Negative Matrix Factorization Research Articles

Identifying SLC2A6 as the novel protective factor in breast cancer by TP53-related genes affecting M1 macrophage infiltration.

The high heterogeneity of breast cancer (BC) caused by pathogenic gene mutations poses a challenge to immunotherapy, but the underlying mechanism remains unknown. The difference in the infiltration of M1 macrophages induced by TP53 mutations has a significant impact on BC immunotherapy. The aim of this study was to develop a TP53-related M1 macrophage infiltration molecular typing risk signature in BC and evaluate the biological functions of the key gene to find new immunotherapy biomarkers. Weighted correlation network analysis (WGCNA) and negative matrix factorization (NMF) were used for distinguishing BC subtypes. The signature and the nomogram were both constructed and evaluated. Biological functions of the novel signature gene SLC2A6 were confirmed through in vitro and in vivo experiments. RNA-Sequencing and protein profiling were used for detecting the possible mechanism of SLC2A6. The results suggested that four BC subtypes were distinguished by TP53-related genes that affect M1 macrophage infiltration. The signature constructed by molecular typing characteristics could evaluate BC's clinical features and tumor microenvironment. The nomogram could accurately predict the prognosis. The signature gene SLC2A6 was found to have an abnormally low expression in tumor tissues. Overexpression of SLC2A6 could inhibit proliferation, promote mitochondrial damage, and result in apoptosis of tumor cells. The HSP70 family member protein HSPA6 could bind with SLC2A6 and increase with the increased expression of SLC2A6. In summary, the risk signature provides a reference for BC risk assessment, and the signature gene SLC2A6 could act as a tumor suppressor in BC.

Abstract 3494: KINOMO: A non-negative matrix factorization framework for recovering intra- and inter-tumoral heterogeneity from single-cell RNA-seq data

Abstract Single-cell RNA-sequencing (scRNA-seq) is a powerful technology to uncover cellular heterogeneity in tumor ecosystems. Due to differences in underlying gene load, direct comparison between patient samples is challenging, and this is further complicated by the sparsity of data matrices in scRNA-seq. Here, we present a factorization method called KINOMO (Kernel dIfferentiability correlation-based NOn- negative Matrix factorization algorithm using Kullback-Leibler divergence loss Optimization). This tool uses quadratic approximation approach for error correction and an iterative multiplicative approach, which improves the quality assessment of NMF-identified factorization, while mitigating biases introduced by inter-patient genomic variability. We benchmarked this new approach against nine different methods across 150 scRNA-seq experiments and find that KINOMO outperforms prior methods when evaluated with an adjusted Rand index (ARI), ranging 0.82-0.91 compared to 0.68-0.77. Thus, KINOMO provides an improved approach for determining coherent transcriptional programs (and meta-programs) from scRNA- seq data of cancer tissues, enabling comparison of patients with variable genomic backgrounds. Citation Format: Somnath Tagore, Yiping Wang, Jana Biermann, Edridge K. D'Souza, Karan D. Luthria, Raul Rabadan, Elham Azizi, Benjamin Izar. KINOMO: A non-negative matrix factorization framework for recovering intra- and inter-tumoral heterogeneity from single-cell RNA-seq data [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2024; Part 1 (Regular Abstracts); 2024 Apr 5-10; San Diego, CA. Philadelphia (PA): AACR; Cancer Res 2024;84(6_Suppl):Abstract nr 3494.

