Omics analysis: from traditional statistics to AI-driven insights

Given the central role of the liver in metabolic and immune functions, the integration of different omics modalities could provide a comprehensive understanding of the underlying regulatory mechanisms. Whereas statistical approaches were traditionally employed, data-driven AI models have now become more ubiquitous for omics data analysis. These models are adept at handling the intricate, high-dimensional and noisy omics datasets. DL models, a class of neural network-based AI models that use many layers, are particularly effective at learning complex relationships within high-dimensional data and are increasingly used for omics data. Further, integration with clinical and structural data with omics can enable the identification of molecular subtypes and patient strata, leading to more targeted and effective treatments.5 6 Current AI models are capable of learning complex functions that map omics and clinical features to outcomes. Figure 1 illustrates the emerging approach where an AI model prediction is treated as a biomarker.

• Download figure
• Open in new tab
• Download powerpoint

Figure 1

Artificial intelligence (AI) model prediction as a biomarker. The left panel depicts the pressing need for non-invasive tests in diagnosing and staging liver disease (LD), emphasising the pivotal role of AI. With sufficient data, current AI models have the capacity to evaluate many features and learn complex functions that can accurately predict outcomes and recognise the degree of severity leading to end stage disease. The panel on the right contrasts this with the examining of liver tissue which remains a gold standard for many aspects of diagnosis and treatment of patients with LD. ML, machine learning.

The most common approach in the application of AI to omics data is to first use statistical methods such as differential analysis to identify a subset of relevant features, followed by AI model development to integrate prediction based on these features. Feature selection or dimensionality reduction is often necessary given the high-dimensional nature of omics data. The choice of algorithms for feature selection and modelling is typically influenced by the data’s structure and its specific characteristics, and the analysis objective. Table 1 presents some algorithm choices that align with these data nuances and problem contexts. While table 1 is not comprehensive and does not recommend using these algorithms exclusively for the conditions highlighted, it demonstrates that a variety of potential solutions are available for a given problem. For instance, when the number of samples is small, linear models, which are less prone to overfitting, serve as a good starting point. However, with appropriate data augmentation to generate synthetic samples, other alternatives, such as random forests or deep neural networks, can also be considered. Data augmentation, though, is not straightforward and requires careful consideration to ensure that the synthetic samples are realistic within the context of the domain and the problem. While this practice is common in image and natural language processing (NLP), the complexity of omics data presents a challenge that necessitates further research. Therefore, it is essential to consider the trade-offs between computational efficiency, interpretability and predictive performance when selecting algorithms. Evaluating multiple methods and validating them through techniques like cross-validation can help achieve reliable and robust models.

While end-to-end AI approaches that combine the feature selection and modelling steps have been proposed, statistical feature selection often infuses domain knowledge such as pathways and gene networks that are challenging to integrate without a strong prior. Statistical methods such as differential analysis consider each feature independently. Whereas AI models attempt to learn a non-linear relationship between multiple features, their combinations and the outcome. Figure 2 describes a common data analysis pipeline, where the modelling step integrates a small subset of selected features across several modalities in the molecular hierarchy. Interpretability and generalisability are two essential aspects of the AI model in clinical applications (figure 2) and are further discussed in the ‘Common pitfalls of applying AI to omics data’ section. In scenarios where model interpretability is important, preprocessing steps like feature selection or dimensionality reduction are still valuable.

• Download figure
• Open in new tab
• Download powerpoint

Figure 2

Multiomics integration pipeline. Illustrates the most prevalent multiomics data analysis and integration pipeline. Most studies interrogate a subset of omics modalities from the molecular hierarchy. The data analysis is typically split into two steps: a feature selection step and a modelling step. The objective of the feature selection step is to reduce the dimensionality of the data by identifying key features relevant to the problem. Prior biological knowledge is also commonly integrated to improve feature selection. In the AI modelling step, the model learns to integrate the multimodal features selected through the statistical feature selection step and map them to outcomes. The choice of modelling algorithm hinges on the type of outcome: prognostic, diagnostic and therapeutic applications usually employ supervised learning, while molecular subtyping and biomarker discovery often use unsupervised learning. However, these distinctions are not always clearly defined, and many studies opt for a combination of both approaches. AI, artificial intelligence.

