MILO

MILO can be used for predicting Microsatellite Instability in LOw-quality samples. MILO takes indel profiles (83-channel mutation spectrum, see here for more information) of given samples and predicts their MSI status.

Low-quality samples could be shallow WGS (e.g., ~0.1X) and/or low-purity FF/FFPE samples (>2% for FF and >5% for FFPE). No matched normal is required.

1 Installation

MILO script can be downloaded here. You also need to download the two trained classifiers (FFPE_rf and FF_rf) to the same folder.

2 Dependencies

1. Python - MILO is tested on python 3.11.0.
2. Python packages used by MILO

• joblib (1.2.0)
• numpy (1.23.5)
• pandas (2.0.0)
• seaborn (0.13.2)
• sys, getopt, datetime, sklearn

3 Usage

3.1 Example commands

1. On FF samples without noise correction:

python MILO_setup.py [-I|--Input] ./MILO_test_data.csv [-T|--TissueType] FF

2. On FF samples with noise correction with plot using default noise profile:

python MILO_setup.py [-I|--Input] ./MILO_test_data.csv [-T|--TissueType] FF [-C|--NoiseCorrection] True [-P|--Plot] True

3. On FF samples with noise correction with user-specified noise profile:

python MILO_setup.py [-I|--Input] ./MILO_test_data.csv [-T|--TissueType] FF [-C|--NoiseCorrection] True [-N|--Noise_file] ./test_noise.csv

4. On FFPE samples in above examples, using [-T|--TissueType] FFPE.

5. To print out help informaiton:

python MILO_setup.py [-h|--help]

3.2 Our tests

We tested MILO on our FF sWGS samples. Click here for the test data and results. Please format your input data as descirbed in the following section.

3.3 Arguments

1. Required

• [-I|--Input] <path_of_input_file> : the location of your CSV file containing the 83-channel mutational counts for your sample(s).

  • CSV Format: The file should be in CSV format with comma separators (,) between values.
  • Column Structure:
    • The file should contain 84 columns.
    • The first column must be the sample ID.
    • The following 83 columns should represent the 83-channel mutation profile.
  • Row Structure: Each row in the file represents a unique sample.

• [-T|TissueType] <FFPE|FF>: MILO trained separate classifiers for FFPE and fresh frozen (FF) samples. Therefore, it is important that you specify your tissue type.

2. Optional

• [-C|--NoiseCorrection] <False|True>: determine if the noise correction will be carried out by MILO
• [-N|--Noise_file] <path_to_noise_file>: users can specify a noise profile using this argument. If none is provided, MILO defaults to using cohort-specific noise profiles from high-confidence MMRp samples (Prob(MSI) < 0.1). If no such profiles are available, MILO uses the FF/FFPE noise pattern from our sWGS cohorts as described in our paper.
• [-P|--Plot] <False|True>: determine if the PDFs of indel profiles will be generated
• [-PC|--Prob_cutoff] : specify the threshold for Prob(MSI) to determine high-confidence MSI samples. The default value is 0.75, meaning only samples with Prob(MSI) > 0.75 are considered high-confidence MSI predictions.

3.4 Output files

Here are the output files from MILO's prediction:

• CSV file with indel profiles and the MILO predictions
• (Optional) CSV file of MILO predicted long-deletion intensity in MSI positive files
• (Optional) PDFs of indel profiles in predicted MMRd samples (and/or noise-corrected MMRd profiles) can be found in the folder ./plots/.

4 Citation

Guo, Q. et al. Long deletion signatures in repetitive genomic regions track somatic evolution and enable sensitive detection of microsatellite instability. bioRxiv 2024.10.03.616572 (2024) doi:10.1101/2024.10.03.616572.

Check out out Preprint!