Spectroscopic analysis improvement using convolutional neural networks

Removing noisy components of signals coming from edge tokamak plasmas, astrochemical organic matter or astronomical objects is one of the promising path to improve the underlying elements identification. Methods such as the penalized semi-supervised non negative matrix factorization (PSNMF) used to extract such prominent elements perform well on complex signals. However, it is results’ confidence decreases as the noise increases. In this context, we have tried to address this limitation by removing part of the undesired noise in atomic spectra using artificial intelligence (AI) method based on convolutional neural networks (CNNs). More specifically, we have tested different architectures of CNN classically used in denoising task: residual CNNs and auto-encoders, to benchmark their respective denoising capacity. The dataset used is made of high resolution atom and ion spectra extracted from the NIST Atomic Spectra Database. In the case of ions, we have considered several degrees of ionization. The synthetic added noise is generated from a typical experimental noise profile randomly modified for each signal. Performance of each AI-based noise remover is measured by analyzing the increase of the element identification precision obtained by the PSNMF. Auto-encoders and residual CNN are both suitable to remove the undesired noise and increase the PSNMF efficiency even for noise-prevailing signals. In this particular case, the auto-encoder architecture seems to be globally more accurate and should be selected when considering noisy multi-element atomic spectra.

A novel PD-1/PD-L1 pathway molecular typing-related signature for predicting prognosis and the tumor microenvironment in breast cancer

BackgroundCurrently, the development of breast cancer immunotherapy based on the PD-1/PD-L1 pathway is relatively slow, and the specific mechanism affecting the immunotherapy efficacy in breast cancer is still unclear.MethodsWeighted correlation network analysis (WGCNA) and the negative matrix factorization (NMF) were used to distinguish subtypes related to the PD-1/PD-L1 pathway in breast cancer. Then univariate Cox, least absolute shrinkage and selection operator (LASSO), and multivariate Cox regression were used to construct the prognostic signature. A nomogram was established based on the signature. The relationship between the signature gene IFNG and breast cancer tumor microenvironment was analyzed.ResultsFour PD-1/PD-L1 pathway-related subtypes were distinguished. A prognostic signature related to PD-1/PD-L1 pathway typing was constructed to evaluate breast cancer’s clinical characteristics and tumor microenvironment. The nomogram based on the RiskScore could be used to accurately predict breast cancer patients’ 1-year, 3-year, and 5-year survival probability. The expression of IFNG was positively correlated with CD8+ T cell infiltration in the breast cancer tumor microenvironment.ConclusionA prognostic signature is constructed based on the PD-1/PD-L1 pathway typing in breast cancer, which can guide the precise treatment of breast cancer. The signature gene IFNG is positively related to CD8+ T cell infiltration in breast cancer.

Model selection and robust inference of mutational signatures using Negative Binomial non-negative matrix factorization

BackgroundThe spectrum of mutations in a collection of cancer genomes can be described by a mixture of a few mutational signatures. The mutational signatures can be found using non-negative matrix factorization (NMF). To extract the mutational signatures we have to assume a distribution for the observed mutational counts and a number of mutational signatures. In most applications, the mutational counts are assumed to be Poisson distributed, and the rank is chosen by comparing the fit of several models with the same underlying distribution and different values for the rank using classical model selection procedures. However, the counts are often overdispersed, and thus the Negative Binomial distribution is more appropriate.ResultsWe propose a Negative Binomial NMF with a patient specific dispersion parameter to capture the variation across patients and derive the corresponding update rules for parameter estimation. We also introduce a novel model selection procedure inspired by cross-validation to determine the number of signatures. Using simulations, we study the influence of the distributional assumption on our method together with other classical model selection procedures. We also present a simulation study with a method comparison where we show that state-of-the-art methods are highly overestimating the number of signatures when overdispersion is present. We apply our proposed analysis on a wide range of simulated data and on two real data sets from breast and prostate cancer patients. On the real data we describe a residual analysis to investigate and validate the model choice.ConclusionsWith our results on simulated and real data we show that our model selection procedure is more robust at determining the correct number of signatures under model misspecification. We also show that our model selection procedure is more accurate than the available methods in the literature for finding the true number of signatures. Lastly, the residual analysis clearly emphasizes the overdispersion in the mutational count data. The code for our model selection procedure and Negative Binomial NMF is available in the R package SigMoS and can be found at https://github.com/MartaPelizzola/SigMoS.

