The emergence of biobank-level datasets offers new opportunities to discover novel biomarkers and develop predictive algorithms for human disease. Here, we present an ensemble machine-learning framework (machine learning with phenotype associations, MILTON) utilizing a range of biomarkers to predict 3,213 diseases in the UK Biobank. Leveraging the UK Biobank’s longitudinal health record data, MILTON predicts incident disease cases undiagnosed at time of recruitment, largely outperforming available polygenic risk scores. We further demonstrate the utility of MILTON in augmenting genetic association analyses in a phenome-wide association study of 484,230 genome-sequenced samples, along with 46,327 samples with matched plasma proteomics data. This resulted in improved signals for 88 known (P < 1 × 10⁻⁸) gene–disease relationships alongside 182 gene–disease relationships that did not achieve genome-wide significance in the nonaugmented baseline cohorts. We validated these discoveries in the FinnGen biobank alongside two orthogonal machine-learning methods built for gene–disease prioritization. All extracted gene–disease associations and incident disease predictive biomarkers are publicly available (http://milton.public.cgr.astrazeneca.com).

Adjuvant systemic therapies are now routinely used following resection of stage III melanoma, however accurate prognostic information is needed to better stratify patients. We use differential expression analyses of primary tumours from 204 RNA-sequenced melanomas within a large adjuvant trial, identifying a 121 metastasis-associated gene signature. This signature strongly associated with progression-free (HR = 1.63, p = 5.24 × 10⁻⁵) and overall survival (HR = 1.61, p = 1.67 × 10⁻⁴), was validated in 175 regional lymph nodes metastasis as well as two externally ascertained datasets. The machine learning classification models trained using the signature genes performed significantly better in predicting metastases than models trained with clinical covariates (p_AUROC = 7.03 × 10⁻⁴), or published prognostic signatures (p_AUROC < 0.05). The signature score negatively correlated with measures of immune cell infiltration (ρ = −0.75, p < 2.2 × 10⁻¹⁶), with a higher score representing reduced lymphocyte infiltration and a higher 5-year risk of death in stage II melanoma. Our expression signature identifies melanoma patients at higher risk of metastases and warrants further evaluation in adjuvant clinical trials.

Genomics can provide insight into the etiology of type 2 diabetes and its comorbidities, but assigning functionality to non-coding variants remains challenging. Polygenic scores, which aggregate variant effects, can uncover mechanisms when paired with molecular data. Here, we test polygenic scores for type 2 diabetes and cardiometabolic comorbidities for associations with 2,922 circulating proteins in the UK Biobank. The genome-wide type 2 diabetes polygenic score associates with 617 proteins, of which 75% also associate with another cardiometabolic score. Partitioned type 2 diabetes scores, which capture distinct disease biology, associate with 342 proteins (20% unique). In this work, we identify key pathways (e.g., complement cascade), potential therapeutic targets (e.g., FAM3D in type 2 diabetes), and biomarkers of diabetic comorbidities (e.g., EFEMP1 and IGFBP2) through causal inference, pathway enrichment, and Cox regression of clinical trial outcomes. Our results are available via an interactive portal (https://public.cgr.astrazeneca.com/t2d-pgs/v1/).

Artificial intelligence (AI) is transforming many fields, including healthcare and medicine. In biomarker discovery, AI algorithms have had a profound impact, thanks to their ability to derive insights from complex high-dimensional datasets and integrate multi-modal datatypes (such as omics, electronic health records, imaging or sensor and wearable data). However, despite the proliferation of AI-powered biomarkers, significant hurdles still remain in translating them to the clinic and driving adoption, including lack of population diversity, difficulties accessing harmonised data, costly and time-consuming clinical studies, evolving AI regulatory frameworks and absence of scalable diagnostic infrastructure. Here, we provide an overview of the AI toolkit available for biomarker discovery, and we discuss exciting examples of AI-powered biomarkers across therapeutic areas. Finally, we address the challenges ahead of us to ensure that these technologies reach patients and users globally and unlock a new era of fast innovation for precision medicine.

In melanoma, predicting which tumors will ultimately metastasize guides treatment decisions. Transcriptional signatures of primary tumors have been utilized to predict metastasis, but which among these are driver or passenger events remains unclear. We used data from the adjuvant AVAST-M trial to identify a predictive gene signature in localized tumors that ultimately metastasized. Using a zebrafish model of primary melanoma, we interrogated the top genes from the AVAST-M signature in vivo. This identified GRAMD1B, a cholesterol transfer protein, as a bona fide metastasis suppressor, with a majority of knockout animals rapidly developing metastasis. Mechanistically, excess free cholesterol or its metabolite 27-hydroxycholesterol promotes invasiveness via activation of an AP-1 program, which is associated with increased metastasis in humans. Our data demonstrate that the transcriptional seeds of metastasis are embedded within localized tumors, suggesting that early targeting of these programs can be used to prevent metastatic relapse. Significance: We analyzed human melanoma transcriptomics data to identify a gene signature predictive of metastasis. To rapidly test clinical signatures, we built a genetic metastasis platform in adult zebrafish and identified GRAMD1B as a suppressor of melanoma metastasis. GRAMD1B-associated cholesterol overload activates an AP-1 program to promote melanoma invasion. This article is highlighted in the In This Issue feature, p. 1

We characterized the longitudinal serum protein signatures of women 6 and 10 years after having gestational diabetes mellitus (GDM) to identify factors associated with the development of type 2 diabetes mellitus (T2D) and prediabetes in this at-risk post-GDM population, aiming to discover potential biomarkers for early diagnosis and prevention of T2D. Our study identified 75 T2D-associated serum proteins and 23 prediabetes-associated proteins, some of which were validated in an independent T2D cohort. Machine learning (ML) performed on the longitudinal proteomics highlighted protein signatures associated with progression to post-GDM diabetes. We also proposed prognostic biomarker candidates that were differentially regulated in healthy participants at 6 years postpartum who later progressed to having T2D. Our longitudinal study revealed T2D risk factors for post-GDM populations who are relatively young and healthy, providing insights for clinical decisions and early lifestyle interventions.