#single-cell #preprocessing #rna-seq #single-nucleus #rna-velocity

bin+lib alevin-fry

A suite of tools for the rapid, accurate and memory-frugal processing single-cell and single-nucleus sequencing data

10 unstable releases (4 breaking)

0.8.1 Jan 13, 2023
0.8.0 Oct 11, 2022
0.7.0 Aug 2, 2022
0.6.0 Jun 1, 2022
0.4.1 Jul 22, 2021

#8 in Science

Download history 2/week @ 2022-10-15 5/week @ 2022-10-22 19/week @ 2022-10-29 25/week @ 2022-11-05 1/week @ 2022-11-12 4/week @ 2022-11-19 1/week @ 2022-11-26 10/week @ 2022-12-03 22/week @ 2022-12-10 14/week @ 2022-12-17 3/week @ 2022-12-31 29/week @ 2023-01-07 7/week @ 2023-01-14 5/week @ 2023-01-21 19/week @ 2023-01-28

60 downloads per month

Custom license

6.5K SLoC


alevin-fry Rust Anaconda-Server Badge Anaconda-Server Badge Anaconda-Server Badge GitHub tag (latest SemVer)

alevin-fry is a suite of tools for the rapid, accurate and memory-frugal processing single-cell and single-nucleus sequencing data. It consumes RAD files generated by salmon alevin, and performs common operations like generating permit lists, and estimating the number of distinct molecules from each gene within each cell. The focus in alevin-fry is on safety, accuracy and efficiency (in terms of both time and memory usage).

You can read the paper describing alevin fry, "Alevin-fry unlocks rapid, accurate, and memory-frugal quantification of single-cell RNA-seq data" here, and the pre-print on bioRxiv.

Getting started with alevin-fry

There are many resources to help you get started with alevin-fry, and to demonstrate how to perform various tasks and analyses in an idiomatic way with alevin-fry --- for example, you can find a lot of information below in this README. Nonetheless, as we attempt to restructure, centralize, and refresh the documentation for alevin-fry, we intend for the the landing page and centralized jumping-off point to be the Read The Docs page for alevin-fry.

More information

  • Quickstart guide with a unified singularity container

  • Relationship to alevin: Alevin-fry has been designed as the successor to alevin. It subsumes the core features of alevin, while also providing important new capabilities and considerably improving the performance profile. We anticipate that new method development and feature additions will take place primarily within the alevin-fry codebase. Thus, we encourage users of alevin to migrate to alevin-fry when feasible. That being said, alevin is still actively-maintained and supported, so if you are using it and not ready to migrate you can continue to ask questions and post issues in the salmon repository.


Alevin-fry is under active development. However, you can find the documentation on read the docs. We try to keep the documentation up to date with the latest developments in the software. Additionally, there is a series of tutorial for using alevin-fry for processing different types of data that you can find here.


Are you curious about processing details like whether to use a sparse or dense index? Do you have a question that isn't necessarily a bug report or feature request, and that isn't readily answered by the documentation or tutorials? Then please feel free to ask over in the Q&A.

Sister repositories

The generation of the reduced alignment data (RAD) files processed by alevin-fry is done by salmon. The latest version of salmon is available on GitHub, via bioconda, and on dockerhub.

The usefulaf repository contains scripts in functions that are useful in helping to prepare input for alevin-fry processing, importing alevin-fry output into downstream analysis evnironemnts, and even running common configurations of alevin-fry more simply. This repository also contains the relevant Python function for loading fry output (specifically in USA mode) in a convenient way into scanpy (i.e. as AnnData objects) for subsequent Python-based processing in scanpy.

The roe and pyroe repositories provide tools to help easily construct a splici transcriptome from a reference genome and GTF file in R and python respectively.

The fishpond package — maintained by @mikelove and his lab — contains the recommended relevant functions for reading alevin-fry output (particularly USA-mode output) into the R ecosystem, in the form of a singleCellExperiment object.

The alevinqc package — maintained by @csoneson — provides tool and functions for performing quality control and assessment downstream of alevin-fry.

Installing from bioconda

Alevin-fry is available for both x86 linux and OSX platforms using bioconda.

With bioconda in the appropriate place in your channel list, you should simply be able to install via:

$ conda install alevin-fry

Installing from crates.io

Alevin-fry can also be installed from crates.io using cargo. This can be done with the following command:

$ cargo install alevin-fry

Building from source

If you want to use features or fixes that may only be available in the latest develop branch (or want to build for a different architecture), then you have to build from source. Luckily, cargo makes that easy; see below.

