to adopt and equitably implement these solutions
Kickstarting the discussion
We convened an expert group of stakeholders to discuss what AI-driven multiomics could look like in healthcare settings, what is needed to support this vision and what challenges are anticipated. This article captures their perspectives and is intended to get the discussion started on integrating AI-driven multiomics in healthcare for personalised medicine. But first – what is ‘multiomics’?
A definition of multiomics
Initially, omics referred to large-scale molecular characterisation of biological systems using DNA (genomics), RNA (transcriptomics), and proteins (proteomics). This landscape expanded to include metabolomics, epigenomics, and lipidomics, among others. More recently, physiological and imaging-derived data have inspired new fields such as radiomics and pathomics, which extract quantitative biomarkers from radiological and histopathological images, respectively. One might also come across new emerging subfields like cardiogenomics and neurogenomics which illustrate how omic technologies are being tailored to specific organ systems and clinical domains.
We define multiomics as the use of more than one type of omic data to make biological inferences, which can be achieved either by combined analysis or via one-at-a-time addition of each omic layer.
How AI can help with the data deluge
Each omic layer generates enormous volumes of data making the analysis far beyond the capacity of traditional or manual analytical methods. Often omic data are in incompatible formats to each other, which further complicates analysis.
. This is where artificial intelligence (AI) is poised to play a transformative role as it can help integrate, analyse and make sense of complex multiomic data.
AI can support multiomic analysis mechanistically and generatively:
- Mechanistically: When two patient groups respond differently to the same treatment, AI-driven multiomics can help identify the biological mechanisms that distinguish them, going beyond what genomics alone can reveal.
- Generatively: AI models can detect or even hypothesise new patterns across large datasets, as seen in digital pathology tools such as PathChat and HistoGPT, which assist in analysing histology images and generating structured pathology reports.
To maximise the effectiveness of AI tools in health services, it is essential to ensure not only the generation of the necessary amount of data but also the ‘right’ kind of data.
Healthcare data is frequently highly fragmented, stored in silos, captured in inconsistent formats, and are often unstructured particularly within clinical notes. Ensuring systems are in place to generate high-quality, well-annotated datasets and consolidating these for training and validating AI models is critical. Moreover, ongoing efforts to digitise NHS care as part of the NHS 10-year plan is an important step towards enabling AI integration in the health service. It is important to acknowledge that for some services, IT infrastructure is still very basic, making AI integration feel like a distant future.
Case study: GE Edison Health Services Platform
A promising model is the GE Edison Health Services platform, developed by GE Healthcare, the University of Cambridge and Addenbrooke’s Hospital. This AI-enhanced platform integrates clinical, imaging, and genomic data into a single interface for multidisciplinary teams to simultaneously view a patient’s profile and co-develop personalised treatment strategies.
This collaboration produced the Integrated Radiogenomics for Ovarian neoadjuvant therapy (IRON) tool, since renamed CareIntellect for Oncology (CI4O). This tool predicts ovarian cancer treatment outcomes with ~80% accuracy by combining radiomics, genomics, blood tests, and EHRs. It could demonstrate the power of a multiomic approach, however, clinical deployment is still pending beyond this proof-of-principle study. CI4O is currently connected to Cambridge University Hospitals’ EHR via a cloud server and its functionality is being evaluated.
Powerful tools, such as GE Edison Health Services platform, described above, must be used responsibly. Clinical AI models need to be transparent and interpretable models (white-box) so clinicians can understand how a conclusion was reached. AI in healthcare should be a collaborator, not a replacement for human decision-making.
Where white-box models are not possible, there must be sufficient confidence in the models such that regulators and clinicians are assured that the AI is functioning safely and effectively for its intended purpose. It is likely to be important that explanations can be provided about how an AI tool reached its conclusion and that a human can scrutinise aspects of its processing.
Multiomics in clinics
Multiomic approaches hold great promise for improving diagnostic yield. This is especially true for cases where simple, single-omic-based biomarkers do not exist. Decisions about which multiomic analyses should be implemented in clinical settings must be guided by clinical need, not by the availability of technology/research. Innovation in this space should be problem-driven, not research/technology-led.