Due to the challenges and costs associated with collecting and processing patient samples, omics datasets are often smaller compared with those used in other successful AI application areas such as computer vision and NLP. Developing AI methods for such high-dimensional data with small sample sizes is challenging. Self-supervised and transfer learning approaches can address some of these challenges. Figure 3 illustrates a few of the most widely used learning paradigms that are often referred to in this paper. Some other popular paradigms, such as reinforcement learning (RL) and federated learning (FL), are not shown here. In practice, these approaches are often combined to achieve specific results. For instance, one can use unsupervised pretraining followed by supervised fine-tuning for transfer learning when a large unlabelled dataset and a small labelled dataset are available.

• Download figure
• Open in new tab
• Download powerpoint

Figure 3

An overview of common machine learning paradigms used to train deep learning-based models. Illustrates the core ideas behind several machine learning paradigms frequently mentioned in this paper. (A) Supervised learning: The most successful learning approach, particularly effective when sufficient labelled training and validation data are available. (B) Unsupervised learning: In this paradigm, labels are not available for the samples. The illustration uses an autoencoder, a specific neural network architecture, to learn a compressed (latent) representation of the data. The aim of this dimensionality reduction is to capture features relevant to the ultimate task (eg, labelling or classification). Clustering algorithms such as k-means, DBSCAN, Gaussian mixture models, etc are used to group samples in the latent space. (C) Self-supervised learning: Combines ideas from supervised and unsupervised learning, with labels generated from the data itself. In this example, contrastive learning is used to implement self-supervised learning. The labels generated are ‘similar’ and ‘dissimilar’, referring to pairs of samples generated from the same or different samples, respectively. Self-supervised learning formulates the objective to more explicitly learn such latent spaces, unlike unsupervised learning. (D) Transfer learning is particularly useful when limited data is available for a task. Here, a model is first pretrained to differentiate between red squares and blue bars and then fine-tuned to classify red stars and blue circles, for which there are limited examples. The pretraining helps the model learn latent dimensions useful for the eventual task, and fine-tuning further optimises these dimensions for the specific task. DBSCAN, density-based spatial clustering of applications with noise.

Data-driven feature selection and biomarker discovery

Knowledge-driven statistical feature selection is widely popular in analysing omics data. Linear techniques such as principal component analysis7 and linear discriminant analysis (LDA)8 are commonly used to address multicollinearity. Multidimensional scaling (MDS)9 aims to maintain local relationships (neighbours) in a lower-dimensional space. Non-linear dimensionality techniques, such as t-distributed Stochastic Neighbour Embedding and Uniform Manifold Approximation and Projection, seek a balance between preserving the global shape of the data and local relationships within the data in the projected lower-dimensional space.10 However, the resulting projected feature spaces, whether linear or non-linear, are often less interpretable, and thus primarily used for visualisation and validation purposes.

Linear variable selection methods11 12 alongside non-linear strategies such as tree-based methods,13 support vector machine (SVM) and feature importance14 offer the advantage of preserving features and are therefore widely used as a precursor step to the modelling step.

Unsupervised DL-based autoencoders (AEs) are designed to learn a compressed, lower-dimensional representation of input data. The compressed data are then used to reconstruct the original input, ensuring that vital features necessary for reconstruction are captured by the lower-dimensional feature space. AE variants, such as variational AE (VAE), have shown promise in extracting more interpretable low-dimensional latent spaces from gene expression profiles for cancer classification.15 AEs are often used to pretrain DL models ultimately used for supervised, semisupervised or transfer learning.16 17

Molecular subtyping

Supervised AI models can be trained on molecular, clinical and phenotypic data using known subtypes to classify patients into diagnostic categories. Such subtyping can aid in biomarker discovery and potentially guide personalised preventive measures.18 Gao et al19 developed a DL-based supervised cancer subtyping framework trained on established subtypes based on the consensus clustering molecular subtyping approaches.