COHERENCE NETWORKS OF BILATERAL UPPER LIMB MOTIONS IN CHRONIC STROKE PATIENTS

Understanding the mechanisms of how the skeletal system reduces multiple degrees of freedom in motor control is very important in the field of neurorehabilitation. We therefore need to understand the functional connectivity of these synergistic patterns in bilateral upper limb of stroke patients from different Brunnstrom Stages. The intermuscular coherence at different frequency bands is decomposed to obtain muscle synchronization. Muscle networks were then applied to describe the functional connectivity between muscle pairs. We recruited 35 stroke patients with Brunnstrom Stages III-VI. Muscle activity was acquired from seven muscles around forearm while participants performed voluntary upward reaching with both affected and unaffected sides. We decomposed IMC spectra by negative matrix factorization to identify shared frequencies and co-modulated muscles. Two muscle networks were found and functional muscle networks are evaluated by strength, transitivity and global efficiency. Results showed that the network topology was different significantly between the two sides and between Brunnstrom stages of stroke patients. These findings can throw light on the understanding of motor control and rehabilitation of motor impairment for stroke survivors, which may promote the post-stroke rehabilitation process by detecting impaired neuromuscular coordination in frequency domain.

A non-convex optimization framework for large-scale low-rank matrix factorization

Low-rank matrix factorization problems such as non negative matrix factorization (NMF) can be categorized as a clustering or dimension reduction technique. The latter denotes techniques designed to find representations of some high dimensional dataset in a lower dimensional manifold without a significant loss of information. If such a representation exists, the features ought to contain the most relevant features of the dataset. Many linear dimensionality reduction techniques can be formulated as a matrix factorization. In this paper, we combine the conjugate gradient (CG) method with the Barzilai and Borwein (BB) gradient method, and propose a BB scaling CG method for NMF problems. The new method does not require to compute and store matrices associated with Hessian of the objective functions. Moreover, adopting a suitable BB step size along with a proper nonmonotone strategy which comes by the size of convex parameter ηk, results in a new algorithm that can significantly improve the CPU time, efficiency, the number of function evaluation. Convergence result is established and numerical comparisons of methods on both synthetic and real-world datasets show that the proposed method is efficient in comparison with existing methods and demonstrate the superiority of our algorithms.

Exhaust and non-exhaust contributions from road transport to PM10 at a Southern European traffic site

It is a well – established fact that road traffic is one of the main contributors to ambient levels of airborne particulate matter (APM). This study was carried out at a traffic site in which the PM10 levels are monitored all year round. A trend analysis of these levels revealed that over a decade there was no discernible trend, with the PM10 concentrations normally hovering around the EU limit values. In 2018, one of these limit values was exceeded. The contribution of traffic at the site was therefore investigated through a chemical speciation of 209 PM10 samples collected throughout this year. The speciation data were used in a source apportionment exercise in which the output of the PMF model was further refined using the lesser–known, constraint weighted non – negative matrix factorization (CW – NMF) model. This technique enabled the isolation of two factors clearly related to traffic, which were labelled as “exhaust contribution” (responsible for 3.4% of the PM10), “tire/brake wear contribution” (contributing 17% of the PM10). Additionally, a factor including both traffic resuspended dust and crustal material was also isolated and labelled “road dust/crustal” factor. The two contributors to the factor jointly contribute 18% to the PM10 and the contribution of the traffic resuspended dust was estimated at 7.3%. The traffic resuspended component of this factor together with the “tire/brake wear contribution” jointly make up the non-exhaust contribution of traffic – derived dust. Consonant with what has been known for quite some time, the exhaust fraction is the minor component of traffic PM10. It is therefore, clear that policies aimed at controlling traffic derived PM10 pollution at the receptor will have a minimal effect unless the non – exhaust emissions are adequately controlled.