The role of genomics
Genomics, as the most mature omic discipline with proven clinical impact, is often seen as central to multiomics, providing the biological foundation and translational model for other omic modalities. But does giving it a dominant role risk constraining the development of other omic disciplines? Treating genomics as the blueprint can overlook the distinct data characteristics and utility of transcriptomics, proteomics, and metabolomics. A truly multiomic approach, therefore, requires assigning equal or context-specific value to each dataset.
In practice, genomics remains the most widely adopted omic analysis in healthcare. Within the NHS, transcriptomic testing is already used clinically, but often as a second-line tool deployed when genomic analysis fails to yield conclusive results or to provide additional layers of insight. This hierarchy reflects both technical readiness and economic considerations as large-scale proteomic or metabolomic testing remains more costly and less standardised than genomic sequencing. As such, it is plausible that future deployment of non-genomic omics may, at least initially, be embedded within existing Genomic Medicine Service (GMS) pathways, supported by specialist omic centres.
This landscape is evolving rapidly. For example, Nightingale Health has developed a metabolomic blood-based test that quantifies relative disease risk at a cost lower than current whole-genome sequencing efforts. Such innovations suggest a potential rebalancing of the omic hierarchy where other omics could soon match or even surpass genomics in scalability and clinical value.
So, what are the pressing clinical problems that multiomics might help address?
One example lies in disease monitoring. Unlike genomics, which provides a largely static view of inherited mutations, metabolomic or proteomic assays can be repeated over time to capture dynamic changes in disease progression or treatment response. This makes them powerful tools for longitudinal patient management and adaptive therapeutic interventions.
Another area of need is the interpretation of variants of uncertain significance (VUSs) which are genetic findings whose clinical relevance is unclear. Integrating additional molecular layers, such as transcriptomic or proteomic data, could help clarify whether a given variant is disease-causing or benign. For clinicians, this represents a tangible and immediate benefit. It enables them to make more confident, evidence-based decisions about diagnosis, prognosis, and treatment planning.
What needs to happen?
In addition to AI-driven multiomic solutions that address specific clinical needs, stakeholder specific concerns need to be addressed.
- Clinicians: Time pressures and liability concerns could be key barriers to multiomic adoption. When new technologies emerge faster than clinical consensus can form, clear guidance becomes essential. Without it, clinicians risk being caught between two difficult positions: acting prematurely on inconclusive information or failing to act on data that later proves significant. National or institutional clinical guidance can help define when, where, and how multiomic tests should be used, providing both a framework for adoption and legal protection for practitioners.
- Patients: Public awareness of genomics has grown substantially, making communication of clinical findings to patients that bit simpler. But test results based on probabilities remain complex or ambiguous. These are familiar challenges for genetic counsellors, who routinely help patients interpret such probabilistic findings. Multiomics introduces new layers of complexity. Perhaps multiomic counsellors – trained to communicate uncertainty, risk, and the limitations of emerging evidence – could help patients navigate their multiomic test results.
- Industry: Companies face challenges accessing NHS and academic datasets, which are often fragmented or of variable quality. Moreover, the absence of a clear deployment and procurement pathway makes it difficult for developers to design and sell clinically relevant tools. They need guidance on fundamental questions such as:
- What constitutes a “good” product in the clinical context?
- What level and type of evidence are required to show clinical validity?
- Which clinical use cases are viable and supported?
- Which patient groups are most appropriate?
- How can they navigate regulation effectively?
- Regulation: As AI-driven multiomics nears clinical use, regulatory frameworks will need to evolve accordingly. AI-driven multiomic tools will likely fall under either laboratory-based or manufacture-based tests, both requiring oversight to ensure safety and performance. There will also be a need for post-market surveillance to monitor effectiveness and safeguard patient trust as these tools enter real-world use.
- Multi-stakeholder coordinated effort: Multiomics is inherently multidisciplinary, spanning biology, data science, clinical medicine, ethics, and regulation. Leadership to implement multiomics in clinics may therefore be best provided by a multi-sector consortium or oversight board that includes representatives from the government, the NHS, academia, and industry. This body could coordinate validation studies, promote data-sharing standards, and align incentives for implementation.
