Single-Cell Genomics: Enabling the Functional Elucidation of Infectious Diseases in Multi-Cell Genomes

Abstract

Since the time when detection of gene expression in single cells by microarrays to the Next Generation Sequencing (NGS) enabled Single Cell Genomics (SCG), it has played a pivotal role to understand and elucidate the functional role of cellular heterogeneity. Along this journey to becoming a key player in the capture of the individuality of cells, SCG overcame many milestones, including scale, speed, sensitivity and sample costs (4S). There have been many important experimental and computational innovations in the efficient analysis and interpretation of SCG data. The increasing role of AI in SCG data analysis has further enhanced its applicability in building models for clinical intervention. Furthermore, SCG has been instrumental in the delineation of the role of cellular heterogeneity in specific diseases, including cancer and infectious diseases. The understanding of the role of differential immune responses in driving coronavirus disease-2019 (COVID-19) disease severity and clinical outcomes has been greatly aided by SCG. With many variants of concern (VOC) in sight, it would be of great importance to further understand the immune response specificity vis-a-vis the immune cell repertoire, the identification of novel cell types, and antibody response. Given the potential of SCG to play an integral part in the multi-omics approach to the study of the host–pathogen interaction and its outcomes, our review attempts to highlight its strengths, its implications for infectious disease biology, and its current limitations. We conclude that the application of SCG would be a critical step towards future pandemic preparedness.

MatchNAS: Optimizing Edge AI in Sparse-Label Data Contexts via Automating Deep Neural Network Porting for Mobile Deployment

“…The general idea behind NAS is to explore network architecture from a space of different architectural choices, such as the number of layers and operation types. This enables the creation of resource-insensitive models for mobile deployment [5,30,34]. Early NAS [29,32,43,44] suffers from prohibitive resource consumption of training and evaluating every candidate networks, while recent one-shot NAS and zero-shot NAS alleviate this burden by supernet training scheme and architectural scoring scheme, respectively.…”

“…Let 𝑅 represent the resource constraints (e.g., model size limitation or latency requirements) One-shot NAS. One-shot NAS techniques [5,14,30,37,40] proposed to reduce NAS cost by using weight-sharing for all networks in the search space A. One-shot NAS has two important concepts: the "supernet" and the "subnet".…”

“…where ACC(•) refers to the validation accuracy on D 𝑣𝑎𝑙 . To reduce the evaluation cost, recent methods [5,30,37,40] tend to train an accuracy predictor by evaluating a small set of subnets. One-shot NAS provides a low-cost scheme for training and searching lightweight networks for mobile deployment.…”

“…Recent work [30] noted that, compared to training a huge network family, training a smaller network family can alleviate the inference among different size subnets. We hypothesise that our semi-supervised-supernet training will also benefit from a smaller search space.…”

“…Zero-shot NAS [7,22,26] designs specific metrics for network performance evaluation, which can effectively obtain optimum network architectures for different platforms, avoiding manual network designing. One-shot NAS [5,14,30,34] trains a large family of networks with similar architecture simultaneously, thereby avoiding the need to train and fine-tune networks separately for each platform. The neural architecture search technology enables researchers to obtain a large number of trained DNN models for mobile porting in the short term with less human effort and lower computing resource overhead.…”