strkit 0.13.1__tar.gz → 0.15.0__tar.gz

{strkit-0.13.1/strkit.egg-info → strkit-0.15.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: strkit
-Version: 0.13.1
+Version: 0.15.0
 Summary: A toolkit for analyzing variation in short(ish) tandem repeats.
 Home-page: https://github.com/davidlougheed/strkit
 Author: David Lougheed
@@ -13,22 +13,22 @@ Requires-Python: ~=3.9
 Description-Content-Type: text/markdown
 License-File: LICENSE
 Requires-Dist: Flask<3.1,>=2.2.5
+Requires-Dist: orjson<4,>=3.9.15
 Requires-Dist: pysam<0.23,>=0.19
 Requires-Dist: numpy<1.27,>=1.23.4
 Requires-Dist: parasail<1.4,>=1.2.4
-Requires-Dist: pyfamsa<0.4,>=0.3.0
 Requires-Dist: scikit-learn<1.5,>=1.2.1
-Requires-Dist: scipy<1.13,>=1.10
+Requires-Dist: scipy<1.14,>=1.10
 Requires-Dist: statsmodels<0.15,>=0.14.0
-Provides-Extra: rustdeps
-Requires-Dist: orjson<4,>=3.8.5; extra == "rustdeps"
-Requires-Dist: strkit_rust_ext==0.5.0; extra == "rustdeps"
+Requires-Dist: strkit_rust_ext==0.12.2
 
-# STRkit
+# STRkit - short tandem repeat genotyping with long reads
 
 [![PyPI version](https://badge.fury.io/py/strkit.svg)](https://badge.fury.io/py/strkit)
 
-A genotyping and analysis toolkit for short(ish) tandem repeats.
+STRkit is a short tandem repeat (STR) genotyping and analysis toolkit for long read sequencing data, especially 
+PacBio HiFi data. The STRkit software package is written in Python and is available in the PyPI package registry or as
+a Docker container.
 
 <img src="./docs/images/strkit_logo_small.png" alt="" width="500" height="324" />
 
@@ -39,8 +39,8 @@ A genotyping and analysis toolkit for short(ish) tandem repeats.
   created during graduate research by DL.
 * 2023+: (versions beyond `0.8.0a1`):
   * Portions &copy; DL and McGill University 2021-2023
-  * Portions &copy; McGill University 2023
-  * Portions &copy; DL 2023
+  * Portions &copy; McGill University 2024
+  * Portions &copy; DL 2024
 
 
 ### Notice
@@ -75,6 +75,8 @@ Used under the terms of the ISC license.
 
 ## Installation
 
+### Via PyPI
+
 STRkit requires Python 3.9+ and can be installed from PyPI via `pip` 
 with the following command:
 
@@ -82,18 +84,38 @@ with the following command:
 python -m pip install strkit
 ```
 
-For **faster genotyping** (via some optimized utility functions) and the faster `orjson` library 
-for JSON parsing and serialization, specify the `rustdeps` extra:
+You may need to install the [Rust toolchain](https://www.rust-lang.org/tools/install)
+to compile the `strkit_rust_ext` wheels, although prebuilt wheels for this module are available
+for some platforms.
+
+On Digital Research Alliance of Canada/Compute Canada clusters, this involves loading a few modules:
+
+```bash
+module load rust/1.76.0 clang/17.0.6 python/3.11 scipy-stack/2023b parasail/2.6.2
+python -m pip install strkit
+```
+
+STRkit should then be available in your Python environment as a command-line tool:
 
 ```bash
-python -m pip install strkit[rustdeps]
+strkit --help
 ```
 
-On some Alliance (DRAC/CC) clusters, this may require loading Rust first, to compile the wheels:
+### As a Docker container
+
+STRkit is also available as a [Docker container](https://github.com/davidlougheed/strkit/pkgs/container/strkit), stored 
+in the GitHub Container Registry.
+
+It can be pulled using the following command:
 