The need for a multiomics strategy
Our discussions revealed a broad consensus around a need for a national multiomic strategy. This strategy could comprise a coordinated framework specifying multiomic test authorisation pathways, reimbursement mechanisms, communication standards, and professional responsibilities. It could build on existing guidance for AI and genomics (e.g., from NICE, the Genomics and Artificial Intelligence Network (GAIN), and the Royal Colleges) and extend them to address the specific challenges of AI-driven multiomic tools.
The path ahead
While AI-driven multiomics remains largely confined to research and is not yet translatable into clinical practice, it is now crucial to prepare for its future even as its ultimate clinical value remains to be fully determined. This is because AI and multiomics together hold the potential to redefine healthcare. It can help move from treatment based on averages to care that reflects the unique biology of every patient. Without a strategic approach, healthcare implementation risks being fragmented, inconsistent and inequitable.
Realising this vision will require coordinated action: building clinical and digital infrastructure, establishing clear guidance and regulation, fostering public trust, and ensuring collaboration across government, industry, and research.
The question is no longer if this transformation will happen, but how prepared we are to shape it.
If this topic interests you, please do get in touch, we are always keen to discuss opportunities for collaborations. Email: [email protected]
Acknowledgements
We would like to thank the Multiomics, AI and Health 2025 roundtable participants and the following individuals, whose valuable time and insights helped shape this briefing: Prof Mike Inouye, Dr Alexander T. Deng, Dr Sheuli Porkess, Dr. Charles Alessi, Robin Carpenter, Diane Gaston, Prof Louise Jones, Dr Sian Morgan, Professor Sarah Coupland.
A promising approach towards personalised care is AI-driven multiomics – the combined analysis of genomic, transcriptomic, proteomic, metabolomic, and other biological layers of information facilitated by AI. As each omic dataset provides a unique lens into patient biology, such combined analyses yield powerful insights that no single dataset can capture. They could be immensely valuable in clinical settings as they hold the potential to inform clinical decisions, for example, in disease risk prediction, disease subtyping, and predicting response to treatment.
At PHG we believe now is the time to prepare for wide scale translation of these important technologies into health systems. This means ensuring that
- AI-driven multiomic solutions in development are addressing real-world clinical need
- There is system-wide readiness to adopt and equitably implement these solutions
Kickstarting the discussion
We convened an expert group of stakeholders to discuss what AI-driven multiomics could look like in healthcare settings, what is needed to support this vision and what challenges are anticipated. This article captures their perspectives and is intended to get the discussion started on integrating AI-driven multiomics in healthcare for personalised medicine. But first – what is ‘multiomics’?
A definition of multiomics
Initially, omics referred to large-scale molecular characterisation of biological systems using DNA (genomics), RNA (transcriptomics), and proteins (proteomics). This landscape expanded to include metabolomics, epigenomics, and lipidomics, among others. More recently, physiological and imaging-derived data have inspired new fields such as radiomics and pathomics, which extract quantitative biomarkers from radiological and histopathological images, respectively. One might also come across new emerging subfields like cardiogenomics and neurogenomics which illustrate how omic technologies are being tailored to specific organ systems and clinical domains.
We define multiomics as the use of more than one type of omic data to make biological inferences, which can be achieved either by combined analysis or via one-at-a-time addition of each omic layer.
How AI can help with the data deluge
Each omic layer generates enormous volumes of data making the analysis far beyond the capacity of traditional or manual analytical methods. Often omic data are in incompatible formats to each other, which further complicates analysis.
. This is where artificial intelligence (AI) is poised to play a transformative role as it can help integrate, analyse and make sense of complex multiomic data.
AI can support multiomic analysis mechanistically and generatively:
- Mechanistically: When two patient groups respond differently to the same treatment, AI-driven multiomics can help identify the biological mechanisms that distinguish them, going beyond what genomics alone can reveal.
- Generatively: AI models can detect or even hypothesise new patterns across large datasets, as seen in digital pathology tools such as PathChat and HistoGPT, which assist in analysing histology images and generating structured pathology reports.