A Review on Speech Emotion Recognition Using Machine Learning

This paper focuses on the development of a robust speech emotion recognition system using a combination of different speech features with feature optimization techniques and speech de-noising technique to acquire improved emotion classification accuracy, decreasing the system complexity and obtain noise robustness. Additionally, we create original methods for SER to merge features. We employ feature optimization methods that are based on the feature transformation and feature selection machine learning techniques in order to build SER. The following is a list of the upcoming events. A neural network can use either of these two techniques. As more feelings are taken into account, the feature fusion-acquired SER accuracy falls short of expectations, and the plague of dimensionality starts to spread due to the addition of speech features, which makes the SER system work harder to complete its task. This is due to the SER system becoming more complicated when voice elements are added. Therefore, it is crucial to create a SER system that is more trustworthy, has the most practical features, and uses the least amount of computing power possible. By using strategies that maximize current features, it is possible to streamline the feature selection process by reducing the total number of accessible choices to a more reasonable level. This piece employs a method known as Semi-Non Negative Matrix Factorization to lessen the amount of processing trash that the SER system generates. (Semi-NMF). This approach can be used to change traits that are capable of learning on their own.

Immune-and Metabolism-Associated Molecular Classification of Ovarian Cancer.

Ovarian cancer (OV) is a complex gynecological disease, and its molecular characteristics are not clear. In this study, the molecular characteristics of OV subtypes based on metabolic genes were explored through the comprehensive analysis of genomic data. A set of transcriptome data of 2752 known metabolic genes was used as a seed for performing non negative matrix factorization (NMF) clustering. Three subtypes of OV (C1, C2 and C3) were found in analysis. The proportion of various immune cells in C1 was higher than that in C2 and C3 subtypes. The expression level of immune checkpoint genes TNFRSF9 in C1 was higher than that of other subtypes. The activation scores of cell cycle, RTK-RAS, Wnt and angiogenesis pathway and ESTIMATE immune scores in C1 group were higher than those in C2 and C3 groups. In the validation set, grade was significantly correlated with OV subtype C1. Functional analysis showed that the extracellular matrix related items in C1 subtype were significantly different from other subtypes. Drug sensitivity analysis showed that C2 subtype was more sensitive to immunotherapy. Survival analysis of differential genes showed that the expression of PXDN and CXCL11 was significantly correlated with survival. The results of tissue microarray immunohistochemistry showed that the expression of PXDN was significantly correlated with tumor size and pathological grade. Based on the genomics of metabolic genes, a new OV typing method was developed, which improved our understanding of the molecular characteristics of human OV.

Comprehensive Genomic and Epigenomic Analyses on Transcriptomic Regulation in Stomach Adenocarcinoma.

Background: DNA methylation (MET)–mediated transcriptomic disturbance and copy number variations (CNVs) exert a significant influence in stimulating the heterogeneous progression of stomach adenocarcinoma (STAD). Nevertheless, the relation of DNA MET with CNVs, together with its impact on tumor occurrence, is still unclear. Methods: The messenger RNA (mRNA) expression (EXP) profiles, DNA MET, and DNA copy numbers, together with STAD mutation data, were collected from the TCGA official data portal. We employed circular binary segmentation algorithm in “DNAcopy.” library of R package for mapping DNA CNV data at genetic level for all samples based on the segmented CNV data. Stable clusters of samples were recognized using negative matrix factorization cluster analysis based on 50 iterations and the “brunet” method using the MET-correlated (METcor) and CNV-correlated (CNVcor) genes. The R package “iCluster” method was utilized to comprehensively analyze the EXP, MET, and DNA CNV profiles. Results: A total of 313 STAD samples were isolated for checking DNA copy numbers and MET and for measuring EXP. In accordance with our results, we discovered obvious co-regulation of CNVcor genes and METcor counterparts. Apart from that, these genes were subject to multi-omics integration. Meanwhile, three subtypes of STAD were detected and confirmed based on independent data. Among them, the subtype with increased aggressiveness was related to decreased mutation frequencies of ARID1A, PIK3CA, ZFHX3, SPECC1, OBSCN, KMT2D, FSIP2, ZBTB20, TTN, and RANBP2, together with the abnormal levels of JPH3, KCNB1, and PLCXD3. Conclusion: According to the results, these aforementioned genes exerted crucial roles in the development of invasive STAD. Our findings on transcriptomic regulation genomically and epigenetically facilitate the understanding of the STAD pathology from different aspects, which help to develop efficient anti-STAD therapy.