 ```bash
-module load rust/1.65.0
-python -m pip install strkit[rustdeps]
+docker pull ghcr.io/davidlougheed/strkit:latest
+```
+
+Then, STRkit commands can be run mostly as normal using the Docker image:
+
+```bash
+docker run -it ghcr.io/davidlougheed/strkit --help
 ```
 
 
@@ -116,8 +138,8 @@ long read data should still work.
   * Whole-genome and targeted genotyping modes to adjust this re-weighting.
 * Incorporation of single-nucleotide variation (SNVs) for better and faster calling plus 
   additional downstream analysis possibilities.
-  * Recommended for **HiFi data only**. In my testing, this worsens runtime and call quality for 
-    ONT-UL data, but speeds up the tool and improves call quality for HiFi data. 
+  * Recommended for **HiFi data and ONT duplex data only**. In my testing, this worsens runtime and call quality for 
+    ONT ultra-long-read data, but speeds up the tool and improves call quality for HiFi/ONT duplex data. 
 * Parallelized for faster computing on clusters and for ad-hoc fast analysis of single samples.
 * 95% confidence intervals on calls via a user-configurable optional parametric bootstrapping process.
 
@@ -126,11 +148,11 @@ long read data should still work.
 
 ```bash
 strkit call \
-  path/to/read/file.bam \  # [REQUIRED] At least one indexed read file (BAM/CRAM)
-  --hq \  # If using PacBio HiFi reads, enable this to get better genotyping & more robust expansion detection
-  --realign \  # If using PacBio HiFi reads, enable this to enable local realignment / read recovery
+  path/to/read/file.bam \  # [REQUIRED] One indexed read file (BAM/CRAM)
+  --hq \  # If using PacBio HiFi or ONT duplex reads, enable this to get better genotyping & more robust expansion detection
+  --realign \  # If using PacBio HiFi reads, enable this to enable local realignment / read recovery. Good for detecting expansions, but slows down calling.
   --ref path/to/reference.fa.gz \  # [REQUIRED] Indexed FASTA-formatted reference genome
-  --loci path/to/loci.bed \  # [REQUIRED] TRF-formatted (or 4-col, with motif as last column) list of loci to genotype
+  --loci path/to/loci.bed \  # [REQUIRED] TRF-formatted (or 4-col, with motif as last column) sorted list of loci to genotype
   --incorporate-snvs path/to/dbsnp.vcf.gz \   # [RECOMMENDED FOR HIFI ONLY] If you want SNV calls to help phase STRs & speed up calling
   --min-reads 4 \  # Minimum number of supporting reads needed to make a call
   --min-allele-reads 2 \  # Minimum number of supporting reads needed to call a specific allele size 
@@ -149,9 +171,9 @@ If you're using HiFi reads as input, **use the `--hq` and `--realign` options**
 genotype calculation and a greater proportion of reads incorporated into the computed genotypes, 
 respectively. These should not add much performance overhead.
 
-If more than one read file is specified, the reads will be pooled. This can come in handy if you
-have e.g. multiple flow cells of the same sample split into different BAM files, or the reads are
-split by chromosome.
+If you want to **incorporate haplotagging from an alignment file (`HP` tags)** into the 
+process, which should speed up runtime and potentially improve calling results, you must pass 
+the `--use-hp` flag. **This flag is experimental, and has not been tested extensively.**
 
 If you want to **incorporate SNV calling** into the process, which speeds up runtime and gives
 marginally better calling results, you must provide an indexed, `bgzip`-compressed SNV catalog 
@@ -165,9 +187,11 @@ If you want to output a full call report, you can use the `--json output-file.js
 specify a path to output a more detailed JSON document to. This document contains 99% CIs, peak
 labels, and some other information that isn't included in the normal TSV file. If you want this
 file to be indented and human-readable, use the `--indent-json` flag in addition to `--json ...`.
+Note that **memory usage will be higher** when JSON output is enabled.
 
 If you want to output a VCF file (STRs and SNVs if called; currently not phased), use the
-`--vcf ...` argument.
+`--vcf ...` argument. If you pass `--vcf stdout`, the VCF will be written to `stdout` instead of a 
+file.
 
