Progressive Genome Segment Enhancement (PGSE)

Overview

PGSE is an algorithm for predicting phenotypes from whole genome sequencing (WGS) data. It was intiially developed for the prediction of antimicrobial minimum inhibitory concentration (MIC) in bacterial strains. PGSE has higher accuracy, lower memory consumption, and shorter runtime compared to traditional $k$-mer based XGBoost models. PGSE is also able to run on distributed systems.

Contributors

Dr Yinzheng (William) Zhong, Univerisity of Liverpool (algorithm design & implementation)

Dr Alessandro Gerada, University of Liverpool (conceptualisation, R package, funding)

Prof William Hope, University of Liverpool (conceptualisation, funding, supervision)

Citation

@article{gerada2026prediction,
  title={Prediction of antimicrobial minimum inhibitory concentration from bacterial genomes using a scalable and interpretable machine learning approach},
  author={Gerada, Alessandro and Zhong, Yinzheng and Harper, Nicholas and Velluva, Anoop and Reza, Nada and Dubey, Vineet and Howard, Alex and Green, Peter L and Paterson, Steve and Hope, William},
  journal={npj Antimicrobials and Resistance},
  year={2026},
  publisher={Nature Publishing Group}
}

License

This project is licensed under the GNU Affero General Public License v3.0 (AGPL-3.0). See the LICENSE file for details.

Installation

PyPI

Make sure Python 3.10 or later is installed, then install pgse from PyPI with uv:

uv pip install pgse

or with pip:

pip install pgse

The published wheels bundle the compiled native counting kernel (a Rust extension) for Linux, macOS, and Windows. Segment counting runs through this kernel; if it is ever unavailable (e.g. a source install without a Rust toolchain) PGSE falls back to a slower pure-Python counter.

Conda

To use in a conda environment:

conda create -n pgse python=3.11
conda activate pgse
python -m pip install pgse

pgse is now available to import.

R

To use PGSE through R, install the package in an R session using:

install.packages("devtools")
devtools::install_github("yinzheng-zhong/PGSE", subdir = "R-package")

Usage

PGSE can be used either as a library, from your own Python program, or as a standalone command-line program. Both do the same work; the sections below give each in turn, for training and then for prediction.

Training

As a library

TrainingPipeline.train() runs the whole pipeline in memory and hands back everything it produced. It writes nothing: no models, no segment lists, no results, no log files and no progress files, whatever paths were passed to the constructor. Use it when PGSE is one step inside a larger program.

from pgse import TrainingPipeline

result = TrainingPipeline(
    data_dir='genomes/',
    label_file='labels.csv',
    folds=5,
    metric='r2',
).train()

# 1. A predictive model, ready to use
model = result.model                      # the best-scoring fold; result.models has them all
predictions = model.predict(['new_1.fna', 'new_2.fna'])
predictions = model.predict_sequences(['ATGCATTACA...'])   # or straight from memory
features = model.count(files=['new_1.fna'])                # the raw segment-count matrix

# 2. The discovered segments and how important each one was
result.segments.top(10).to_frame()        # Segment / Importance, most important first
result.segments.segments                  # the segments alone, in count-matrix order
result.segments.importances               # the matching scores as a NumPy array

# 3. Per-fold detail
result.scores                             # the metric for each fold
result.predictions                        # held-out Prediction/Actual/Fold for every fold
result.to_frame()                         # one row per fold: score and segment count
for fold in result.folds:
    print(fold.index, fold.score, len(fold.segments))

A model carries its own segments, alphabet and count settings, so nothing has to be passed alongside it and several models can be held at once. Predicting uses neither Ray nor any global state. Saving is explicit, and everything needed to reload is written:

from pgse import PGSEModel

model.save('artifacts/ecoli-caz')     # ecoli-caz.json, ecoli-caz_segs.csv, ecoli-caz_meta.json
reloaded = PGSEModel.load('artifacts/ecoli-caz')
reloaded.predict(['new_1.fna'])       # the alphabet and count settings come back with it

run() is the same run with its output written out: use it when you want the artefacts on disk but are still driving PGSE from Python. Each fold's model, segments and metadata go under export_file, and a resume point to save_file; whichever of the two is left unset is simply not written.

result = TrainingPipeline(
    data_dir='genomes/',
    label_file='labels.csv',
    save_file='run.save',         # resume point, removed once a fold finishes
    export_file='out/ecoli-caz',  # out/ecoli-caz_fold_0.json, _segs.csv, _meta.json, ...
    folds=5,
).run()

Those per-fold files are exactly what save() writes, so PGSEModel.load('out/ecoli-caz_fold_0') picks up anything a run() or a pgse-train produced.

As a standalone program

To run PGSE as a standalone program on a local machine, install the package and use the following command as an example:

pgse-train \
        --label-file "../<path_to>/<you_labels>.csv" \
        --data-dir "../<you_data_dir>/" \
        --pre-kfold-info-file "../<k_fold_information>.json" \
        --save-file "../<saved progress>.save" \
        --export-file "../<exported files>" \
        --workers 8 \
        --features 10000 \
        --partition-size-target 5000 \
        --dist 0 \
        --k 6 \
        --target 70 \
        --ext 2 \
        --lr 0.001 \
        --num-rounds 6000 \
        --folds 5 \
        --ea-max 64 \
        --ea-min 0

• --label-file (Required): path to the .csv label file

  Here the label file is a csv file with the following format:

  | labels | files     |
  | ------ | --------- |
  | 7      | file1.fna |
  | 7      | file2.fna |
  | 6      | file3.fna |
  

  The labels are the target values for the prediction task. The files are the file names (.fna files under --data-dir) containing the genome sequences.