Unsupervised AI techniques, such as clustering algorithms, are vital for identifying novel subgroups within heterogeneous datasets.16 20–23 The Cancer Genome Atlas (TCGA) network used unsupervised clustering methods on multiplatform data to identify molecular subtypes in breast cancer samples.24 Similar clustering approaches have been employed to integrate multiple omics modalities from primary tumours and paired para-tumoural tissue for the identification of three hepatocellular carcinoma (HCC) subtypes with distinct clinical prognoses.18

Metabolic dysfunction-associated steatotic liver disease

Metabolic dysfunction-associated steatotic liver disease (MASLD), in which fat accumulates in the liver, is one of the most prevalent liver diseases worldwide. In recent years, metabolite and serum lipid profiling have emerged as important tools to diagnose metabolic dysfunction and assess disease severity in MASLD, including conditions such as active inflammation seen in metabolic dysfunction-associated steatohepatitis (MASH) and fibrosis.

Machine learning methodologies have been used to investigate metabolomics and pathophysiology of MASLD. McGlinchey et al25 used Gaussian mixture model-based clustering to group lipids and investigated their association with metabolites and clinical variables using a correlation network. Using their network, they identified unique metabolites linked to steatosis, MASH and fibrosis. They developed RF binary classifiers to model fibrosis progression, distinguishing between early and advanced stages, with a median AUROC of 0.77 for F0–F1 vs F2–F4 task and a median AUROC of 0.78 for F0–F2 vs F3–F4 task. Their findings highlighted a critical metabolic transition from stages F2 to F3 in MASLD pathogenesis, emphasising the significance of oxidative stress.

Quesada-Vázquez et al26 employed the Boruta algorithm, a supervised feature selection method based on RF, to identify metabolomic signatures of steatosis and used the limma linear regression model27 for transcriptomic analysis. Through SHAPley additive explanation (SHAP) analysis, they found plasma histidine strongly inversely associated with steatosis (SHAP mean value=0.777) and linked it to a hepatic transcriptomic signature involving insulin signalling, inflammation and trace amine-associated receptor 1.

Perakakis et al28 developed non-invasive diagnostic models that integrate lipids, glycans and hormones to diagnose MASH using SVM with recursive feature elimination. They identified a combination of lipids that could accurately diagnose the presence of liver fibrosis with 98% accuracy. Chiappini et al29 also developed an RF model to differentiate MASH from various MASLD stages using lipid profiling. They highlighted dysregulation in the fatty acid synthesis metabolic pathway as a diagnostic of MASH.

In a paediatric study, Khusial et al30 combined the plasma metabolome with clinical variables with an RF model, identifying a panel of amino acids consisting of serine, leucine/isoleucine and tryptophan to screen for MASLD in youth (AUROC=0.94). Apart from plasma metabolome, recent studies have systematically compared lipidomic profiles from urinary extracellular vesicles (EV), identifying lipid signatures diagnostic of MASH.31 Moolla et al32 also differentiated early (F0–F2) from advanced (F3–F4) fibrosis using a generalised matrix learning vector quantisation analysis on urinary metabolomic data (AUROC=0.92).

The association between the human gut microbiota and its contribution to the pathogenesis of MASLD has also been explored through machine learning methodologies. In a study by Wang et al,33 predictive taxonomic and functional signatures were identified from publicly available MASLD shotgun metagenomic datasets to elucidate links to MASH progression. To further investigate the diagnostic accuracy of their microbiome signature, stochastic gradient-boosting machine (GBM) learning models were trained to identify the severity of MASH, achieving performance ranging from AUROC of 0.7 to 0.94. Oh et al34 identified a gut microbiome-derived signature, combined with patient age, accurately detecting cirrhosis in a diverse MASH-cirrhosis population compared with non-MASH controls. Their RF model, trained on shotgun metagenomic and untargeted metabolomic profiles, achieved an AUROC of 0.91.