 See the '[Caller catalog format & choosing a catalog](./docs/caller_catalog.md)' page for more on
 how to format a locus catalog or choose from existing available catalogs.
@@ -180,19 +204,6 @@ bgzip my-reference.fa  # Replaces .fa with a .fa.gz file
 samtools faidx my-reference.fa.gz  # Generates a .fai index file
 ```
 
-##### Note on OpenMP Threading
-
-Slow performance can result from running `strkit call` on a system with OpenMP, 
-due to a misguided  attempt at multithreading under the hood somewhere in Numpy/Scipy (which doesn't work 
-here due to  repeated initializations of the Gaussian mixture model.) To fix this, the following
-environment variable is auto-set (hardcoded) before running:
-
-```bash
-export OMP_NUM_THREADS=1
-```
-
-If this hard-coded value interferes with your use case, please open an issue.
-
 
 #### Further documentation on the STRkit caller, including output format:
 

{strkit-0.13.1 → strkit-0.15.0}/README.md

@@ -1,8 +1,10 @@
-# STRkit
+# STRkit - short tandem repeat genotyping with long reads
 
 [![PyPI version](https://badge.fury.io/py/strkit.svg)](https://badge.fury.io/py/strkit)
 
-A genotyping and analysis toolkit for short(ish) tandem repeats.
+STRkit is a short tandem repeat (STR) genotyping and analysis toolkit for long read sequencing data, especially 
+PacBio HiFi data. The STRkit software package is written in Python and is available in the PyPI package registry or as
+a Docker container.
 
 <img src="./docs/images/strkit_logo_small.png" alt="" width="500" height="324" />
 
@@ -13,8 +15,8 @@ A genotyping and analysis toolkit for short(ish) tandem repeats.
   created during graduate research by DL.
 * 2023+: (versions beyond `0.8.0a1`):
   * Portions &copy; DL and McGill University 2021-2023
-  * Portions &copy; McGill University 2023
-  * Portions &copy; DL 2023
+  * Portions &copy; McGill University 2024
+  * Portions &copy; DL 2024
 
 
 ### Notice
@@ -49,6 +51,8 @@ Used under the terms of the ISC license.
 
 ## Installation
 
+### Via PyPI
+
 STRkit requires Python 3.9+ and can be installed from PyPI via `pip` 
 with the following command:
 
@@ -56,18 +60,38 @@ with the following command:
 python -m pip install strkit
 ```
 
-For **faster genotyping** (via some optimized utility functions) and the faster `orjson` library 
-for JSON parsing and serialization, specify the `rustdeps` extra:
+You may need to install the [Rust toolchain](https://www.rust-lang.org/tools/install)
+to compile the `strkit_rust_ext` wheels, although prebuilt wheels for this module are available
+for some platforms.
+
+On Digital Research Alliance of Canada/Compute Canada clusters, this involves loading a few modules:
+
+```bash
+module load rust/1.76.0 clang/17.0.6 python/3.11 scipy-stack/2023b parasail/2.6.2
+python -m pip install strkit
+```
+
+STRkit should then be available in your Python environment as a command-line tool:
 
 ```bash
-python -m pip install strkit[rustdeps]
+strkit --help
 ```
 
-On some Alliance (DRAC/CC) clusters, this may require loading Rust first, to compile the wheels:
+### As a Docker container
+
+STRkit is also available as a [Docker container](https://github.com/davidlougheed/strkit/pkgs/container/strkit), stored 
+in the GitHub Container Registry.
+
+It can be pulled using the following command:
 
 ```bash
-module load rust/1.65.0
-python -m pip install strkit[rustdeps]
+docker pull ghcr.io/davidlougheed/strkit:latest
+```
+
+Then, STRkit commands can be run mostly as normal using the Docker image:
+
+```bash
+docker run -it ghcr.io/davidlougheed/strkit --help
 ```
 
