README.md

slide-recon

Sample command to run pipeline for generation of sparse count matrix from a pair of Illumina fastq files:

python /path/to/slide-recon/bc_umi_pipeline.py \
        -c cores \
        -i input_directory \
        -s sample_name \
        -ma max_r1_bcs \
        -mt max_r2_bcs \
        -r1 read1_structure \
        -r2 read2_structure

This pipeline expects two fastq files with format sample_name_R[1-2]_001.fastq.gz inside input_direcory. These are steps in the pipeline which uses functions within bc_umi_utils.py module.

1. Extracting and splitting the fastq files into an intermediate directory with path input_direcory/sample_name/split. The number of reads per chunk can be modified with paramter lines within bc_umi_utils.split_fastq_by_lines function and has default of 10m reads, or 40m lines.

2. Function extract_bc_umi_dict perform main pre-processing step on the raw reads

   • Fucntion parse_read_struct takes read1_structure and read2_structure from the input and converts the sequences into intervals for different landmark on the read, namely, barcode, UMI, polyT/A sequences, and other adapters placed in the bead oligo structure.

   • Function edit_match identifies reads containing appropriate landmarks by matching constant sequences of apadpters in known locations by calculating edit distances using edlib package.

   • Passing raw reads are written to the disk in csv format. Each line is a quadruple of R1_BC,R1_UMI,R2_BC,R2_UMI. also json files are stored for all R1_BC_R1_UMI combinations and R2_BC_R2_UMI combinations.

   • This is done in parallel for each fastq part using multiprocessing module.

3. Function aggregate_dicts aggregates R1 and R2 json files from all parts and stores in somewhat huge new json files.

   • This can be slow and final json file can be quite large due to potentially 100s of millions of raw barcodes. One solution is to ignore UMI information at this stage.

   • Function aggregate_stat_dicts aggregaitng barcode matching statistic for each of the adapters. This statitics is edit distance in the following format:

     r1_adapter1_..._r1_adapterN__r2_adapter1_..._r2_adapterN.

     For example -1_1__0_3 means r1_adapter1 didn't match, r1_adapter2 matched with 1 distance, and so on.

4. Function whitelist_rankplot uses R1 aggregated json files to estimate whitelist barcodes using a histogram-thresholding strategy on UMI counts of all raw barcode.

   • Same process is reapeated for R2 json files.

   • It also makes plots of the histrograms and rankplots underlying raw barcodes.

   • A dup-rate histogram for all whitelist barcodes is also generated.

5. A second round of processing takes place on the quadruple csv files and they are converted into dictionary elements with key-value paring in 'R1_BC_whitelist':['R1_UMI','R2_BC_whitelist'] format.

6. Each of these part dictionaries is chunked into batches of R1_BC_whitelist barcodes, using function save_barcode_batch_json. Currently this is hardcoded to 30k barcodes in each batch.

7. Function aggregate_barcode_batches aggregates all parts for each of the batches and sotes them in new json files.

8. Function make_count_sparse_mtx_batch takes each batch and stores it in a sparse count matrix.

sbatch ~/reconstruction/scripts/SLURM_extract_10x.sh . xBO153_SB_240612_S1 \
        100000 \
        100000 \
        JJJJJJJJJJJJJJJJNNNNNNNNNNNNTTTTTTTTT \
        JJJJJJJJTCTTCAGCGTTCCCGAGAJJJJJJJNNNNNNNNNAAAAAAA


sbatch ~/reconstruction/scripts/SLURM_extract.sh . H8_3_recon \
        1000000 \
        1000000 \
        JJJJJJJJTCTTCAGCGTTCCCGAGAJJJJJJJNNNNNNNNNTTTTTTTTTT \
        JJJJJJJJTCTTCAGCGTTCCCGAGAJJJJJJJNNNNNNNNNAAAAAAAAAA