Owing to the advent of high throughput single cell transcriptomics, past few years have seen exponential growth in production of gene expression data. Recently efforts have been made by various research groups to homogenize and store single cell expression from a large number of studies. The true value of this ever increasing data deluge can be unlocked by making it searchable. To this end, we propose CellAtlasSearch, a novel search architecture for high dimensional expression data, which is massively parallel as well as light-weight, thus infinitely scalable. In CellAtlasSearch, we use a Graphical Processing Unit (GPU) friendly version of Locality Sensitive Hashing (LSH) for unmatched speedup in data processing and query. Currently, CellAtlasSearch features over 300 000 reference expression profiles including both bulk and single-cell data. It enables the user query individual single cell transcriptomes and finds matching samples from the database along with necessary meta information. CellAtlasSearch aims to assist researchers and clinicians in characterizing unannotated single cells. It also facilitates noise free, low dimensional representation of single-cell expression profiles by projecting them on a wide variety of reference samples. The web-server is accessible at: http://www.cellatlassearch.com.

No abstract

Topic modeling is frequently employed for discovering structures (or patterns) in a corpus of documents. Its utility in text-mining and document retrieval tasks in various fields of scientific research is rather well known. An unsupervised machine learning approach, Latent Dirichlet Allocation (LDA) has particularly been utilized for identifying latent (or hidden) topics in document collections and for deciphering the words that define one or more topics using a generative statistical model. Here we describe how SARS-CoV-2 genomic mutation profiles can be structured into a Bag of Words (BoW) to enable identification of signatures (topics) and their probabilistic distribution across various genomes using LDA. Topic models were generated using ~47000 novel corona virus genomes (considered as documents), leading to identification of 16 amino acid mutation signatures and 18 nucleotide mutation signatures (equivalent to topics) in the corpus of chosen genomes through coherence optimization. The document assumption for genomes also helped in identification of contextual nucleotide mutation signatures in the form of conventional N-grams (e.g. bi-grams and tri-grams). We validated the signatures obtained using LDA driven method against the previously reported phylogenetic clades for genomes. Additionally, we report the distribution of the identified mutation signatures on the global map of SARS-CoV-2 genomes. Use of the non-phylogenetic albeit classical approaches like topic modeling and other data centric pattern mining algorithms is therefore proposed for supplementing the efforts towards understanding the genomic diversity of the evolving SARS-CoV-2 genomes (and other pathogens/microbes).

Motivation: Continuous emergence of new variants through appearance, accumulation and disappearance of mutations in viruses is a hallmark of many viral diseases. SARS-CoV-2 and its variants have particularly exerted tremendous pressure on global healthcare system owing to their life threatening and debilitating implications. The sheer plurality of the variants and huge scale of genome sequence data available for Covid19 have added to the challenges of traceability of mutations of concern. The latter however provides an opportunity to utilize SARS-CoV-2 genomes and the mutations therein as "big data records" to comprehensively classify the variants through the (machine) learning of mutation patterns. The unprecedented sequencing effort and tracing of dis-ease outcomes provide an excellent ground for identifying important mutations by developing ma-chine learnt models or severity classifiers using mutation profile of SARS-CoV-2. This is expected to provide a significant impetus to the efforts towards not only identifying the mutations of concern but also exploring the potential of mutation driven predictive prognosis of SARS-CoV-2. Results: We describe how a graduated approach of building various severity specific machine learning classifiers, using only the mutation corpus of SARS-CoV-2 genomes, can potentially lead to the identification of important mutations and guide potential prognosis of infection. We demonstrate the applicability of model derived important mutations and use of Shapley values in order to identify the significant mutations of concern as well as for developing sparse models of outcome classification. A total of 77,284 outcome traced SARS-CoV-2 genomes were employed in this study which represented a total corpus of 30346 unique nucleotide mutations and 18647 amino acid mutations. Machine learning models pertaining to graduated classifiers of target outcomes namely "Asymptomatic, Mild, Symptomatic/Moderate, Severe and Fatal" were built considering the TRIPOD guidelines for predictive prognosis. Shapley values for model linked important mutations were employed to select significant mutations leading to identification of less than 20 outcome driving mutations from each classifier. We additionally describe the significance of adopting a "temporal modeling approach" to benchmark the predictive prognosis linked with continuously evolving pathogens. A chronologically distinct sampling is important in evaluating the performance of models trained on "past data" in accurately classifying prognosis linked with genomes of future (observed with new mutations). We conclude that while machine learning approach can play a vital role in identifying relevant mutations, caution should be exercised in using the mutation signatures for predictive prognosis in cases where new mutations have accumulated along with the previously observed mutations of concern.

No abstract