 
@@ -90,8 +114,8 @@ long read data should still work.
   * Whole-genome and targeted genotyping modes to adjust this re-weighting.
 * Incorporation of single-nucleotide variation (SNVs) for better and faster calling plus 
   additional downstream analysis possibilities.
-  * Recommended for **HiFi data only**. In my testing, this worsens runtime and call quality for 
-    ONT-UL data, but speeds up the tool and improves call quality for HiFi data. 
+  * Recommended for **HiFi data and ONT duplex data only**. In my testing, this worsens runtime and call quality for 
+    ONT ultra-long-read data, but speeds up the tool and improves call quality for HiFi/ONT duplex data. 
 * Parallelized for faster computing on clusters and for ad-hoc fast analysis of single samples.
 * 95% confidence intervals on calls via a user-configurable optional parametric bootstrapping process.
 
@@ -100,11 +124,11 @@ long read data should still work.
 
 ```bash
 strkit call \
-  path/to/read/file.bam \  # [REQUIRED] At least one indexed read file (BAM/CRAM)
-  --hq \  # If using PacBio HiFi reads, enable this to get better genotyping & more robust expansion detection
-  --realign \  # If using PacBio HiFi reads, enable this to enable local realignment / read recovery
+  path/to/read/file.bam \  # [REQUIRED] One indexed read file (BAM/CRAM)
+  --hq \  # If using PacBio HiFi or ONT duplex reads, enable this to get better genotyping & more robust expansion detection
+  --realign \  # If using PacBio HiFi reads, enable this to enable local realignment / read recovery. Good for detecting expansions, but slows down calling.
   --ref path/to/reference.fa.gz \  # [REQUIRED] Indexed FASTA-formatted reference genome
-  --loci path/to/loci.bed \  # [REQUIRED] TRF-formatted (or 4-col, with motif as last column) list of loci to genotype
+  --loci path/to/loci.bed \  # [REQUIRED] TRF-formatted (or 4-col, with motif as last column) sorted list of loci to genotype
   --incorporate-snvs path/to/dbsnp.vcf.gz \   # [RECOMMENDED FOR HIFI ONLY] If you want SNV calls to help phase STRs & speed up calling
   --min-reads 4 \  # Minimum number of supporting reads needed to make a call
   --min-allele-reads 2 \  # Minimum number of supporting reads needed to call a specific allele size 
@@ -123,9 +147,9 @@ If you're using HiFi reads as input, **use the `--hq` and `--realign` options**
 genotype calculation and a greater proportion of reads incorporated into the computed genotypes, 
 respectively. These should not add much performance overhead.
 
-If more than one read file is specified, the reads will be pooled. This can come in handy if you
-have e.g. multiple flow cells of the same sample split into different BAM files, or the reads are
-split by chromosome.
+If you want to **incorporate haplotagging from an alignment file (`HP` tags)** into the 
+process, which should speed up runtime and potentially improve calling results, you must pass 
+the `--use-hp` flag. **This flag is experimental, and has not been tested extensively.**
 
 If you want to **incorporate SNV calling** into the process, which speeds up runtime and gives
 marginally better calling results, you must provide an indexed, `bgzip`-compressed SNV catalog 
@@ -139,9 +163,11 @@ If you want to output a full call report, you can use the `--json output-file.js
 specify a path to output a more detailed JSON document to. This document contains 99% CIs, peak
 labels, and some other information that isn't included in the normal TSV file. If you want this
 file to be indented and human-readable, use the `--indent-json` flag in addition to `--json ...`.
+Note that **memory usage will be higher** when JSON output is enabled.
 
 If you want to output a VCF file (STRs and SNVs if called; currently not phased), use the
-`--vcf ...` argument.
+`--vcf ...` argument. If you pass `--vcf stdout`, the VCF will be written to `stdout` instead of a 
+file.
 
 See the '[Caller catalog format & choosing a catalog](./docs/caller_catalog.md)' page for more on
 how to format a locus catalog or choose from existing available catalogs.
@@ -154,19 +180,6 @@ bgzip my-reference.fa  # Replaces .fa with a .fa.gz file
 samtools faidx my-reference.fa.gz  # Generates a .fai index file
 ```
 