Using RF modelling, Sharpton et al35 identified a stool metagenomic signature comprising 13 discriminatory species to detect decompensated MASH cirrhosis. In their study, Saboo et al36 analysed stool and saliva microbiomes of cirrhosis patients, highlighting the superior predictive value of stool microbiota for cirrhosis progression over saliva microbiota. Their RF-based classification model based on stool microbiota achieved the highest AUROC of 0.78 in distinguishing patients with cirrhosis from controls.

The gut microbiome has also been investigated in a prognostic model to understand MASH development. Leung et al37 examined serum and stool samples and characterised both metagenomic and metabolomic changes during MASH progression in a community-based cohort. They developed interpretable models using RF, with AUROCs of 0.72–0.80, integrating baseline microbial signatures to identify patients at risk of developing MASLD 4 years after baseline.

Other omics modalities have emerged in recent literature as potential diagnostic and prognostic markers in MASLD. Eslam et al38 combined routinely available clinical and laboratory data with IFNL genotype for fibrosis prediction using decision trees. Govaere et al39 used logistic regression (LR) to model the progression of MASH, discriminating mild from advanced disease (MASH F0/F1 vs MASH F≥2) with an AUROC of 0.86. Baboota et al40 identified GREM1 as a strong predictor of fibrosis using conditional RF and gradient boosting, highlighting senescence markers as crucial drivers of MASH in humans.

In proteomic-based studies, Feng et al41 used LASSO regression with a 10-fold cross-validation to pinpoint urinary proteins associated with MASH and fibrosis stages. Their model distinguished advanced fibrosis with a better AUROC of 0.92 compared with early fibrosis stages with AUROC of 0.86. Luo et al42 achieved an AUROC of 0.78 using an elastic-net algorithm based on a 4-protein or 12-protein panel to distinguish advanced fibrosis among MASH patients.

An increasing number of studies have combined multiomics technologies to find diagnostic markers of MASLD.43 44 Atabaki-Pasdar et al43 applied machine learning methods, on patients with or without type 2 diabetes from the Innovative Medicines Initiative diabetes research on patient stratification Consortium cohort,45 to devise an early diagnostic model for MASLD. A Pearson correlation matrix, with a pairwise correlation threshold of r>0.8, was used for the feature selection of clinical variables while LASSO-based feature selection was performed for dimension reduction at each of the omics levels. An RF model developed using the preselected features, which consisted of both clinical and omics variables, achieved an AUROC of 0.84. Wood et al44 further assessed the performance of steatosis classifier models in the setting of obesity using LR on each omics domain alone and found that proteomics achieved the highest AUROC of 0.913 followed by phenomic data of 0.892 and PNPLA3 genotyping data of only 0.596. Combining all markers selected from each individual data domain achieved an AUROC of 0.935 for the diagnosis of hepatic steatosis.

Overall, AI algorithms integrating omics and clinical data show significant potential to inform personalised diagnostic, prognostic and therapeutic strategies in MASH.

Alcohol-associated liver disease

Alcohol-associated liver disease (ALD) encompasses a range of liver injuries due to alcohol consumption, spanning from hepatic steatosis to advanced stages like alcoholic hepatitis (AH), alcohol-associated cirrhosis (AC) leading to acute-on-chronic liver failure (ACLF). Despite advances, the molecular pathophysiology of ALD remains incompletely understood, highlighting the necessity for non-invasive diagnostic methods to screen individuals at risk, notably those with a history of alcohol misuse. Omics data and machine learning methodologies have been explored as potential diagnostic markers in ALD.

In a cohort of 7115 individuals over ~15 years, Liu et al46 explored the use of microbiome-augmented gradient boosting algorithms to predict liver disease incidence, particularly ALD. Gut microbial composition was assessed across various taxonomic levels. Prediction models for incident ALD were developed, using microbial features down to the species level, achieving an AUROC of 0.895.

