Foundation model for cancer imaging biomarkers.

Foundation models in deep learning are characterized by a single large-scale model trained on vast amounts of data serving as the foundation for various downstream tasks. Foundation models are generally trained using self-supervised learning and excel in reducing the demand for training samples in downstream applications. This is especially important in medicine, where large labelled datasets are often scarce. Here, we developed a foundation model for cancer imaging biomarker discovery by training a convolutional encoder through self-supervised learning using a comprehensive dataset of 11,467 radiographic lesions. The foundation model was evaluated in distinct and clinically relevant applications of cancer imaging-based biomarkers. We found that it facilitated better and more efficient learning of imaging biomarkers and yielded task-specific models that significantly outperformed conventional supervised and other state-of-the-art pretrained implementations on downstream tasks, especially when training dataset sizes were very limited. Furthermore, the foundation model was more stable to input variations and showed strong associations with underlying biology. Our results demonstrate the tremendous potential of foundation models in discovering new imaging biomarkers that may extend to other clinical use cases and can accelerate the widespread translation of imaging biomarkers into clinical settings.

Spatial transformation of multi-omics data unlocks novel insights into cancer biology.

The application of next-generation sequencing (NGS) has transformed cancer research. As costs have decreased, NGS has increasingly been applied to generate multiple layers of molecular data from the same samples, covering genomics, transcriptomics, and methylomics. Integrating these types of multi-omics data in a combined analysis is now becoming a common issue with no obvious solution, often handled on an ad hoc basis, with multi-omics data arriving in a tabular format and analyzed using computationally intensive statistical methods. These methods particularly ignore the spatial orientation of the genome and often apply stringent p-value corrections that likely result in the loss of true positive associations. Here, we present GENIUS (GEnome traNsformatIon and spatial representation of mUltiomicS data), a framework for integrating multi-omics data using deep learning models developed for advanced image analysis. The GENIUS framework is able to transform multi-omics data into images with genes displayed as spatially connected pixels and successfully extract relevant information with respect to the desired output. We demonstrate the utility of GENIUS by applying the framework to multi-omics datasets from the Cancer Genome Atlas. Our results are focused on predicting the development of metastatic cancer from primary tumors, and demonstrate how through model inference, we are able to extract the genes which are driving the model prediction and are likely associated with metastatic disease progression. We anticipate our framework to be a starting point and strong proof of concept for multi-omics data transformation and analysis without the need for statistical correction.

The ratio of adaptive to innate immune cells differs between genders and associates with improved prognosis and response to immunotherapy.

Immunotherapy has revolutionised cancer treatment. However, not all cancer patients benefit, and current stratification strategies based primarily on PD1 status and mutation burden are far from perfect. We hypothesised that high activation of an innate response relative to the adaptive response may prevent proper tumour neoantigen identification and decrease the specific anticancer response, both in the presence and absence of immunotherapy. To investigate this, we obtained transcriptomic data from three large publicly available cancer datasets, the Cancer Genome Atlas (TCGA), the Hartwig Medical Foundation (HMF), and a recently published cohort of metastatic bladder cancer patients treated with immunotherapy. To analyse immune infiltration into bulk tumours, we developed an RNAseq-based model based on previously published definitions to estimate the overall level of infiltrating innate and adaptive immune cells from bulk tumour RNAseq data. From these, the adaptive-to-innate immune ratio (A/I ratio) was defined. A meta-analysis of 32 cancer types from TCGA overall showed improved overall survival in patients with an A/I ratio above median (Hazard ratio (HR) females 0.73, HR males 0.86, P < 0.05). Of particular interest, we found that the association was different for males and females for eight cancer types, demonstrating a gender bias in the relative balance of the infiltration of innate and adaptive immune cells. For patients with metastatic disease, we found that responders to immunotherapy had a significantly higher A/I ratio than non-responders in HMF (P = 0.036) and a significantly higher ratio in complete responders in a separate metastatic bladder cancer dataset (P = 0.022). Overall, the adaptive-to-innate immune ratio seems to define separate states of immune activation, likely linked to fundamental immunological reactions to cancer. This ratio was associated with improved prognosis and improved response to immunotherapy, demonstrating potential relevance to patient stratification. Furthermore, by demonstrating a significant difference between males and females that associates with response, we highlight an important gender bias which likely has direct clinical relevance.

Foundation Models for Quantitative Biomarker Discovery in Cancer Imaging.

Foundation models represent a recent paradigm shift in deep learning, where a single large-scale model trained on vast amounts of data can serve as the foundation for various downstream tasks. Foundation models are generally trained using self-supervised learning and excel in reducing the demand for training samples in downstream applications. This is especially important in medicine, where large labeled datasets are often scarce. Here, we developed a foundation model for imaging biomarker discovery by training a convolutional encoder through self-supervised learning using a comprehensive dataset of 11,467 radiographic lesions. The foundation model was evaluated in distinct and clinically relevant applications of imaging-based biomarkers. We found that they facilitated better and more efficient learning of imaging biomarkers and yielded task-specific models that significantly outperformed their conventional supervised counterparts on downstream tasks. The performance gain was most prominent when training dataset sizes were very limited. Furthermore, foundation models were more stable to input and inter-reader variations and showed stronger associations with underlying biology. Our results demonstrate the tremendous potential of foundation models in discovering novel imaging biomarkers that may extend to other clinical use cases and can accelerate the widespread translation of imaging biomarkers into clinical settings.