-##### Note on OpenMP Threading
-
-Slow performance can result from running `strkit call` on a system with OpenMP, 
-due to a misguided  attempt at multithreading under the hood somewhere in Numpy/Scipy (which doesn't work 
-here due to  repeated initializations of the Gaussian mixture model.) To fix this, the following
-environment variable is auto-set (hardcoded) before running:
-
-```bash
-export OMP_NUM_THREADS=1
-```
-
-If this hard-coded value interferes with your use case, please open an issue.
-
 
 #### Further documentation on the STRkit caller, including output format:
 

{strkit-0.13.1 → strkit-0.15.0}/setup.py

@@ -15,20 +15,15 @@ setup(
     python_requires="~=3.9",
     install_requires=[
         "Flask>=2.2.5,<3.1",
+        "orjson>=3.9.15,<4",
         "pysam>=0.19,<0.23",
         "numpy>=1.23.4,<1.27",
         "parasail>=1.2.4,<1.4",
-        "pyfamsa>=0.3.0,<0.4",
         "scikit-learn>=1.2.1,<1.5",
-        "scipy>=1.10,<1.13",
+        "scipy>=1.10,<1.14",
         "statsmodels>=0.14.0,<0.15",
+        "strkit_rust_ext==0.12.2",
     ],
-    extras_require={
-        "rustdeps": [
-            "orjson>=3.8.5,<4",
-            "strkit_rust_ext==0.5.0",
-        ],
-    },
 
     description="A toolkit for analyzing variation in short(ish) tandem repeats.",
     long_description=long_description,

strkit-0.15.0/strkit/VERSION

@@ -0,0 +1 @@
+0.15.0

{strkit-0.13.1 → strkit-0.15.0}/strkit/call/__init__.py

@@ -2,8 +2,10 @@ from __future__ import annotations
 
 from .allele import call_alleles
 from .call_sample import call_sample
+from .params import CallParams
 
 __all__ = [
     "call_alleles",
     "call_sample",
+    "CallParams",
 ]

{strkit-0.13.1 → strkit-0.15.0}/strkit/call/allele.py

@@ -1,8 +1,12 @@
 from __future__ import annotations
 
-# Disable OpenMP multithreading since it adds enormous overhead when multiprocessing
+# Disable OpenMP/other multithreading since it adds enormous overhead when multiprocessing
 import os
 os.environ["OMP_NUM_THREADS"] = "1"
+os.environ["OPENBLAS_NUM_THREADS"] = "1"
+os.environ["MKL_NUM_THREADS"] = "1"
+os.environ["VECLIB_MAXIMUM_THREADS"] = "1"
+os.environ["NUMEXPR_NUM_THREADS"] = "1"
 
 # ----------------------------------------------------------------------------------------------------------------------
 
@@ -20,6 +24,7 @@ from typing import Iterable, Literal, Optional, TypedDict, Union
 
 import strkit.constants as cc
 
+from .params import CallParams
 from .utils import get_new_seed
 
 __all__ = [
@@ -40,6 +45,7 @@ small_allele_min = 8
 expansion_ratio = 5
 N_GM_INIT = 3
 
+WEIGHT_1_0 = np.array([[1.0]])
 FLOAT_32_EPSILON = np.finfo(np.float32).eps
 
 CI_PERCENTILE_RANGES = {
@@ -52,15 +58,11 @@ def _array_as_int(n: Union[NDArray[np.int_], NDArray[np.float_]]) -> NDArray[np.
     return np.rint(n).astype(np.int32)
 
 
-def _calculate_cis(
-    samples,
-    force_int: bool = False,
-    ci: str = Literal["95", "99"],
-) -> Union[NDArray[np.int32], NDArray[np.float_]]:
+def _calculate_cis(samples, ci: str = Literal["95", "99"]) -> Union[NDArray[np.int32], NDArray[np.float_]]:
     percentiles = np.percentile(
         samples, CI_PERCENTILE_RANGES[ci], axis=1, method="interpolated_inverted_cdf"
     ).transpose()
-    return _array_as_int(percentiles) if force_int else percentiles
+    return _array_as_int(percentiles)
 
 
 def get_n_alleles(default_n_alleles: int, sample_sex_chroms: Optional[str], contig: str) -> Optional[int]:
@@ -100,7 +102,7 @@ def fit_gmm(
             # I've confirmed this gives an ~identical result to fitting a GMM with one parameter.
             fake_g: object = type("", (), {})()
             fake_g.means_ = np.array([[np.mean(sample_rs)]])
-            fake_g.weights_ = np.array([[1.0]])
+            fake_g.weights_ = WEIGHT_1_0
             fake_g.covariances_ = np.array([[np.var(sample_rs)]])
             return fake_g
 