Raman spectroscopy and group and basis-restricted non negative matrix factorisation identifies radiation induced metabolic changes in human cancer cells

This work combines single cell Raman spectroscopy (RS) with group and basis restricted non-negative matrix factorisation (GBR-NMF) to identify individual biochemical changes associated with radiation exposure in three human cancer cell lines. The cell lines analysed were derived from lung (H460), breast (MCF7) and prostate (LNCaP) tissue and are known to display varying degrees of radio sensitivity due to the inherent properties of each cell type. The GBR-NMF approach involves the deconstruction of Raman spectra into component biochemical bases using a library of Raman spectra of known biochemicals present in the cells. Subsequently, scores are obtained on each of these bases which can be directly correlated with the contribution of each chemical to the overall Raman spectrum. We validated GBR-NMF through the correlation of GBR-NMF-derived glycogen scores with scores that were previously observed using principal component analysis (PCA). Phosphatidylcholine, glucose, arginine and asparagine showed a distinct differential score pattern between radio-resistant and radio-sensitive cell types. In summary, the GBR-NMF approach allows for the monitoring of individual biochemical radiation-response dynamics previously unattainable with more traditional PCA-based approaches.

Spiny Lobster Sound Identification Based on Blind Source Separation (BSS) for Passive Acoustic Monitoring (PAM)

Spiny lobsters are targeted in capture fisheries due the high market demand and this situation has led to overfishing activities. Therefore, a smart monitoring system is imperative to be developed in handling this issue in order to store the distribution of marine spiny lobsters’ information such as habitat preferences, population density and biology aspects of the lobsters. This study focused on separating the mixed underwater acoustic sound. An output of the separation system was used for Passive Acoustic Monitoring (PAM) application. The objective of this study is to provide estimated source signals from a recorded mixed acoustic signal. The feasibility of extracting spiny lobster sound as the target sound was observed during the experiment. In this study, Blind Source Separation (BSS) approach was employed to estimate the target sound (i.e., spiny lobsters) from the mixed underwater acoustic signal consisting of the sound of spiny lobsters and several man-made and natural sounds. In this investigation, two different blind source separation methods namely Fast Fixed Point Independent Component Analysis (FastICA) and Non Negative Matrix Factorization (NMF) were implemented. The mixture of one target sound which was spiny lobster and four interferences signals were used as input in this experiment,. The separated source sound by using FastICA with Negentropy, FastICA with Kurtosis and NMF algorithms were compared and evaluated based on bss_eval_sources toolbox metrics. These metrics consisted of signal to distortion ratio (SDR), signal to interferences ratio (SIR) and signal to artifacts ratio (SAR). In conclusion, the FastICA with Negentropy technique provides the best performance in separating mixed signal based on SIR, SAR and SDR measurement results that showed the FastICA with Negentropy generated the highest average values compared to the FastICA with Kurtosis and NMF.

Non-negative matrix factorization: A useful method for two-way life tables

In statistical applications, two-way tables of numeric data, samples by variables, are often analyzed using principal component analysis (PCA), a special case of singular value decomposition (SVD). The scores from PCA are often useful for helping to understand relationships among the samples of the table. We would also like to understand the relationships among the variables and for that purpose PCA loadings are usually difficult to interpret. If the entries in the table are positive or zero, then non-negative matrix factorization (NMF) allows better interpretations of the variables. In this paper, we compare PCA, SVD, and NMF using USA mortality data covering the years 1968 to 2010. The goal is to explicate negative matrix factorization using a data set familiar to actuaries.

