Constructor for Class

Description

Constructor to create an instance of FiRE class. There are two ways in which constructor can be created.

Arguments

Details

For usage see example.

Examples

     L <- 100
     M <- 50
     H <- 107881
     seed <- 0
     verbose <- 1

     ## Creating class object with required arguments
     model <- new(FiRE::FiRE, L, M)

     ## Creating class object with all arguments
     model <- new(FiRE::FiRE, L, M, H, seed, verbose)

Finder of Rare Entities (FiRE) - To assign rareness/ outlierness score to every sample

Description

This package sorts samples as per their rareness/ oultierness. Instead of dichotomized decisions, FiRE assigns rareness/ outlierness score to every sample. These scores can then be used to identify rare samples or outliers, with varying degrees of rareness. FiRE takes multiple estimations of the proximity between a pair of samples, in low-dimensional spaces to compute scores.

Details

FiRE is written in c++ and wrapped with Rcpp to create an R interface. To use FiRE, an object with the number of estimators (L), number of dimensions to be sampled per estimator (M), number of bins per estimator (H = 1017881), seed for random number generator (seed = 0) and verbose level (verbose = 0/1) needs to be created. Once the object is created, fit function needs to be called to hash all the samples into bins. Once hashing is done, score function needs to be called to retrieve score for every sample. Sample commands may be seen in the Examples section. The resulting model has four model parameters which may be accessed, dimensions (d) of size M sampled per estimator, thresholds (ths) of size M per d, weights(w) of size M generated per estimator and bin (b) of size H per estimator.

Author(s)

Prashant Gupta, prashant10991@gmail.com

Aashi Jindal, jindal.aashi21@gmail.com

Jayadeva, iitjd4@gmail.com

Debarka Sengupta, debarka@gmail.com

Maintainer: Prashant Gupta <prashant10991@gmail.com>

References

[1]Jindal, A., Gupta, P., Jayadeva and Sengupta, D., 2018. Discovery of rare cells from voluminous single cell expression data. Nature Communications, 9(1), p.4719.

[2]Wang, Z., Dong, W., Josephson, W., Lv, Q., Charikar, M. and Li, K., 2007, June. Sizing sketches: a rank-based analysis for similarity search. In ACM SIGMETRICS Performance Evaluation Review (Vol. 35, No. 1, pp. 157-168). ACM.

Examples

     ## L <- number of estimators
     ## M <- Number of dims to sample
     ## H <- Number of bins
     ## seed <- seed for random number generator
     ## verbose <- verbose level

     library('FiRE')
     data(sample_data) #Samples * Features

     data(sample_label)
     ## Samples with label '1' represent abundant,
     ## while samples with label '2' represent rare.

     ## Saving rownames and colnames
     rnames <- rownames(sample_data)
     cnames <- colnames(sample_data)

     ## Converting data.frame to matrix
     sample_data <- as.matrix(sample_data)
     sample_label <- as.matrix(sample_label)

     L <- 100 # Number of estimators
     M <- 50 # Dims to be sampled

     # Model creation without optional parameter
     model <- new(FiRE::FiRE, L, M)

     ## There are 3 more optional parameters they can be passed as
     ## model <- new(FiRE::FiRE, L, M, H, seed, verbose)

     ## Hashing all samples
     model$fit(sample_data)

     ## Computing score of each sample
     rareness_score <- model$score(sample_data)

     ## Apply IQR-based criteria to identify rare samples for further downstream analysis.
     q3 <- quantile(rareness_score, 0.75)
     iqr <- IQR(rareness_score)
     th <- q3 + (1.5*iqr)

     ## Select indexes that satisfy IQR-based thresholding criteria.
     indIqr <- which(rareness_score >= th)

     ## Create a vector for binary predictions
     predictions <- rep(1, dim(sample_data)[1])
     predictions[indIqr] <- 2 #Replace predictions for rare samples with '2'.

     ## To access model parameters
     ## Parameters - generated weights
     # model$w

     ## Parameters - sample dimensions
     # model$d

     ## Parameters - generated thresholds
     # model$ths

     ## Parameters - Bins
     # model$b

Hash samples in bins

Description

Hashing samples into bins, using Sketching as hash function. This function repeats hashing process L times by sampling M dimensions at a time and hashes samples into one of H bins.

Arguments

Details

For usage see example.

Note

This function does not do any preprocessing, so preprocessed data must be passed if required.

Examples

  ## Not run: 

     ## Creating class object with required arguments
     model <- new(FiRE::FiRE, L, M)
     model$fit(data)

  
## End(Not run)

Preprocessed 293T–Jurkat scRNA-seq dataset

Description

This data set has been generated from 293T and Jurkat cell data containing a total of ~3200 cells, with an almost equal number of representative transcriptomes of each type. The cells were mixed in vitro at equal proportions. Authors of the study resolved the cell types bioinformatically exploiting their SNV profiles. With this data, we mimicked the rare cell phenomenon by bioinformatically diluting Jurkat cell proportion to 2.5% in the data.

Usage

data(sample_data)

Format

A matrix containing 1580 observations (in rows), with 1000 features per observation (in columns).

Source

10xgenomics, https://support.10xgenomics.com/single-cell-gene-expression/datasets

References

Zheng, G. X. et al. Massively parallel digital transcriptional profiling of single cells. Nat. Commun. 8, 14049 (2017).

Labels of preprocessed 293T–Jurkat scRNA-seq dataset

Description

This contains labels for 1580 observations. There are two different clusters, one is abundant (denoted by 1) and the other is rare (denoted by 2). Rare cluster consists of Jurkat cells, which form 2.5% of the total population. The abundant cluster consists of 293T cells.

Usage

data(sample_label)

Format

A vector containing 1580 observations.

Source

10xgenomics, https://support.10xgenomics.com/single-cell-gene-expression/datasets

References

Zheng, G. X. et al. Massively parallel digital transcriptional profiling of single cells. Nat. Commun. 8, 14049 (2017).

Compute score on hashed sample.

Description

Once hashing is done using fit, compute score via proximity estimation.

Arguments

Details

For usage see example.

Examples

  ## Not run: 

     ## Creating class object with required arguments
     model <- new(FiRE::FiRE, L, M)
     model$fit(data)
     score <- model$score(data)

  
## End(Not run)