Structure-preserving visualisation of high dimensional single-cell datasets

Abstract

Single-cell technologies offer an unprecedented opportunity to effectively characterize cellular heterogeneity in health and disease. Nevertheless, visualisation and interpretation of these multi-dimensional datasets remains a challenge. We present a novel framework, ivis, for dimensionality reduction of single-cell expression data. ivis utilizes a siamese neural network architecture that is trained using a novel triplet loss function. Results on simulated and real datasets demonstrate that ivis preserves global data structures in a low-dimensional space, adds new data points to existing embeddings using a parametric mapping function, and scales linearly to hundreds of thousands of cells. ivis is made publicly available through Python and R interfaces on https://github.com/beringresearch/ivis .

Supervised and unsupervised language modelling in Chest X-Ray radiological reports

“…Performance of this three-step approach was comparable to the CheXpert Labeller, which utilises hand-crafted rules tailored for CXR report annotation [4,23]. We have previously applied ivis to structured single-cell datasets [24], demonstrating that the algorithm reliably preserves local and global distances in a low-dimensional space. Briefly, ivis employs a Siamese Neural Network architecture that learns to discriminate between similar and dissimilar fastText vectors without imposing strong priors.…”

“…Dimensionality reduction using siamese neural networks. The unsupervised ivis algorithm [24] was used to reduce dimensionality of 50-dimensional fastText embeddings of unlabelled reports within the GGC corpus. To obtain report-level embeddings, fastText word vectors within each report were averaged [37] and the resulting 50-dimensional vector was used to as inputs into the ivis algorithm.…”

“…The embedding layers use a linear activation and have their weights initialized using Glorot's uniform distribution. The network was trained using a triplet loss function, whilst Euclidean distance was used to establish similarity between points in the embedding space [24]. Nearest neighbour selection was limited to 130 points and the training was halted early if the triplet loss did not improve for five epochs.…”

“…We evaluate the utility of classical supervised machine learning techniques as well as state-of-the-art Long Short-Term Memory networks (LSTM) in the context of large corpora of free-text reports from Greater Glasgow and Clyde Health Board (n = 500,000) and the Beth Israel Deaconess Medical Center (MIMIC-CXR database [n = 227,835]) [23]. Additionally, we use ivis, an unsupervised Siamese Neural Network-based algorithm [24], which accurately classifies radiological reports and visualises document embeddings. Finally, we explore generalisability of machine learning techniques across European and North American radiological report corpora.…”

scGAE: topology-preserving dimensionality reduction for single-cell RNA-seq data using graph autoencoder

“…Many dimensionality reduction methods have been developed or introduced for scRNA-seq data analyses in the past several years. Recently developed competitive methods include DCA 7 , SCVI 8 , scDeepCluster 9 , PHATE 10 , SAUCIE 11 , and Ivis 12 . Among them, deep learning showed the greatest potentials.…”