Positive and Negative Label-Driven Nonnegative Matrix Factorization

Positive label is often used as the supervisory information in the learning scenario, which refers to the category that a sample is assigned to. However, another side information lying in the labels, which describes the categories that a sample is exclusive of, have been largely ignored. In this paper, we propose a nonnegative matrix factorization (NMF) based classification method leveraging both positive and negative label information, which is termed as positive and negative label-driven NMF (PNLD-NMF). The proposed scheme concurrently accomplishes data representation and classification in a joint manner. Owing to the complementary characteristics between positive and negative labels, we further design a new regularization framework to take advantage of these two label types. Extensive experiments on six image classification benchmark datasets show that the proposed scheme is able to consistently deliver better classification accuracy.

31P Accurate detection of HRD status in multiple cancer types using somatic mutation pattern

Homologous Recombination Deficiency (HRD) is caused by various molecular alterations in genes that participate in HR repair. However, HR is a complicated biological process and it is unlikely that all HRD causing alterations have been uncovered so far. We hypothesized that HRD detection based on mutation pattern can accurately detect HRD status with better patient stratification. We retrieved mutations of ovarian cancer from TCGA repository. Samples treated with platinum chemotherapy were selected for further analysis. Mutation pattern were profiled using negative matrix factorization (NMF) method. Samples were tiered based on weight of HRD-related mutation signature. Cox model was applied to compare survival of samples of each tier against non-HRD samples separately. To test the feasibility of our method on targeted sequencing data, mutation profile was shrunk to selected genomic region composed of 825 genes (∼ 2.2Mb). We retrieved 192 in-house clinical samples and labeled HRD status based on known HRD causing variants. We applied our method on these samples to inspect the consistency. 375 samples were profiled for HRD status and were divided into four tiers based on the weight of HRD-related mutation pattern. Cox model showed significant survival difference (p-value = 0.001) proportional to the weight of HRD mutation pattern with higher weight being the longest survival and vice versa. 223 and 152 samples were categorized as HRD and non-HRD based on optimized threshold and survival difference remains significant (p-value = 0.0017). Similar survival difference was observed (p-value=0.0013) suggesting the validity of the method on targeted sequencing. Out of 124, 49 and 19 breast, ovarian and prostate cancer samples, 13, 11 and 4 of them were labeled as HRD positive based on known HRD causing variants. All but one HRD positive samples were consistently classified by our method. Our results showed that NMF method is able to faithfully detect HRD samples. The method can be applied to WES as well as targeted sequencing and achieve equally significant survival difference between samples with different HR status. Upon detecting HRD, this method can identify additional HRD patients and therefore make better patient stratification.

Analysis of Oral Carcinogenesis Factors Based on Non Negative Matrix Evolutionary Spectrum Clustering Decomposition

To explore the risk factors of Oral Submucosal fibrous carcinogenesis and provide theoretical basis for selfprevention and clinical treatment. First of all, an evolutionary non negative matrix factorization framework based on clustering quality is proposed for dynamic oral carcinogenesis detection. Then, the equivalence among evolutionary spectral clustering, ENMF and density optimization of evolutionary modules is proved theoretically. At the same time, based on this equivalence, a new semi-supervised ENMF algorithm is proposed by adding a priori information to ENMF without increasing the time complexity. Using this algorithm, 42 patients with Oral Submucosal Fibrosis and 40 non-cancerous patients with oral submucosal fibrosis were analyzed by single factor and multi-factor analysis to study the quantitative relationship between Oral Submucosal Fibrosis and 14 related factors. The results suggested that age, betel nut chewing duration, smoking duration and leukoplakia or lichen planus were the four factors which were introduced into the cluster quality analysis. Age, betel nut chewing duration, smoking duration and complications of leukoplakia or lichen planus are risk factors for Oral Submucosal fibrous carcinogenesis.