To maximise the effectiveness of AI tools in health services, it is essential to ensure not only the generation of the necessary amount of data but also the ‘right’ kind of data.
Healthcare data is frequently highly fragmented, stored in silos, captured in inconsistent formats, and are often unstructured particularly within clinical notes. Ensuring systems are in place to generate high-quality, well-annotated datasets and consolidating these for training and validating AI models is critical. Moreover, ongoing efforts to digitise NHS care as part of the NHS 10-year plan is an important step towards enabling AI integration in the health service. It is important to acknowledge that for some services, IT infrastructure is still very basic, making AI integration feel like a distant future.
Case study: GE Edison Health Services Platform
A promising model is the GE Edison Health Services platform, developed by GE Healthcare, the University of Cambridge and Addenbrooke’s Hospital. This AI-enhanced platform integrates clinical, imaging, and genomic data into a single interface for multidisciplinary teams to simultaneously view a patient’s profile and co-develop personalised treatment strategies.
This collaboration produced the Integrated Radiogenomics for Ovarian neoadjuvant therapy (IRON) tool, since renamed CareIntellect for Oncology (CI4O). This tool predicts ovarian cancer treatment outcomes with ~80% accuracy by combining radiomics, genomics, blood tests, and EHRs. It could demonstrate the power of a multiomic approach, however, clinical deployment is still pending beyond this proof-of-principle study. CI4O is currently connected to Cambridge University Hospitals’ EHR via a cloud server and its functionality is being evaluated.
Powerful tools, such as GE Edison Health Services platform, described above, must be used responsibly. Clinical AI models need to be transparent and interpretable models (white-box) so clinicians can understand how a conclusion was reached. AI in healthcare should be a collaborator, not a replacement for human decision-making.
Where white-box models are not possible, there must be sufficient confidence in the models such that regulators and clinicians are assured that the AI is functioning safely and effectively for its intended purpose. It is likely to be important that explanations can be provided about how an AI tool reached its conclusion and that a human can scrutinise aspects of its processing.
Multiomics in clinics
Multiomic approaches hold great promise for improving diagnostic yield. This is especially true for cases where simple, single-omic-based biomarkers do not exist. Decisions about which multiomic analyses should be implemented in clinical settings must be guided by clinical need, not by the availability of technology/research. Innovation in this space should be problem-driven, not research/technology-led.
The role of genomics
Genomics, as the most mature omic discipline with proven clinical impact, is often seen as central to multiomics, providing the biological foundation and translational model for other omic modalities. But does giving it a dominant role risk constraining the development of other omic disciplines? Treating genomics as the blueprint can overlook the distinct data characteristics and utility of transcriptomics, proteomics, and metabolomics. A truly multiomic approach, therefore, requires assigning equal or context-specific value to each dataset.
In practice, genomics remains the most widely adopted omic analysis in healthcare. Within the NHS, transcriptomic testing is already used clinically, but often as a second-line tool deployed when genomic analysis fails to yield conclusive results or to provide additional layers of insight. This hierarchy reflects both technical readiness and economic considerations as large-scale proteomic or metabolomic testing remains more costly and less standardised than genomic sequencing. As such, it is plausible that future deployment of non-genomic omics may, at least initially, be embedded within existing Genomic Medicine Service (GMS) pathways, supported by specialist omic centres.
This landscape is evolving rapidly. For example, Nightingale Health has developed a metabolomic blood-based test that quantifies relative disease risk at a cost lower than current whole-genome sequencing efforts. Such innovations suggest a potential rebalancing of the omic hierarchy where other omics could soon match or even surpass genomics in scalability and clinical value.
So, what are the pressing clinical problems that multiomics might help address?
One example lies in disease monitoring. Unlike genomics, which provides a largely static view of inherited mutations, metabolomic or proteomic assays can be repeated over time to capture dynamic changes in disease progression or treatment response. This makes them powerful tools for longitudinal patient management and adaptive therapeutic interventions.
