Multimodal data fusion for cancer biomarker discovery with deep learning.

Technological advances now make it possible to study a patient from multiple angles with high-dimensional, high-throughput multi-scale biomedical data. In oncology, massive amounts of data are being generated ranging from molecular, histopathology, radiology to clinical records. The introduction of deep learning has significantly advanced the analysis of biomedical data. However, most approaches focus on single data modalities leading to slow progress in methods to integrate complementary data types. Development of effective multimodal fusion approaches is becoming increasingly important as a single modality might not be consistent and sufficient to capture the heterogeneity of complex diseases to tailor medical care and improve personalised medicine. Many initiatives now focus on integrating these disparate modalities to unravel the biological processes involved in multifactorial diseases such as cancer. However, many obstacles remain, including lack of usable data as well as methods for clinical validation and interpretation. Here, we cover these current challenges and reflect on opportunities through deep learning to tackle data sparsity and scarcity, multimodal interpretability, and standardisation of datasets.

Development and Assessment of an Artificial Intelligence-Based Tool for Ptosis Measurement in Adult Myasthenia Gravis Patients Using Selfie Video Clips Recorded on Smartphones.

Introduction: Myasthenia gravis (MG) is a rare autoimmune disease characterized by muscle weakness and fatigue. Ptosis (eyelid drooping) occurs due to fatigue of the muscles for eyelid elevation and is one symptom widely used by patients and healthcare providers to track progression of the disease. Margin reflex distance 1 (MRD1) is an accepted clinical measure of ptosis and is typically assessed using a hand-held ruler. In this work, we develop an AI model that enables automated measurement of MRD1 in self-recorded video clips collected using patient smartphones. Methods: A 3-month prospective observational study collected a dataset of video clips from patients with MG. Study participants were asked to perform an eyelid fatigability exercise to elicit ptosis while filming "selfie" videos on their smartphones. These images were collected in nonclinical settings, with no in-person training. The dataset was annotated by non-clinicians for (1) eye landmarks to establish ground truth MRD1 and (2) the quality of the video frames. The ground truth MRD1 (in millimeters, mm) was calculated from eye landmark annotations in the video frames using a standard conversion factor, the horizontal visible iris diameter of the human eye. To develop the model, we trained a neural network for eye landmark detection consisting of a ResNet50 backbone plus two dense layers of 78 dimensions on publicly available datasets. Only the ResNet50 backbone was used, discarding the last two layers. The embeddings from the ResNet50 were used as features for a support vector regressor (SVR) using a linear kernel, for regression to MRD1, in mm. The SVR was trained on data collected remotely from MG patients in the prospective study, split into training and development folds. The model's performance for MRD1 estimation was evaluated on a separate test fold from the study dataset. Results: On the full test fold (N = 664 images), the correlation between the ground truth and predicted MRD1 values was strong (r = 0.732). The mean absolute error was 0.822 mm; the mean of differences was -0.256 mm; and 95% limits of agreement (LOA) were -0.214-1.768 mm. Model performance showed no improvement when test data were gated to exclude "poor" quality images. Conclusions: On data generated under highly challenging real-world conditions from a variety of different smartphone devices, the model predicts MRD1 with a strong correlation (r = 0.732) between ground truth and predicted MRD1.

A decentralized, prospective, observational study to collect real-world data from patients with myasthenia gravis using smartphones.

Introduction: We conducted a 3-month, prospective study in a population of patients with Myasthenia Gravis (MG), utilizing a fully decentralized approach for recruitment and monitoring (ClinicalTrials.gov Identifier: NCT04590716). The study objectives were to assess the feasibility of collecting real-world data through a smartphone-based research platform, in order to characterize symptom involvement during MG exacerbations. Methods: Primary data collection included daily electronically recorded patient-reported outcomes (ePROs) on the presence of MG symptoms, the level of symptom severity (using the MG-Activities of Daily Living assessment, MG-ADL), and exacerbation status. Participants were also given the option to contribute data on their physical activity levels from their own wearable devices. Results: The study enrolled and onboarded 113 participants across 37 US states, and 73% (N= 82) completed the study. The mean age of participants was 53.6 years, 60% were female. Participants were representative of a moderate to severe MG phenotype, with frequent exacerbations, high symptom burden and multiple comorbidities. 55% of participants (N=45) reported MG exacerbations during the study, with an average of 6.3 exacerbation days per participant. Median average MG-ADL scores for participants during self-reported exacerbation and non-exacerbation periods were 7 (interquartile range 4-9, range 1-19) and 0.3 (interquartile range 0-0.8, range 0-9), respectively. Analyses examining relationships between patient-reported and patient-generated health data streams and exacerbation status demonstrated concordance between self-reported MG-ADL scores and exacerbation status, and identified features that may be used to understand and predict the onset of MG symptom exacerbations, including: 1.) dynamic changes in day-to-day symptom reporting and severity 2.) daily step counts as a measure of physical activity and 3.) clinical characteristics of the patient, including the amount of time since their initial diagnosis and their active medications related to MG treatment. Finally, application of unsupervised machine learning methods identified unique clusters of exacerbation subtypes, each with their own specific representation of symptoms and symptom severity. Conclusion: While these symptom signatures require further study and validation, our results suggest that digital phenotyping, characterized by increased multidimensionality and frequency of the data collection, holds promise for furthering our understanding of clinically significant exacerbations and reimagining the approach to treating MG as a heterogeneous condition.

