COSMOS: a platform for real-time morphology-based, label-free cell sorting using deep learning.

Cells are the singular building blocks of life, and a comprehensive understanding of morphology, among other properties, is crucial to the assessment of underlying heterogeneity. We developed Computational Sorting and Mapping of Single Cells (COSMOS), a platform based on Artificial Intelligence (AI) and microfluidics to characterize and sort single cells based on real-time deep learning interpretation of high-resolution brightfield images. Supervised deep learning models were applied to characterize and sort cell lines and dissociated primary tissue based on high-dimensional embedding vectors of morphology without the need for biomarker labels and stains/dyes. We demonstrate COSMOS capabilities with multiple human cell lines and tissue samples. These early results suggest that our neural networks embedding space can capture and recapitulate deep visual characteristics and can be used to efficiently purify unlabeled viable cells with desired morphological traits. Our approach resolves a technical gap in the ability to perform real-time deep learning assessment and sorting of cells based on high-resolution brightfield images.

Accurate whole-genome sequencing and haplotyping from 10 to 20 human cells.

Recent advances in whole-genome sequencing have brought the vision of personal genomics and genomic medicine closer to reality. However, current methods lack clinical accuracy and the ability to describe the context (haplotypes) in which genome variants co-occur in a cost-effective manner. Here we describe a low-cost DNA sequencing and haplotyping process, long fragment read (LFR) technology, which is similar to sequencing long single DNA molecules without cloning or separation of metaphase chromosomes. In this study, ten LFR libraries were made using only ∼100 picograms of human DNA per sample. Up to 97% of the heterozygous single nucleotide variants were assembled into long haplotype contigs. Removal of false positive single nucleotide variants not phased by multiple LFR haplotypes resulted in a final genome error rate of 1 in 10 megabases. Cost-effective and accurate genome sequencing and haplotyping from 10-20 human cells, as demonstrated here, will enable comprehensive genetic studies and diverse clinical applications.

The effects of hepatitis B virus integration into the genomes of hepatocellular carcinoma patients.

Hepatitis B virus (HBV) infection is a leading risk factor for hepatocellular carcinoma (HCC). HBV integration into the host genome has been reported, but its scale, impact and contribution to HCC development is not clear. Here, we sequenced the tumor and nontumor genomes (>80× coverage) and transcriptomes of four HCC patients and identified 255 HBV integration sites. Increased sequencing to 240× coverage revealed a proportionally higher number of integration sites. Clonal expansion of HBV-integrated hepatocytes was found specifically in tumor samples. We observe a diverse collection of genomic perturbations near viral integration sites, including direct gene disruption, viral promoter-driven human transcription, viral-human transcript fusion, and DNA copy number alteration. Thus, we report the most comprehensive characterization of HBV integration in hepatocellular carcinoma patients. Such widespread random viral integration will likely increase carcinogenic opportunities in HBV-infected individuals.

Computational techniques for human genome resequencing using mated gapped reads.

Unchained base reads on self-assembling DNA nanoarrays have recently emerged as a promising approach to low-cost, high-quality resequencing of human genomes. Because of unique characteristics of these mated pair reads, existing computational methods for resequencing assembly, such as those based on map-consensus calling, are not adequate for accurate variant calling. We describe novel computational methods developed for accurate calling of SNPs and short substitutions and indels (<100 bp); the same methods apply to evaluation of hypothesized larger, structural variations. We use an optimization process that iteratively adjusts the genome sequence to maximize its a posteriori probability given the observed reads. For each candidate sequence, this probability is computed using Bayesian statistics with a simple read generation model and simplifying assumptions that make the problem computationally tractable. The optimization process iteratively applies one-base substitutions, insertions, and deletions until convergence is achieved to an optimum diploid sequence. A local de novo assembly procedure that generalizes approaches based on De Bruijn graphs is used to seed the optimization process in order to reduce the chance of converging to local optima. Finally, a correlation-based filter is applied to reduce the false positive rate caused by the presence of repetitive regions in the reference genome.

The mutation spectrum revealed by paired genome sequences from a lung cancer patient.

Lung cancer is the leading cause of cancer-related mortality worldwide, with non-small-cell lung carcinomas in smokers being the predominant form of the disease. Although previous studies have identified important common somatic mutations in lung cancers, they have primarily focused on a limited set of genes and have thus provided a constrained view of the mutational spectrum. Recent cancer sequencing efforts have used next-generation sequencing technologies to provide a genome-wide view of mutations in leukaemia, breast cancer and cancer cell lines. Here we present the complete sequences of a primary lung tumour (60x coverage) and adjacent normal tissue (46x). Comparing the two genomes, we identify a wide variety of somatic variations, including >50,000 high-confidence single nucleotide variants. We validated 530 somatic single nucleotide variants in this tumour, including one in the KRAS proto-oncogene and 391 others in coding regions, as well as 43 large-scale structural variations. These constitute a large set of new somatic mutations and yield an estimated 17.7 per megabase genome-wide somatic mutation rate. Notably, we observe a distinct pattern of selection against mutations within expressed genes compared to non-expressed genes and in promoter regions up to 5 kilobases upstream of all protein-coding genes. Furthermore, we observe a higher rate of amino acid-changing mutations in kinase genes. We present a comprehensive view of somatic alterations in a single lung tumour, and provide the first evidence, to our knowledge, of distinct selective pressures present within the tumour environment.

Analysis of genetic inheritance in a family quartet by whole-genome sequencing.

We analyzed the whole-genome sequences of a family of four, consisting of two siblings and their parents. Family-based sequencing allowed us to delineate recombination sites precisely, identify 70% of the sequencing errors (resulting in > 99.999% accuracy), and identify very rare single-nucleotide polymorphisms. We also directly estimated a human intergeneration mutation rate of approximately 1.1 x 10(-8) per position per haploid genome. Both offspring in this family have two recessive disorders: Miller syndrome, for which the gene was concurrently identified, and primary ciliary dyskinesia, for which causative genes have been previously identified. Family-based genome analysis enabled us to narrow the candidate genes for both of these Mendelian disorders to only four. Our results demonstrate the value of complete genome sequencing in families.

Human genome sequencing using unchained base reads on self-assembling DNA nanoarrays.

Genome sequencing of large numbers of individuals promises to advance the understanding, treatment, and prevention of human diseases, among other applications. We describe a genome sequencing platform that achieves efficient imaging and low reagent consumption with combinatorial probe anchor ligation chemistry to independently assay each base from patterned nanoarrays of self-assembling DNA nanoballs. We sequenced three human genomes with this platform, generating an average of 45- to 87-fold coverage per genome and identifying 3.2 to 4.5 million sequence variants per genome. Validation of one genome data set demonstrates a sequence accuracy of about 1 false variant per 100 kilobases. The high accuracy, affordable cost of $4400 for sequencing consumables, and scalability of this platform enable complete human genome sequencing for the detection of rare variants in large-scale genetic studies.