Niu et al47 used plasma proteomic profiling through mass spectrometry to develop a diagnostic test for early fibrosis detection in patients with ALD. Their LR models, trained on a subset of 22 proteins, achieved better performance than 15 standard clinical tests in predicting significant fibrosis, mild inflammatory activity and any liver steatosis, achieving a mean AUROC of 0.90.

In their study, Listopad et al48 applied multiple machine learning models, such as LR, SVM and k-nearest neighbours (k-NN) to transcriptomic and proteomic data from liver tissue and peripheral blood mononuclear cells (PBMCs) to differentiate AH, AC and healthy controls. Integrating transcriptomic and proteomic data yielded superior performance in both liver tissue and PBMCs compared with individual omics-based classification methods. In PBMCs, the combined model achieved an AUROC of 0.96, outperforming the AUROCs of 0.89 for each individual omics-based classification.

Trépo et al49 developed gene-signature Model for End-Stage Liver Disease (gs-MELD), a scoring system combining baseline liver gene expression patterns of 123 genes with the MELD score to differentiate severe AH patients based on survival risk. The gs-MELD score was crafted using regression coefficients obtained from a multivariable Cox regression model.

Primary sclerosing cholangitis

Primary sclerosing cholangitis (PSC) is a rare cholestatic liver disease with a poorly understood pathogenesis and limited therapeutic options. It is characterised by bile stasis and progresses to fibrosis, cirrhosis, at times culminating in hepatic insufficiency.

Iwasawa et al50 investigated the salivary microbiota to distinguish salivary microbial communities of paediatric-onset PSC patients from healthy controls. For the classification problem, a combination of genera and species was used in RF models to distinguish the PSC group with an AUROC of 0.74.

Altered bile acids play a pivotal role in nutrient absorption and act as crucial signalling molecules, regulating hepatic metabolism and modulating immune responses. While altered BA homeostasis is an intrinsic facet of cholestatic liver diseases, the clinical utility of plasma BA assessment in PSC remains understudied.

Mousa et al51 used a combination of univariate Cox proportional hazard and multivariable gradient boosting machine models to evaluate whether plasma BA was an effective predictor of hepatic decompensation in PSC patients. Their model achieved a high concordance statistic (C-statistic) of 0.86 on a validation cohort.

Hepatitis B virus and hepatitis C virus

Although viral hepatitis incidence has decreased due to curative treatments and vaccines for hepatitis C virus (HCV) and hepatitis B virus (HBV), respectively, they remain the leading causes of chronic liver diseases. Viral hepatitis progresses rapidly, often culminating in cirrhosis, HCC and end-stage liver disease.

Treatment for chronic HCV infection has improved rapidly since the introduction of all-oral, interferon-free direct-acting antiviral (DAA) regimens. DAAs have significantly high cure rates (sustained virological response, SVR) compared with previous interferon-based therapies. Cure rates with DAAs exceed 90%–95% across different HCV genotypes and patient populations, including those with cirrhosis and other comorbidities. However, patient-specific comorbidities burden, prior treatment exposures and challenges with adherence, which can affect the effectiveness of DAAs. Park et al evaluated multiple machine learning approaches and found GBM the most effective in generating prediction risk scores for identifying patients with chronic hepatitis C infection at high risk of DAA treatment failure and to determine predictors associated with DAA treatment failure.52 Feldman et al present a machine-learning-based approach to predict patient characteristics associated with the need for extended duration for DAA therapy given a particular DAA regimen. Their optimal predictive model using XGBoost had an AUC of 0.745.53

In recent times, AI-assisted drug discovery has streamlined and enhanced the drug discovery process. Particularly, AI has been used to identify new uses for existing drugs. One such study by Kamboj et al used SVM, RF, k-NN and artificial neural network and quantitative structure–activity relationship approaches to predict repurposed drugs targeting HCV non-structural proteins.54

Wu et al55 implemented an attention-based DL model to predict HBV genomic integration sites, unveiling novel insights into HBV-induced cancer. Their binary classifier, incorporating convolutional layers and attention modules, effectively predicted HBV integration sites by identifying key genomic regions crucial for classification.