Multimodal deep learning to predict prognosis in adult and pediatric brain tumors.

BACKGROUND: The introduction of deep learning in both imaging and genomics has significantly advanced the analysis of biomedical data. For complex diseases such as cancer, different data modalities may reveal different disease characteristics, and the integration of imaging with genomic data has the potential to unravel additional information than when using these data sources in isolation. Here, we propose a DL framework that combines these two modalities with the aim to predict brain tumor prognosis. METHODS: Using two separate glioma cohorts of 783 adults and 305 pediatric patients we developed a DL framework that can fuse histopathology images with gene expression profiles. Three strategies for data fusion were implemented and compared: early, late, and joint fusion. Additional validation of the adult glioma models was done on an independent cohort of 97 adult patients. RESULTS: Here we show that the developed multimodal data models achieve better prediction results compared to the single data models, but also lead to the identification of more relevant biological pathways. When testing our adult models on a third brain tumor dataset, we show our multimodal framework is able to generalize and performs better on new data from different cohorts. Leveraging the concept of transfer learning, we demonstrate how our pediatric multimodal models can be used to predict prognosis for two more rare (less available samples) pediatric brain tumors. CONCLUSIONS: Our study illustrates that a multimodal data fusion approach can be successfully implemented and customized to model clinical outcome of adult and pediatric brain tumors.

An increasing amount of complex patient data is generated when treating patients with cancer, including histopathology data (where the appearance of a tumor is examined under a microscope) and molecular data (such as analysis of a tumor's genetic material). Computational methods to integrate these data types might help us to predict outcomes in patients with cancer. Here, we propose a deep learning method which involves computer software learning from patterns in the data, to combine histopathology and molecular data to predict outcomes in patients with brain cancers. Using three cohorts of patients, we show that our method combining the different datasets performs better than models using one data type. Methods like ours might help clinicians to better inform patients about their prognosis and make decisions about their care.

mRNA Capture Sequencing and RT-qPCR for the Detection of Pathognomonic, Novel, and Secondary Fusion Transcripts in FFPE Tissue: A Sarcoma Showcase.

We assess the performance of mRNA capture sequencing to identify fusion transcripts in FFPE tissue of different sarcoma types, followed by RT-qPCR confirmation. To validate our workflow, six positive control tumors with a specific chromosomal rearrangement were analyzed using the TruSight RNA Pan-Cancer Panel. Fusion transcript calling by FusionCatcher confirmed these aberrations and enabled the identification of both fusion gene partners and breakpoints. Next, whole-transcriptome TruSeq RNA Exome sequencing was applied to 17 fusion gene-negative alveolar rhabdomyosarcoma (ARMS) or undifferentiated round cell sarcoma (URCS) tumors, for whom fluorescence in situ hybridization (FISH) did not identify the classical pathognomonic rearrangements. For six patients, a pathognomonic fusion transcript was readily detected, i.e., PAX3-FOXO1 in two ARMS patients, and EWSR1-FLI1, EWSR1-ERG, or EWSR1-NFATC2 in four URCS patients. For the 11 remaining patients, 11 newly identified fusion transcripts were confirmed by RT-qPCR, including COPS3-TOM1L2, NCOA1-DTNB, WWTR1-LINC01986, PLAA-MOB3B, AP1B1-CHEK2, and BRD4-LEUTX fusion transcripts in ARMS patients. Additionally, recurrently detected secondary fusion transcripts in patients diagnosed with EWSR1-NFATC2-positive sarcoma were confirmed (COPS4-TBC1D9, PICALM-SYTL2, SMG6-VPS53, and UBE2F-ALS2). In conclusion, this study shows that mRNA capture sequencing enhances the detection rate of pathognomonic fusions and enables the identification of novel and secondary fusion transcripts in sarcomas.