@@ -143,7 +145,7 @@ def fit_gmm(
     return g
 
 
-class CallDict(TypedDict):
+class BaseCallDict(TypedDict):
     call: Union[NDArray[np.int32], NDArray[np.float_]]
     call_95_cis: Union[NDArray[np.int32], NDArray[np.float_]]  # 2D arrays
     call_99_cis: Union[NDArray[np.int32], NDArray[np.float_]]  # 2D arrays
@@ -153,6 +155,10 @@ class CallDict(TypedDict):
     modal_n_peaks: int
 
 
+class CallDict(BaseCallDict, total=False):
+    ps: int
+
+
 def make_read_weights(read_weights: Optional[Iterable[float]], num_reads: int) -> NDArray[np.float_]:
     return np.array(
         read_weights if read_weights is not None else np.array(([1/num_reads] * num_reads) if num_reads else []))
@@ -163,15 +169,12 @@ def call_alleles(
     repeats_rev: RepeatCounts,
     read_weights_fwd: Optional[Iterable[float]],
     read_weights_rev: Optional[Iterable[float]],
-    bootstrap_iterations: int,
+    params: CallParams,
     min_reads: int,
-    min_allele_reads: int,
     n_alleles: int,
     separate_strands: bool,
     read_bias_corr_min: int,
     gm_filter_factor: int,
-    hq: bool,
-    force_int: bool,
     seed: Optional[int],
     logger_: logging.Logger,
     debug_str: str,
@@ -200,18 +203,18 @@ def call_alleles(
         logger_.debug(f"{debug_str} - skipping bootstrap / GMM fitting for allele(s) (single value)")
         cn = combined_reads[0]
 
-        call = np.array([cn] * n_alleles)
-        call_cis = np.array([[cn, cn] for _ in range(n_alleles)])
+        call = _array_as_int(np.full(n_alleles, cn))
+        call_cis = _array_as_int(np.full((n_alleles, 2), cn))
 
-        peaks: NDArray[np.float_] = np.array([cn] * n_alleles, dtype=np.float_)
+        peaks: NDArray[np.float_] = call.astype(np.float_)
 
         return {
-            "call": _array_as_int(call) if force_int else call,
-            "call_95_cis": _array_as_int(call_cis) if force_int else call,
-            "call_99_cis": _array_as_int(call_cis) if force_int else call,
+            "call": call,
+            "call_95_cis": call_cis,
+            "call_99_cis": call_cis,
             "peaks": peaks,
-            "peak_weights": np.array([1.0] * n_alleles) / n_alleles,
-            "peak_stdevs": np.array([0.0] * n_alleles),
+            "peak_weights": np.full(n_alleles, 1.0 / n_alleles),
+            "peak_stdevs": np.full(n_alleles, 0.0),
             "modal_n_peaks": 1,  # 1 peak, since we have 1 value
         }
 
@@ -234,14 +237,14 @@ def call_alleles(
 
         fwd_strand_samples = rng.choice(
             fwd_strand_reads,
-            size=(bootstrap_iterations, target_length),
+            size=(params.num_bootstrap, target_length),
             replace=True,
             p=fwd_strand_weights,
         )
 