Convergence of Prognostic Gene Signatures Suggests Underlying Mechanisms of Human Prostate Cancer Progression.

The highly heterogeneous clinical course of human prostate cancer has prompted the development of multiple RNA biomarkers and diagnostic tools to predict outcome for individual patients. Biomarker discovery is often unstable with, for example, small changes in discovery dataset configuration resulting in large alterations in biomarker composition. Our hypothesis, which forms the basis of this current study, is that highly significant overlaps occurring between gene signatures obtained using entirely different approaches indicate genes fundamental for controlling cancer progression. For prostate cancer, we found two sets of signatures that had significant overlaps suggesting important genes (p < 10−34 for paired overlaps, hypergeometrical test). These overlapping signatures defined a core set of genes linking hormone signalling (HES6-AR), cell cycle progression (Prolaris) and a molecular subgroup of patients (PCS1) derived by Non Negative Matrix Factorization (NNMF) of control pathways, together designated as SIG-HES6. The second set (designated SIG-DESNT) consisted of the DESNT diagnostic signature and a second NNMF signature PCS3. Stratifications using SIG-HES6 (HES6, PCS1, Prolaris) and SIG-DESNT (DESNT) classifiers frequently detected the same individual high-risk cancers, indicating that the underlying mechanisms associated with SIG-HES6 and SIG-DESNT may act together to promote aggressive cancer development. We show that the use of combinations of a SIG-HES6 signature together with DESNT substantially increases the ability to predict poor outcome, and we propose a model for prostate cancer development involving co-operation between the SIG-HES6 and SIG-DESNT pathways that has implication for therapeutic design.

Statistical 3D watermarking algorithm using non negative matrix factorization

In this paper, we propose a robust blind watermarking method for 3D object based on Non Negative Matrix Factorization (NMF). Our main idea is to use the NMF basis matrix guide the watermark embedding. We first segment the model into partitions based on its salient points. Then, we apply NMF algorithm to the vertex norms. The distribution of vertex is grouped into bins according to the basis matrix X. Finally, we insert watermark bits in the vertex norm distribution. Experimental results demonstrate that the proposed method achieved good visual quality and also robustness against various attacks.

Mutational signature learning with supervised negative binomial non-negative matrix factorization.

MotivationUnderstanding the underlying mutational processes of cancer patients has been a long-standing goal in the community and promises to provide new insights that could improve cancer diagnoses and treatments. Mutational signatures are summaries of the mutational processes, and improving the derivation of mutational signatures can yield new discoveries previously obscured by technical and biological confounders. Results from existing mutational signature extraction methods depend on the size of available patient cohort and solely focus on the analysis of mutation count data without considering the exploitation of metadata.ResultsHere we present a supervised method that utilizes cancer type as metadata to extract more distinctive signatures. More specifically, we use a negative binomial non-negative matrix factorization and add a support vector machine loss. We show that mutational signatures extracted by our proposed method have a lower reconstruction error and are designed to be more predictive of cancer type than those generated by unsupervised methods. This design reduces the need for elaborate post-processing strategies in order to recover most of the known signatures unlike the existing unsupervised signature extraction methods. Signatures extracted by a supervised model used in conjunction with cancer-type labels are also more robust, especially when using small and potentially cancer-type limited patient cohorts. Finally, we adapted our model such that molecular features can be utilized to derive an according mutational signature. We used APOBEC expression and MUTYH mutation status to demonstrate the possibilities that arise from this ability. We conclude that our method, which exploits available metadata, improves the quality of mutational signatures as well as helps derive more interpretable representations.Availability and implementation https://github.com/ratschlab/SNBNMF-mutsig-public.Supplementary information Supplementary data are available at Bioinformatics online.