Clinical use of the mRNA urinary biomarker SelectMDx test for prostate cancer.

BACKGROUND: Molecular biomarker tests are developed as diagnostic tools for prostate cancer (PCa) diagnosis. The SelectMDx (MDxHealth, Nijmegen, The Netherlands) test is a urinary-based biomarker test intended to be used to predict presence of high-grade PCa upon biopsy in men with elevated serum prostate-specific antigen (PSA) levels. Previous validation of the SelectMDx test revealed that 53% of the unnecessary biopsies (biopsies indicating no- or GG1 PCa) could be avoided using the SelectMDx test as a decision-tool to select men for prostate biopsy. The objective of this study is to examine the use of the commercially available SelectMDx test under routine, real-life practice. METHODS: Men that underwent a SelectMDx test between May 2019 and December 2020 and that were originating from countries that perform the SelectMDx test on a regular basis were included in this study, resulting in 5157 cases from 10 European countries. Clinical parameters, urinary RNA scores, and test outcomes were compared between PSA groups, age groups, countries, and the validation cohort (described previously [4]) using the Mann-Whitney U test, Chi-Square test, Benjamini-Hochberg and Kruskal-Wallis tests. RESULTS: 40.72% of the cases received a negative SelectMDx result. The test is also used in patients outside the intended-use population (PSA < 3 and >10 ng/mL). Clinical parameters (age, PSA density, DRE outcome) varied between patient population from individual countries and the validation cohort, resulting in differences in the potential number of saved biopsies using the test. CONCLUSIONS: The potential number of reduced biopsies in clinical use was 40,72% using the SelectMDx test, assuming a negative SelectMDx test resulted in the decision not to biopsy the patient. This is higher compared to the validation cohort, which is explained by differences in patient population.

Clinically relevant aberrant Filip1l DNA methylation detected in a murine model of cutaneous squamous cell carcinoma.

BACKGROUND: Cutaneous squamous cell carcinomas (cSCC) are among the most common and highly mutated human malignancies. Understanding the impact of DNA methylation in cSCC may provide avenues for new therapeutic strategies. METHODS: We used reduced-representation bisulfite sequencing for DNA methylation analysis of murine cSCC. Differential methylation was assessed at the CpG level using limma. Next, we compared with human cSCC Infinium HumanMethylation BeadArray data. Genes were considered to be of major relevance when they featured at least one significantly differentially methylated CpGs (RRBS) / probes (Infinium) with at least a 30% difference between tumour vs. control in both a murine gene and its human orthologue. The human EPIC Infinium data were used to distinguish two cSCC subtypes, stem-cell-like and keratinocyte-like tumours. FINDINGS: We found increased average methylation in mouse cSCC (by 12.8%, p = 0.0011) as well as in stem-cell like (by 3.1%, p=0.002), but not keratinocyte-like (0.2%, p = 0.98), human cSCC. Comparison of differentially methylated genes revealed striking similarities between human and mouse cSCC. Locus specific methylation changes in mouse cSCC often occurred in regions of potential regulatory function, including enhancers and promoters. A key differentially methylated region was located in a potential enhancer of the tumour suppressor gene Filip1l and its expression was reduced in mouse tumours. Moreover, the FILIP1L locus showed hypermethylation in human cSCC and lower expression in human cSCC cell lines. INTERPRETATION: Deregulation of DNA methylation is an important feature of murine and human cSCC that likely contributes to silencing of tumour suppressor genes, as shown for Filip1l. FUNDING: British Skin Foundation, Cancer Research UK.

DNA Methylation Regulates Transcription Factor-Specific Neurodevelopmental but Not Sexually Dimorphic Gene Expression Dynamics in Zebra Finch Telencephalon.

Song learning in zebra finches (Taeniopygia guttata) is a prototypical example of a complex learned behavior, yet knowledge of the underlying molecular processes is limited. Therefore, we characterized transcriptomic (RNA-sequencing) and epigenomic (RRBS, reduced representation bisulfite sequencing; immunofluorescence) dynamics in matched zebra finch telencephalon samples of both sexes from 1 day post hatching (1 dph) to adulthood, spanning the critical period for song learning (20 and 65 dph). We identified extensive transcriptional neurodevelopmental changes during postnatal telencephalon development. DNA methylation was very low, yet increased over time, particularly in song control nuclei. Only a small fraction of the massive differential expression in the developing zebra finch telencephalon could be explained by differential CpG and CpH DNA methylation. However, a strong association between DNA methylation and age-dependent gene expression was found for various transcription factors (i.e., OTX2, AR, and FOS) involved in neurodevelopment. Incomplete dosage compensation, independent of DNA methylation, was found to be largely responsible for sexually dimorphic gene expression, with dosage compensation increasing throughout life. In conclusion, our results indicate that DNA methylation regulates neurodevelopmental gene expression dynamics through steering transcription factor activity, but does not explain sexually dimorphic gene expression patterns in zebra finch telencephalon.