         rev_strand_samples = rng.choice(
             rev_strand_reads,
-            size=(bootstrap_iterations, target_length),
+            size=(params.num_bootstrap, target_length),
             replace=True,
             p=rev_strand_weights,
         )
@@ -254,26 +257,26 @@ def call_alleles(
         concat_samples = np.sort(
             rng.choice(
                 combined_reads,
-                size=(bootstrap_iterations, combined_len),
+                size=(params.num_bootstrap, combined_len),
                 replace=True,
                 p=combined_weights,
-            ) if bootstrap_iterations > 1 else np.array([combined_reads]),
+            ) if params.num_bootstrap > 1 else np.array([combined_reads]),
             kind="stable")
 
     gmm_cache = {}
 
-    def _get_fitted_gmm(s: NDArray[np.int_] | NDArray[np.float_]) -> Optional[object]:
-        if (s_t := tuple(s)) not in gmm_cache:
+    def _get_fitted_gmm(s: Union[NDArray[np.int_], NDArray[np.float_]]) -> Optional[object]:
+        if (s_t := s.tobytes()) not in gmm_cache:
             # Fit Gaussian mixture model to the resampled data
-            gmm_cache[s_t] = fit_gmm(rng, s, n_alleles, allele_filter, hq, gm_filter_factor)
+            gmm_cache[s_t] = fit_gmm(rng, s, n_alleles, allele_filter, params.hq, gm_filter_factor)
 
         return gmm_cache[s_t]
 
     # Filter out peaks that aren't supported by ~min_allele_reads reads by probability, with some delta to
     # allow for peaks supported by "most of a read".
-    allele_filter = (min_allele_reads - 0.1) / concat_samples.shape[0]
+    allele_filter = (params.min_allele_reads - 0.1) / concat_samples.shape[0]
 
-    for i in range(bootstrap_iterations):
+    for i in range(params.num_bootstrap):
         sample = concat_samples[i, :]
 
         g: Optional[object] = _get_fitted_gmm(sample)
@@ -320,8 +323,8 @@ def call_alleles(
     # Calculate 95% and 99% confidence intervals for each allele from the bootstrap distributions.
     allele_samples_argsort = allele_samples.argsort(axis=1, kind="stable")
     allele_samples = np.take_along_axis(allele_samples, allele_samples_argsort, axis=1)
-    allele_cis_95 = _calculate_cis(allele_samples, force_int=force_int, ci="95")
-    allele_cis_99 = _calculate_cis(allele_samples, force_int=force_int, ci="99")
+    allele_cis_95 = _calculate_cis(allele_samples, ci="95")
+    allele_cis_99 = _calculate_cis(allele_samples, ci="99")
     allele_weight_samples = np.take_along_axis(allele_weight_samples, allele_samples_argsort, axis=1)
     allele_stdev_samples = np.take_along_axis(allele_stdev_samples, allele_samples_argsort, axis=1)
 
@@ -335,20 +338,20 @@ def call_alleles(
 
     median_idx = allele_samples.shape[1] // 2  #
     medians_of_means = allele_samples[:, median_idx]
-    medians_of_means_final = medians_of_means
-    if force_int:
-        medians_of_means_final = np.rint(medians_of_means).astype(np.int32)
-    peak_weights = allele_weight_samples[:, median_idx]
+    medians_of_means_final = np.rint(medians_of_means).astype(np.int32)
+    peak_weights = allele_weight_samples[:, median_idx].flatten()
     peak_stdevs = allele_stdev_samples[:, median_idx]
     modal_n_peaks: int = statistics.mode(sample_peaks).item()
 
+    peak_weights /= peak_weights.sum()  # re-normalize weights
+
     return {
         "call": medians_of_means_final.flatten(),
         "call_95_cis": allele_cis_95,
         "call_99_cis": allele_cis_99,
 
-        "peaks": medians_of_means_final.flatten(),  # Don't round, so we can recover original Gaussian model
-        "peak_weights": peak_weights.flatten(),
+        "peaks": medians_of_means.flatten(),  # Don't round, so we can recover original Gaussian model
+        "peak_weights": peak_weights,
         "peak_stdevs": peak_stdevs.flatten(),
         # TODO: should be ok to use this, because resample gets put at end, vertically (3rd allele in a 3-ploid case)
         #  so taking the first 2 alleles still works in terms of stdev/mean estimates? I think?