In a proteomic study by Estevez et al,56 an RF model analysed patterns in 51 common cytokines. Their findings revealed distinct cytokine profiles in patients with HBV or HCV infection, particularly in the context of the presence of HCC.

HCC and intrahepatic cholangiocarcinoma

While the burden of HBV and HCV-associated HCC has been declining with effective antiviral therapy, HCC incidence related to metabolic syndrome is expected to rise due to the dramatic increase in the prevalence of MASH in the general population.57

Omics data and machine learning methodologies are a growing field of study in primary liver malignancies and their potential application in translational research is broad, from screening, to diagnostics, to prognostic clinical tests.

Hershberger et al58 identified machine-learning-derived salivary metabolite combinations as promising biomarkers for HCC detection, highlighting saliva analysis as a potential diagnostic tool for liver diseases. Furthermore, volatile organic compounds (VOCs) in patient breath offer valuable insights into disease processes, capable of distinguishing disease status. Leveraging machine learning with breath VOCs holds the potential for developing improved, non-invasive screening tools for chronic liver disease and primary and secondary liver tumours.59

Lewinska et al60 employed multivariate analysis to evaluate the diagnostic potential of altered serum lipidome in MASLD-associated HCC, revealing depleted unsaturated fatty acids and elevated triglycerides as indicative of deregulated metabolic networks in MASLD-HCC. Wu et al61 enhanced serum lipid profiling accuracy with graphitised carbon matrix, offering superior diagnostic utility for liver cancer compared with conventional matrices.

The primary factors behind HCC are genetic and epigenetic changes. Lu et al62 integrated clinical features and genetic biomarkers to establish a decision tree-based model for HCC detection postviral eradication. Their prospective study included 55 HCV patients with advanced fibrosis achieving sustained virological response. The model achieved 95.7% accuracy in HCC detection, identifying TAS1R3, FOSL1 and ABCA3 as key predictors. The transcriptome serves as a pivotal link between cellular phenotype and genetic tumour biology, containing RNA-encoded information transcribed from DNA. Unlike the relatively stable genome, the transcriptome is dynamic, responding actively to physiological or pathological conditions. Choi et al63 introduced a network-based DL model, adapted from the Word2Vec framework, to analyse gene expression data sourced from TCGA. This approach identified distinct gene modules significantly linked to HCC prognosis. Along with transcriptomic data, Chaudhary et al64 used microRNA and methylation data in an AE-based DL framework to identify two HCC subpopulations with differential survival. Their results consistently linked the more aggressive subtype with higher TP53 inactivation mutation frequencies while the moderate subtype exhibited activated metabolism-related pathways, including drug metabolism, amino acid and fatty acid metabolism.

Recent single-cell transcriptomic studies have enabled in-depth analysis of the immune landscape. LASSO regression was used to identify T-cell exhaustion signatures, characterising the immune landscape and predicting HCC prognosis.65 Gong et al66 uncovered the critical role of neutrophils in the HCC tumour microenvironment. They developed a neutrophil-derived signature via the partition around medoids algorithm), followed by Cox analysis, integrating them by fitting 101 models using the leave-one-out cross-validation framework.

As ageing-related genes are closely related to the occurrence and development of cancer, building prognostic models with them is gaining popularity. Zhang et al67 using gene expression data developed an HCC senescence score model for prognostic prediction. Using a combination of feature selection methods and a LASSO penalised Cox regression model, the genes, CDCA8, CENPA, SPC25 and TTK, associated with HCC progression. In another study, LASSO regression was used to develop a prognostic model consisting of three ageing-related genes, TFDP1, NDRG1 and FXR1.68

Liu et al69 identified a set of nine efferocytosis-related genes and their regulatory influence on HCC immunotherapy to construct a prognostic model for HCC. Ma et al70 developed a consensus clustering approach to identify the impact of tumour cell state heterogeneity and functional clonality on HCC and intrahepatic cholangiocarcinoma (iCCA) patient prognosis.