Multicenter Optimization and Validation of a 2-Gene mRNA Urine Test for Detection of Clinically Significant Prostate Cancer before Initial Prostate Biopsy.

PURPOSE: A 2-gene, urine based molecular test that combines mRNA biomarkers with clinical factors can risk stratify patients for clinically significant prostate cancer. To ensure the generalizability of assay results we optimized and validated the clinical model for men with serum prostate specific antigen less than 10 ng/ml who were undergoing initial prostate biopsy. MATERIALS AND METHODS: Urine samples were collected from 1,955 men from The Netherlands, France and Germany prior to an initial prostate biopsy and study subjects were divided into training and validation cohorts. Urinary HOXC6 and DLX1 mRNA levels were quantified and RNA results were then combined with other risk factors in a clinical model optimized to detect ISUP (International Society of Urological Pathology) Grade Group 2 or greater prostate cancer in men with prostate specific antigen less than 10 ng/ml. Results in the validation cohort were compared with the PCPTRC (Prostate Cancer Prevention Trial Risk Calculator), version 2.0. RESULTS: The optimal clinical model included urinary HOXC6 and DLX1 mRNA levels, patient age, digital rectal examination and prostate specific antigen density (serum prostate specific antigen/prostate volume). In the 715 validation cohort subjects with prostate specific antigen less than 10 ng/ml the AUC was 0.82 with 89% sensitivity, 53% specificity and 95% negative predictive value. The PCPTRC AUC was 0.70. The full validation cohort of 916 men including all prostate specific antigen levels yielded an AUC of 0.85 with 93% sensitivity, 47% specificity and 95% negative predictive value. The PCPTRC AUC was 0.76. CONCLUSIONS: The 2-gene based urine assay, which is optimized for biopsy naïve patients with serum prostate specific antigen less than 10 ng/ml, demonstrated high sensitivity and negative predictive value to detect clinically significant prostate cancer. These data support using the test to help guide initial prostate biopsy decisions.

A comprehensive overview of genomic imprinting in breast and its deregulation in cancer.

Genomic imprinting plays an important role in growth and development. Loss of imprinting (LOI) has been found in cancer, yet systematic studies are impeded by data-analytical challenges. We developed a methodology to detect monoallelically expressed loci without requiring genotyping data, and applied it on The Cancer Genome Atlas (TCGA, discovery) and Genotype-Tissue expression project (GTEx, validation) breast tissue RNA-seq data. Here, we report the identification of 30 putatively imprinted genes in breast. In breast cancer (TCGA), HM13 is featured by LOI and expression upregulation, which is linked to DNA demethylation. Other imprinted genes typically demonstrate lower expression in cancer, often associated with copy number variation and aberrant DNA methylation. Downregulation in cancer frequently leads to higher relative expression of the (imperfectly) silenced allele, yet this is not considered canonical LOI given the lack of (absolute) re-expression. In summary, our novel methodology highlights the massive deregulation of imprinting in breast cancer.

Prevalent, Dynamic, and Conserved R-Loop Structures Associate with Specific Epigenomic Signatures in Mammals.

R-loops are three-stranded nucleic acid structures formed upon annealing of an RNA strand to one strand of duplex DNA. We profiled R-loops using a high-resolution, strand-specific methodology in human and mouse cell types. R-loops are prevalent, collectively occupying up to 5% of mammalian genomes. R-loop formation occurs over conserved genic hotspots such as promoter and terminator regions of poly(A)-dependent genes. In most cases, R-loops occur co-transcriptionally and undergo dynamic turnover. Detailed epigenomic profiling revealed that R-loops associate with specific chromatin signatures. At promoters, R-loops associate with a hyper-accessible state characteristic of unmethylated CpG island promoters. By contrast, terminal R-loops associate with an enhancer- and insulator-like state and define a broad class of transcription terminators. Together, this suggests that the retention of nascent RNA transcripts at their site of expression represents an abundant, dynamic, and programmed component of the mammalian chromatin that affects chromatin patterning and the control of gene expression.

A genome-wide search for epigenetically [corrected] regulated genes in zebra finch using MethylCap-seq and RNA-seq.