Body composition and lung cancer-associated cachexia in TRACERx.

Cancer-associated cachexia (CAC) is a major contributor to morbidity and mortality in individuals with non-small cell lung cancer. Key features of CAC include alterations in body composition and body weight. Here, we explore the association between body composition and body weight with survival and delineate potential biological processes and mediators that contribute to the development of CAC. Computed tomography-based body composition analysis of 651 individuals in the TRACERx (TRAcking non-small cell lung Cancer Evolution through therapy (Rx)) study suggested that individuals in the bottom 20th percentile of the distribution of skeletal muscle or adipose tissue area at the time of lung cancer diagnosis, had significantly shorter lung cancer-specific survival and overall survival. This finding was validated in 420 individuals in the independent Boston Lung Cancer Study. Individuals classified as having developed CAC according to one or more features at relapse encompassing loss of adipose or muscle tissue, or body mass index-adjusted weight loss were found to have distinct tumor genomic and transcriptomic profiles compared with individuals who did not develop such features. Primary non-small cell lung cancers from individuals who developed CAC were characterized by enrichment of inflammatory signaling and epithelial-mesenchymal transitional pathways, and differentially expressed genes upregulated in these tumors included cancer-testis antigen MAGEA6 and matrix metalloproteinases, such as ADAMTS3. In an exploratory proteomic analysis of circulating putative mediators of cachexia performed in a subset of 110 individuals from TRACERx, a significant association between circulating GDF15 and loss of body weight, skeletal muscle and adipose tissue was identified at relapse, supporting the potential therapeutic relevance of targeting GDF15 in the management of CAC.

Computational Analysis Reveals the Temporal Acquisition of Pathway Alterations during the Evolution of Cancer.

Cancer metastasis is the lethal developmental step in cancer, responsible for the majority of cancer deaths. To metastasise, cancer cells must acquire the ability to disseminate systemically and to escape an activated immune response. Here, we endeavoured to investigate if metastatic dissemination reflects acquisition of genomic traits that are selected for. We acquired mutation and copy number data from 8332 tumours representing 19 cancer types acquired from The Cancer Genome Atlas and the Hartwig Medical Foundation. A total of 827,344 non-synonymous mutations across 8332 tumour samples representing 19 cancer types were timed as early or late relative to copy number alterations, and potential driver events were annotated. We found that metastatic cancers had a significantly higher proportion of clonal mutations and a general enrichment of early mutations in p53 and RTK/KRAS pathways. However, while individual pathways demonstrated a clear time-separated preference for specific events, the relative timing did not vary between primary and metastatic cancers. These results indicate that the selective pressure that drives cancer development does not change dramatically between primary and metastatic cancer on a genomic level, and is mainly focused on alterations that increase proliferation.

Classifying cGAS-STING Activity Links Chromosomal Instability with Immunotherapy Response in Metastatic Bladder Cancer.

The cGAS-STING pathway serves a critical role in anticancer therapy. Particularly, response to immunotherapy is likely driven by both active cGAS-STING signaling that attracts immune cells, and by the presence of cancer neoantigens that presents as targets for cytotoxic T cells. Chromosomal instability (CIN) is a hallmark of cancer, but also leads to an accumulation of cytosolic DNA that in turn results in increased cGAS-STING signaling. To avoid triggering the cGAS-STING pathway, it is commonly disrupted by cancer cells, either through mutations in the pathway or through transcriptional silencing. Given its effect on the immune system, determining the cGAS-STING activation status prior to treatment initiation is likely of clinical relevance. Here, we used combined expression data from 2,307 tumors from five cancer types from The Cancer Genome Atlas to define a novel cGAS-STING activity score based on eight genes with a known role in the pathway. Using unsupervised clustering, four distinct categories of cGAS-STING activation were identified. In multivariate models, the cGAS-STING active tumors show improved prognosis. Importantly, in an independent bladder cancer immunotherapy-treated cohort, patients with low cGAS-STING expression showed limited response to treatment, while patients with high expression showed improved response and prognosis, particularly among patients with high CIN and more neoantigens. In a multivariate model, a significant interaction was observed between CIN, neoantigens, and cGAS-STING activation. Together, this suggests a potential role of cGAS-STING activity as a predictive biomarker for the application of immunotherapy. Significance: The cGAS-STING pathway is induced by CIN, triggers inflammation and is often deficient in cancer. We provide a tool to evaluate cGAS-STING activity and demonstrate clinical significance in immunotherapy response.

Treatment Represents a Key Driver of Metastatic Cancer Evolution.