Alevin-fry is built and tested with the latest (major & minor) stable version of Rust. While it will likely compile fine with slightly older versions of Rust, this is not a guarantee and is not a support priority. Unlike with C++, Rust has a frequent and stable release cadence, is designed to be installed and updated from user space, and is easy to keep up to date with rustup. Thanks to cargo, building should be as easy as:

$ cargo build --release

subsequent commands below will assume that the executable is in your path. Temporarily, this can be done (in bash-like shells) using:

$ export PATH=`pwd`/target/release/:$PATH

A quick start run through on sample data

Here, we show how to perform a complete analysis on the 1k PBMCs from a Healthy Donor data from 10X Genomics. This run through includes all steps, even extracting the splici sequence and building the salmon index, which you typically would not do per-sample. To make this sample as easy as possible to follow, we have bundled all of the required software and utilities in a singularity container that we use in the commands below.

Download input data and singularity container

First, create a working directory with sufficient space to download all of the input data and to hold the output (50GB should be sufficient). We alias this directory and use the alias below so that you can easily set it to something else if you want and still copy and paste the later commands.

$ mkdir af_test_workdir
$ export AF_SAMPLE_DIR=$PWD/af_test_workdir

Then, we download all of our test data. This consist of the human reference genome and annotation (based, in this case, on the CellRanger 3.0 reference annotation) and the FASTQ files from the PBMC1k (v3) healthy donor samples.

$ mkdir -p human_CR_3.0/fasta
$ mkdir -p human_CR_3.0/genes
$ wget -v -O human_CR_3.0/fasta/genome.fa -L https://umd.box.com/shared/static/3kuh1lc03bxg1d3hi1jfloez7zoutfjc
$ wget -v -O human_CR_3.0/genes/genes.gtf -L https://umd.box.com/shared/static/tvyg43710ufuuvp8mnuoanowm6xmkbjk
$ mkdir -p data/pbmc_1k_v3_fastqs
$ wget -v -O data/pbmc_1k_v3_fastqs/pbmc_1k_v3_S1_L001_R2_001.fastq.gz -L https://umd.box.com/shared/static/bmhtt9db8ojhmbkb6d98mt7fnsdhsymm
$ wget -v -O data/pbmc_1k_v3_fastqs/pbmc_1k_v3_S1_L002_R2_001.fastq.gz -L https://umd.box.com/shared/static/h8ymvs2njqiygfsu50jla2uce6p6etke
$ wget -v -O data/pbmc_1k_v3_fastqs/pbmc_1k_v3_S1_L001_R1_001.fastq.gz -L https://umd.box.com/shared/static/hi8mkx1yltmhnl9kn22n96xtic2wqm5i
$ wget -v -O data/pbmc_1k_v3_fastqs/pbmc_1k_v3_S1_L002_R1_001.fastq.gz -L https://umd.box.com/shared/static/4sn4pio63kk7pku52eo3xg9ztf5tq1ul

Finally, we'll download the singularity image that contains all of the software we'll need to do our processing.

$ wget -v -O usefulaf.sif https://umd.box.com/shared/static/bcd8io9fbjc321pfgcomues5oe2a12cz

or, alternatively, you can pull the docker image directly from Dockerhub and have singularity convert it for you

$ singularity pull docker://combinelab/usefulaf:latest

Info about the singularity container

The singularity container we just downloaded above contains a recent release of salmon (v1.8.0) and alevin-fry (v0.5.1), as well as an installation of R and all of the packages needed to build the splici index.

To build the reference index (and quantify) we'll use the simpleaf wrapper. This is a shell script written around salmon, alevin-fry, and the splici index construction code that simplifies processing by grouping together related commands, using a fixed directory structure for processing, and also by eliminating some different options that are otherwise exposed by salmon and alevin-fry (e.g. it builds the sparse index, maps in sketch mode etc.). If you would like to run the "raw" commands, the Singularity image contains salmon and alevin-fry in the path, and the R script to construct the splici index at /usefulaf/R/build_splici_ref.R, so you can explore more detailed options.