Data-driven omics-based biomarkers have the potential to guide personalised therapeutic strategies in oncology. Genomic predictors could be used to tailor chemotherapy, immunotherapy and combined treatments for more effective therapeutic response. Retrospective studies have shown that gene expression profiles and other molecular signatures derived from omics data can predict patient response to therapies like anti-PD1 and anti-CTLA-4, which are used in immunotherapy. Hectors et al found that the expression of the immunotherapy target PD-L1 at the protein level, as well as PD-1 and CTLA4 at the mRNA level, correlated with radiomic features of HCC.71 Chen et al evaluated the stemness index based on transcriptome data to find subtypes within patients with HCC.72 Their LR-based classifier achieved an AUC of 0.918 to identify patients who are more probable to have a sensitive response to immunotherapy based on stemness features. These AI tools would need to be further validated to permit deployment into clinical practice, but nonetheless show the promise of integrating clinical and ‘omics data using AI to guide a personalised therapeutic approach. Although not using omics, DL architectures, such as transformers, have been used to study HCC. Sato et al73 used a transformer model to predict overall survival in patients with HCC treated by radiofrequency ablation (RFA). Wang et al74 used a transformer to study associations between MRI and microvascular invasion in patients with solitary HCC. Further, combining different treatment modalities, such as systemic therapies with local treatments like RFA or transarterial chemoembolisation (TACE), is an emerging strategy aimed at improving therapeutic outcomes and overcoming resistance mechanisms in HCC and is discussed in the ‘Omics data integration with biomedical images’ section.

Acute liver failure

Acute liver failure (ALF) is a rare clinical manifestation of rapid and severe liver dysfunction in patients without pre-existing liver disease. Challenges associated with building AI tools for ALF include its sudden onset which limits the availability of comprehensive molecular profiles of affected patients. AI tools for ALF must address the need for short-term risk prediction, patient stratification and evaluation of therapeutic responses to improve patient outcomes.

Sharma et al75 identified the top five metabolites through multiomics analysis and assessed their potential in predicting early mortality or the need for emergency liver transplantation in patients with ALF. Five machine learning algorithms RF, k-NN, SVM, classification and regression trees and LDA were evaluated for classification based on metabolite expression levels. While models trained on individual metabolites exhibited significant variation, the combined model utilising all five metabolites achieved the highest AUC of 0.933 on the validation set.

Drug-induced hepatotoxicity, also a leading cause of ALF, is categorised as acute or chronic depending on the duration of injury and histological damage location. Jin et al76 developed an interpretable XGBoost-SHAP model integrating rat liver gene expression profiles, drug labels and clinical reports to evaluate and predict drug-induced liver injury (DILI). Moore et al77 investigated SNP impact on DILI susceptibility using linear models such as multivariate adaptive regression splines (MARS), multifactor dimensionality reduction (MDR) and LR. They found that SNP–SNP interactions significantly predicted DILI chronicity.

Acute-on-chronic liver failure

ACLF is characterised by rapid clinical deterioration in patients with chronic liver disease. While omics-based studies have increased, the rapidly evolving clinical course of ACLF poses challenges in identifying its pathophysiological process and developing prognostic models.78

A study by Zhang et al79 used an RF model in conjunction with differential analysis (t-test followed by false discovery rate-based sorting) to prioritise metabolite biomarkers that are found relevant in both approaches. For diagnosing pre-ACLF patients, they combined two metabolites, pipecolate and gamma-carboxyethyl hydroxychroman (γ-CEHC), with clinical parameters using an LR model. Their combined diagnostic model achieved an AUROC of 0.88 on a validation dataset.

Further, they developed a prognostic model, with metabolites comprising ureidopropionate, N-acetyl-aspartyl-glutamate and pipecolate, to predict 90-day mortality in patients with ACLF, achieved an AUROC of 0.83 on a validation dataset.