Metastasis is the main cause of cancer death, yet the evolutionary processes behind it remain largely unknown. Here, through analysis of large panel-based genomic datasets from the AACR Genomics Evidence Neoplasia Information Exchange project, including 40,979 primary and metastatic tumors across 25 distinct cancer types, we explore how the evolutionary pressure of cancer metastasis shapes the selection of genomic drivers of cancer. The most commonly affected genes were TP53, MYC, and CDKN2A, with no specific pattern associated with metastatic disease. This suggests that, on a driver mutation level, the selective pressure operating in primary and metastatic tumors is similar. The most highly enriched individual driver mutations in metastatic tumors were mutations known to drive resistance to hormone therapies in breast and prostate cancer (ESR1 and AR), anti-EGFR therapy in non-small cell lung cancer (EGFR T790M), and imatinib in gastrointestinal cancer (KIT V654A). Specific mutational signatures were also associated with treatment in three cancer types, supporting clonal selection following anticancer therapy. Overall, this implies that initial acquisition of driver mutations is predominantly shaped by the tissue of origin, where specific mutations define the developing primary tumor and drive growth, immune escape, and tolerance to chromosomal instability. However, acquisition of driver mutations that contribute to metastatic disease is less specific, with the main genomic drivers of metastatic cancer evolution associating with resistance to therapy. SIGNIFICANCE: This study leverages large datasets to investigate the evolutionary landscape of established cancer genes to shed new light upon the mystery of cancer dissemination and expand the understanding of metastatic cancer biology.

Increased Soluble PD-1 Predicts Response to Nivolumab plus Ipilimumab in Melanoma.

BACKGROUND: Checkpoint inhibitors have revolutionized the treatment of metastatic melanoma, yielding long-term survival in a considerable proportion of the patients. Yet, 40-60% of patients do not achieve a long-term benefit from such therapy, emphasizing the urgent need to identify biomarkers that can predict response to immunotherapy and guide patients for the best possible treatment. Here, we exploited an unsupervised machine learning approach to identify potential inflammatory cytokine signatures from liquid biopsies, which could predict response to immunotherapy in melanoma. METHODS: We studied a cohort of 77 patients diagnosed with unresectable advanced-stage melanoma undergoing treatment with first-line nivolumab plus ipilimumab or pembrolizumab. Baseline and on-treatment plasma samples were tested for levels of PD-1, PD-L1, IFNγ, IFNß, CCL20, CXCL5, CXCL10, IL6, IL8, IL10, MCP1, and TNFα and analyzed by Uniform Manifold Approximation and Projection (UMAP) dimension reduction method and k-means clustering analysis. RESULTS: Interestingly, using UMAP analysis, we found that treatment-induced cytokine changes measured as a ratio between baseline and on-treatment samples correlated significantly to progression-free survival (PFS). For patients treated with nivolumab plus ipilimumab we identified a group of patients with superior PFS that were characterized by significantly higher baseline-to-on-treatment increments of PD-1, PD-L1, IFNγ, IL10, CXCL10, and TNFα compared to patients with worse PFS. Particularly, a high PD-1 increment was a strong individual predictor for superior PFS (HR = 0.13; 95% CI 0.034-0.49; p = 0.0026). In contrast, decreasing levels of IFNγ and IL6 and increasing levels of CXCL5 were associated with superior PFS in the pembrolizumab group, although none of the cytokines were individually predictors for PFS. CONCLUSIONS: In short, our study demonstrates that a high increment of PD-1 is associated with superior PFS in advanced-stage melanoma patients treated with nivolumab plus ipilimumab. In contrast, decreasing levels of IFNγ and IL6, and increasing levels of CXCL5 are associated with response to pembrolizumab. These results suggest that using serial samples to monitor changes in cytokine levels early during treatment is informative for treatment response.

An integrated multi-omics analysis identifies prognostic molecular subtypes of non-muscle-invasive bladder cancer.

The molecular landscape in non-muscle-invasive bladder cancer (NMIBC) is characterized by large biological heterogeneity with variable clinical outcomes. Here, we perform an integrative multi-omics analysis of patients diagnosed with NMIBC (n = 834). Transcriptomic analysis identifies four classes (1, 2a, 2b and 3) reflecting tumor biology and disease aggressiveness. Both transcriptome-based subtyping and the level of chromosomal instability provide independent prognostic value beyond established prognostic clinicopathological parameters. High chromosomal instability, p53-pathway disruption and APOBEC-related mutations are significantly associated with transcriptomic class 2a and poor outcome. RNA-derived immune cell infiltration is associated with chromosomally unstable tumors and enriched in class 2b. Spatial proteomics analysis confirms the higher infiltration of class 2b tumors and demonstrates an association between higher immune cell infiltration and lower recurrence rates. Finally, the independent prognostic value of the transcriptomic classes is documented in 1228 validation samples using a single sample classification tool. The classifier provides a framework for biomarker discovery and for optimizing treatment and surveillance in next-generation clinical trials.