Another area of need is the interpretation of variants of uncertain significance (VUSs) which are genetic findings whose clinical relevance is unclear. Integrating additional molecular layers, such as transcriptomic or proteomic data, could help clarify whether a given variant is disease-causing or benign. For clinicians, this represents a tangible and immediate benefit. It enables them to make more confident, evidence-based decisions about diagnosis, prognosis, and treatment planning.
What needs to happen?
In addition to AI-driven multiomic solutions that address specific clinical needs, stakeholder specific concerns need to be addressed.
- Clinicians: Time pressures and liability concerns could be key barriers to multiomic adoption. When new technologies emerge faster than clinical consensus can form, clear guidance becomes essential. Without it, clinicians risk being caught between two difficult positions: acting prematurely on inconclusive information or failing to act on data that later proves significant. National or institutional clinical guidance can help define when, where, and how multiomic tests should be used, providing both a framework for adoption and legal protection for practitioners.
- Patients: Public awareness of genomics has grown substantially, making communication of clinical findings to patients that bit simpler. But test results based on probabilities remain complex or ambiguous. These are familiar challenges for genetic counsellors, who routinely help patients interpret such probabilistic findings. Multiomics introduces new layers of complexity. Perhaps multiomic counsellors – trained to communicate uncertainty, risk, and the limitations of emerging evidence – could help patients navigate their multiomic test results.
- Industry: Companies face challenges accessing NHS and academic datasets, which are often fragmented or of variable quality. Moreover, the absence of a clear deployment and procurement pathway makes it difficult for developers to design and sell clinically relevant tools. They need guidance on fundamental questions such as:
- What constitutes a “good” product in the clinical context?
- What level and type of evidence are required to show clinical validity?
- Which clinical use cases are viable and supported?
- Which patient groups are most appropriate?
- How can they navigate regulation effectively?
- Regulation: As AI-driven multiomics nears clinical use, regulatory frameworks will need to evolve accordingly. AI-driven multiomic tools will likely fall under either laboratory-based or manufacture-based tests, both requiring oversight to ensure safety and performance. There will also be a need for post-market surveillance to monitor effectiveness and safeguard patient trust as these tools enter real-world use.
- Multi-stakeholder coordinated effort: Multiomics is inherently multidisciplinary, spanning biology, data science, clinical medicine, ethics, and regulation. Leadership to implement multiomics in clinics may therefore be best provided by a multi-sector consortium or oversight board that includes representatives from the government, the NHS, academia, and industry. This body could coordinate validation studies, promote data-sharing standards, and align incentives for implementation.
The need for a multiomics strategy
Our discussions revealed a broad consensus around a need for a national multiomic strategy. This strategy could comprise a coordinated framework specifying multiomic test authorisation pathways, reimbursement mechanisms, communication standards, and professional responsibilities. It could build on existing guidance for AI and genomics (e.g., from NICE, the Genomics and Artificial Intelligence Network (GAIN), and the Royal Colleges) and extend them to address the specific challenges of AI-driven multiomic tools.
The path ahead
While AI-driven multiomics remains largely confined to research and is not yet translatable into clinical practice, it is now crucial to prepare for its future even as its ultimate clinical value remains to be fully determined. This is because AI and multiomics together hold the potential to redefine healthcare. It can help move from treatment based on averages to care that reflects the unique biology of every patient. Without a strategic approach, healthcare implementation risks being fragmented, inconsistent and inequitable.
Realising this vision will require coordinated action: building clinical and digital infrastructure, establishing clear guidance and regulation, fostering public trust, and ensuring collaboration across government, industry, and research.
The question is no longer if this transformation will happen, but how prepared we are to shape it.
If this topic interests you, please do get in touch, we are always keen to discuss opportunities for collaborations. Email: [email protected]
Acknowledgements
We would like to thank the Multiomics, AI and Health 2025 roundtable participants and the following individuals, whose valuable time and insights helped shape this briefing: Prof Mike Inouye, Dr Alexander T. Deng, Dr Sheuli Porkess, Dr. Charles Alessi, Robin Carpenter, Diane Gaston, Prof Louise Jones, Dr Sian Morgan, Professor Sarah Coupland.