Acute cellular rejection

There is a growing body of studies focusing on molecular profiling of liver allografts for early diagnosis of outcomes, such as acute cellular rejection (ACR). ACR commonly occurs within the initial 90 days postliver transplantation, marked by cytotoxic T cell infiltration into the transplanted liver tissue. In a paediatric study, Ningappa et al80 applied an integrative machine learning method merging transcriptomics data with the human protein interactome. Through a LASSO-based feature selection approach, they pinpointed network module signatures linked to rejection in paediatric liver transplantation with a median AUROC of 0.7. Their comparison of network-based signatures to gene-based signatures obtained through the differential expression of genes (DEGs) demonstrates how these network modules may be more suitable for targeting with antirejection drugs.

Omics data integration with clinical variables

Integration of clinical variables with omics data has been shown to significantly improve accuracy in liver disease characterisation and diagnosis across various studies. By combining information from both clinical assessments and omics profiling, researchers have achieved enhanced predictive performance, allowing for more accurate diagnosis and prognosis evaluation.

The RF model, analysing gut microbiome changes during MASLD progression alongside integrated microbial and clinical features (body mass index, age, High-Density Lipoprotein, fasting insulin), demonstrated an improved AUROC of 0.8 compared with both clinical prognostic models (AUROCs of 0.58–0.60) and models relying solely on microbial signatures (AUROC of 0.72).37 Similar improvements were demonstrated with lipids and polar metabolite integration in MASLD25 and microbiome data integration in ALD,46 Oh et al combined gut microbiome-derived metagenomic signature with patient age and serum albumin levels to detect cirrhosis. They found that addition of serum aspartate aminotransferase levels enabled discrimination of cirrhosis from earlier fibrosis stages.34

In their multiomics study, Atabaki-Pasdar et al43 found that integrating genetic, transcriptomic, proteomic and metabolomic with clinical variables yielded better performance (AUROC 0.84) compared with using clinical variables alone (AUROC 0.82) in predicting liver fat quantity.

Omics data integration with biomedical images

Histopathology plays a vital role in LD diagnosis and characterisation. DL-based histopathology image analysis studies often leverage omics data for validation. Huang et al81 used a VGG-16 network-based CNN and Raman spectroscopy to distinguish between HCC and iCCA and leveraged tissue metabolomics for validation. Calderaro et al82 developed an attention-based DL model to reclassify combined HCC-iCCA as either HCC or iCCA using whole slide images. Clinical and biological relevance of the predictions were validated using survival prognosis and spatial transcriptomic analysis. Conway et al83 used a pretrained deep CNN model to extract human interpretable features from histopathology images and correlated them with transcriptomic profiles to reveal a five-gene signature that distinguishes stage F3 from F4 fibrosis.

Radiogenomics, a promising field that correlates radiomics features with other omics modalities, has gained significant attention. Several studies have successfully used radiogenomics to decode the molecular profile of HCC.71 84–86 In a recent study,87 applied radiogenomics to predict mutations in the PI3K signalling pathway.

While surgical resection and liver transplantation remain the primary curative treatments for HCC, radiofrequency ablation (RFA) is a cornerstone for image-guided ablation in non-surgical early-stage HCC. For intermediate-stage HCC, TACE is the most widely used treatment. Boldanova et al59 integrated clinical observations with genetic and radiological insights to forecast patient responses to TACE. They used RF to identify potent features, such as tumour morphology and genetic markers, for predicting TACE response.

For those who cannot undergo the above treatments for HCC, targeted therapies, immunotherapy and other systemic treatment options have become invaluable, particularly for treating advanced stages of HCC. Tian et al combined 14 features—9 selected from a three-dimensional CNN trained on MRI images, and 5 radiomic features—to predict PD-L1 levels using an SVM classifier. Their combined model performed the best with an AUC of 0.897 while the DL-based feature selection model alone and the radiomics-based feature selection model alone had AUCs of 0.852 and 0.794, respectively.88