Building the splici reference and index

To build our reference index (this will both extract the splici fasta and transcript to gene mapping, and build the salmon index on it), use the following command (this should generally take ~1hr or less):

$ singularity exec --cleanenv \
--bind $AF_SAMPLE_DIR:/workdir \
--pwd /usefulaf/bash usefulaf.sif \
./simpleaf index \
-f /workdir/human_CR_3.0/fasta/genome.fa \
-g /workdir/human_CR_3.0/genes/genes.gtf \
-l 91 -t 16 -o /workdir/human_CR_3.0_splici

Quantifying the sample

Given the constructed index (which will be written by the above command to $AF_SAMPLE_DIR/human_CR_3.0_splici/index), the next step is to quantify the sample against this index. This can be done with the following command (this should generally take only a few minutes), which will run salmon to generate the RAD file in sketch mode, perform unfiltered permit-list generation --- automatically downloading the appropriate external permit-list --- collate the RAD file and quantify the gene counts using the cr-like strategy):

$ singularity exec --cleanenv \
--bind $AF_SAMPLE_DIR:/workdir \
--pwd /usefulaf/bash usefulaf.sif \
./simpleaf quant \
-1 /workdir/data/pbmc_1k_v3_fastqs/pbmc_1k_v3_S1_L001_R1_001.fastq.gz,/workdir/data/pbmc_1k_v3_fastqs/pbmc_1k_v3_S1_L002_R1_001.fastq.gz \
-2 /workdir/data/pbmc_1k_v3_fastqs/pbmc_1k_v3_S1_L001_R2_001.fastq.gz,/workdir/data/pbmc_1k_v3_fastqs/pbmc_1k_v3_S1_L002_R2_001.fastq.gz \
-i /workdir/human_CR_3.0_splici/index \
-o /workdir/quants/pbmc1k_v3 \
-f u -c v3 -r cr-like \
-m /workdir/human_CR_3.0_splici/ref/transcriptome_splici_fl86_t2g_3col.tsv \
-t 16


At the end of this process, the directory $AF_SAMPLE_DIR/quants/pbmc1k_v3/quant will have the final output of running alevin-fry's quant command. The alevin subdirectory will include a file specifying the row names (cell barcode), the column names (unspliced, spliced and ambiguous genes) and the counts (in MTX coordinate format). You can load these counts up in your favorite analysis environment to explore further.

R : In R, you can make use of the R load_fry() function here, and read the input with the command:

m <- load_fry("$AF_SAMPLE_DIR/quants/pbmc1k_v3/quant")

where $AF_SAMPLE_DIR is appropriately replaced by the path to the working directory we chose at the start of this exercise. This will return a SingleCellExperiment object containing the counts for this experiment. The stand-alone load_fry() function is part of fishpond, and the function is documented in detail here.

Python : In python, you can make use of the python load_fry() function, which relies on scanpy. To read the input you can use the following command:

m = load_fry("$AF_SAMPLE_DIR/quants/pbmc1k_v3/quant")

where, again $AF_SAMPLE_DIR is appropriately replaced by the path to the working directory we chose at the start of this exercise. This will return a scanpy AnnData object with the counts.

From here, you can use your favorite downstream analysis packages (like Seurat in R or scanpy in python) to perform quality control, filtering and analysis of your data.

A note about preparing a splici (spliced + intron) reference

In the manuscript describing alevin-fry, we primarily make use of an index that is built over spliced + intron sequence, which we refer to as a splici reference. This is also what we build in the quick start example above. To make the construction of the relevant reference sequence (and the 3 column TSV file you will need for Unspliced/Spliced/Ambiguous (USA) quantification) simple, we have written an R script that will process a genome and GTF file and produce the splici reference which you can then index with salmon as normal.

First, checkout the usefulaf repository and navigate to the R directory. Then, we'll run the build_splici_ref.R script.

$ ./build_splici_ref.R <path_to_genome_fasta> <path_to_gtf> <target_read_length> <output_dir>

where $ indicated your command prompt. In addition to these required positional arguments, there are a few optional arguments that you can find by running

$ ./build_splici_ref.R -h

After you have run this script, your output directory should contain 3 files:


The first file contains the splici reference sequence that you should index with salmon, and the third contains the 3-column transcript-to-gene mapping that you should pass to alevin-fry during the quant phase.

If you have any questions about preparing the splici reference, or otherwise about processing your data with alevin-fry please feel free to open an issue here on GitHub!

Citing alevin-fry

If you use alevin-fry in your work, please cite:

He, D., Zakeri, M., Sarkar, H. et al. Alevin-fry unlocks rapid, accurate and memory-frugal quantification of single-cell RNA-seq data. Nat Methods 19, 316322 (2022). https://doi.org/10.1038/s41592-022-01408-3


author={He, Dongze and Zakeri, Mohsen and Sarkar, Hirak and Soneson, Charlotte and Srivastava, Avi and Patro, Rob},
title={Alevin-fry unlocks rapid, accurate and memory-frugal quantification of single-cell RNA-seq data},
journal={Nature Methods},


~737K SLoC