Learning and memory formation are known to require dynamic CpG (de)methylation and gene expression changes. Here, we aimed at establishing a genome-wide DNA methylation map of the zebra finch genome, a model organism in neuroscience, as well as identifying putatively epigenetically regulated genes. RNA- and MethylCap-seq experiments were performed on two zebra finch cell lines in presence or absence of 5-aza-2'-deoxycytidine induced demethylation. First, the MethylCap-seq methodology was validated in zebra finch by comparison with RRBS-generated data. To assess the influence of (variable) methylation on gene expression, RNA-seq experiments were performed as well. Comparison of RNA-seq and MethylCap-seq results showed that at least 357 of the 3,457 AZA-upregulated genes are putatively regulated by methylation in the promoter region, for which a pathway analysis showed remarkable enrichment for neurological networks. A subset of genes was validated using Exon Arrays, quantitative RT-PCR and CpG pyrosequencing on bisulfite-treated samples. To our knowledge, this study provides the first genome-wide DNA methylation map of the zebra finch genome as well as a comprehensive set of genes of which transcription is under putative methylation control.

PROTEOFORMER: deep proteome coverage through ribosome profiling and MS integration.

An increasing amount of studies integrate mRNA sequencing data into MS-based proteomics to complement the translation product search space. However, several factors, including extensive regulation of mRNA translation and the need for three- or six-frame-translation, impede the use of mRNA-seq data for the construction of a protein sequence search database. With that in mind, we developed the PROTEOFORMER tool that automatically processes data of the recently developed ribosome profiling method (sequencing of ribosome-protected mRNA fragments), resulting in genome-wide visualization of ribosome occupancy. Our tool also includes a translation initiation site calling algorithm allowing the delineation of the open reading frames (ORFs) of all translation products. A complete protein synthesis-based sequence database can thus be compiled for mass spectrometry-based identification. This approach increases the overall protein identification rates with 3% and 11% (improved and new identifications) for human and mouse, respectively, and enables proteome-wide detection of 5'-extended proteoforms, upstream ORF translation and near-cognate translation start sites. The PROTEOFORMER tool is available as a stand-alone pipeline and has been implemented in the galaxy framework for ease of use.

SNP-guided identification of monoallelic DNA-methylation events from enrichment-based sequencing data.

Monoallelic gene expression is typically initiated early in the development of an organism. Dysregulation of monoallelic gene expression has already been linked to several non-Mendelian inherited genetic disorders. In humans, DNA-methylation is deemed to be an important regulator of monoallelic gene expression, but only few examples are known. One important reason is that current, cost-affordable truly genome-wide methods to assess DNA-methylation are based on sequencing post-enrichment. Here, we present a new methodology based on classical population genetic theory, i.e. the Hardy-Weinberg theorem, that combines methylomic data from MethylCap-seq with associated SNP profiles to identify monoallelically methylated loci. Applied on 334 MethylCap-seq samples of very diverse origin, this resulted in the identification of 80 genomic regions featured by monoallelic DNA-methylation. Of these 80 loci, 49 are located in genic regions of which 25 have already been linked to imprinting. Further analysis revealed statistically significant enrichment of these loci in promoter regions, further establishing the relevance and usefulness of the method. Additional validation was done using both 14 whole-genome bisulfite sequencing data sets and 16 mRNA-seq data sets. Importantly, the developed approach can be easily applied to other enrichment-based sequencing technologies, like the ChIP-seq-based identification of monoallelic histone modifications.

A proteogenomics approach integrating proteomics and ribosome profiling increases the efficiency of protein identification and enables the discovery of alternative translation start sites.

Next-generation transcriptome sequencing is increasingly integrated with MS to enhance MS-based protein and peptide identification. Recently, a breakthrough in transcriptome analysis was achieved with the development of ribosome profiling (ribo-seq). This technology is based on the deep sequencing of ribosome-protected mRNA fragments, thereby enabling the direct observation of in vivo protein synthesis at the transcript level. In order to explore the impact of a ribo-seq-derived protein sequence search space on MS/MS spectrum identification, we performed a comprehensive proteome study on a human cancer cell line, using both shotgun and N-terminal proteomics, next to ribosome profiling, which was used to delineate (alternative) translational reading frames. By including protein-level evidence of sample-specific genetic variation and alternative translation, this strategy improved the identification score of 69 proteins and identified 22 new proteins in the shotgun experiment. Furthermore, we discovered 18 new alternative translation start sites in the N-terminal proteomics data and observed a correlation between the quantitative measures of ribo-seq and shotgun proteomics with a Pearson correlation coefficient ranging from 0.483 to 0.664. Overall, this study demonstrated the benefits of ribosome profiling for MS-based protein and peptide identification and we believe this approach could develop into a common practice for next-generation proteomics.