• --data-dir (Required): path to the data directory containing the .fna files. PGSE will be able to retrieve the genome sequences using this path and the file names in the label file.

• --pre-kfold-info-file: path to the predefined k-fold info JSON file. This is not required but will be useful if you want to compare PGSE with other systems. Without this, PGSE will split the data into k folds randomly using a fixed seed. E.g.

  {
      "fold_0": [
          "Sample_208-MOLMIC_E33.scaffolds.fna",
          "Sample_726-MOLMIC_F29.scaffolds.fna",
          "Sample_474-MOLMIC_I14.scaffolds.fna",
          "Sample_111-MOLMIC_C61.scaffolds.fna",
          "Sample_087-MOLMIC_C25.scaffolds.fna",
          "Sample_467-MOLMIC_I6.scaffolds.fna",
          "..."
      ],
      "fold_1": [
          "Sample_208-MOLMIC_E33.scaffolds.fna",
          "Sample_726-MOLMIC_F29.scaffolds.fna",
          "Sample_474-MOLMIC_I14.scaffolds.fna",
          "Sample_111-MOLMIC_C61.scaffolds.fna",
          "Sample_087-MOLMIC_C25.scaffolds.fna",
          "Sample_467-MOLMIC_I6.scaffolds.fna",
          "..."
      ],
      "...": [
      "..."
      ]
  }
  

• --save-file: file to save the progress. This is useful if you want to resume the training process. Nothing is written when it is omitted.

• --export-file: path prefix for the results. Normally without an extension. Each fold writes <prefix>_fold_<i>.json (the trained model), <prefix>_fold_<i>_segs.csv (the selected segments and their importance) and <prefix>_fold_<i>_meta.json (the alphabet and count settings needed to reload it), plus <prefix>.csv with the predictions of every fold. Nothing is written when it is omitted.

• --log-file: file to append the log to. The log goes to the console only when it is omitted.

• --workers: number of workers per node.

• --features: Maximum number of features to keep after the feature importance calculation and ranking.

• --partition-size-target: Target number of features in each XGBoost partition during feature selection (default 5000). Partitions are evenly sized, so the actual size lands between this and twice this. See Why do we perform feature partitioning? below. Larger partitions use more memory per worker and give each model a wider view of the features; 0 or less trains a single model over all features.

• --dist: Using distributed computation or not. 0 for running on a single node/machine, 1 for running on multiple nodes.

• --k: initial k-mer size.

• --target: Maximum segment length to extend to.

• --ext: Extension length in each round. Extension parameter p from the paper.

• --lr: learning rate.

• --num-rounds: Maximum rounds for the training process.

• --folds: Number of folds for the k-fold cross-validation.

• --binary: 0 (default) predicts a continuous value; 1 predicts a 0/1 label. See Binary mode below.

• --metric: Validation metric reported for the held-out fold (default essential_agreement, or auroc with --binary 1). See Validation metrics below.

• --ea-max: Maximum number of censored essential agreement values. Don't need this unless you want to see more accurate EA information from the console output during the training. Only read by the essential_agreement metric.

• --ea-min: Minimum number of censored essential agreement values. Similar to --ea-max.

• --alphabet: the set of characters the input is made of, given as a single string. Defaults to DNA (atgc). See Alphabets below.

• --case-sensitive: 0 (default) folds everything to lower case, 1 treats upper and lower case as distinct characters.

• --complement: the complement used to canonicalise segments, one character per character of --alphabet. See Alphabets below.

• --uint16: 0 (default) stores the segment-count matrix as float32; 1 stores it as uint16, halving the memory. Lossless for counts up to 65535 (larger counts are saturated). See Reducing memory usage below.

• --sparse: 0 (default) stores the count matrix densely; 1 stores it as a sparse CSR matrix. For short, sparse inputs (e.g. SMILES strings) the matrix is almost entirely zeros, so this can save orders of magnitude. XGBoost reads the unstored zeros of a CSR matrix as missing values (not as 0), so the same value must be used at prediction time. See Reducing memory usage below.

Every flag above matches a TrainingPipeline argument of the same name, so anything the command line can do the library can do too. main-pgse.py shows the two joined up: it builds the pipeline from pgse.environment.args.get_parser(), the parser pgse-train itself uses, which you can reuse or replace with your own.

Alphabets

PGSE is not limited to DNA. The alphabet is the set of characters the input is made of, and everything outside it is dropped while reading the input files. Any symbolic text can therefore be used, for example plain text:

pgse-train \
        --label-file "../<path_to>/<you_labels>.csv" \
        --data-dir "../<you_data_dir>/" \
        --alphabet "abcdefghijklmnopqrstuvwxyz " \
        --case-sensitive 0 \
        --k 3 \
        --target 12

The same options are available on the Python API:

from pgse import TrainingPipeline

pipeline = TrainingPipeline(
    data_dir='...',
    label_file='...',
    k=3,
    target=12,
    alphabet='abcdefghijklmnopqrstuvwxyz ',
    case_sensitive=False
)

Three things follow from the alphabet:

• Case sensitivity. By default the alphabet is case-insensitive: the input is folded to lower case, so A and a are the same character. With --case-sensitive 1 nothing is folded and the alphabet has to list every character you want to keep, e.g. "abcdefghijklmnopqrstuvwxyzABCDEFGHIJKLMNOPQRSTUVWXYZ ".
• Canonicalisation. DNA segments are counted on both strands, so a segment and its reverse complement are the same feature. This only makes sense when the alphabet has a complement. --complement defaults to reverse complementing for the DNA alphabet and to no canonicalisation for any other alphabet, which is almost always what you want. To set one explicitly, give one character per character of --alphabet (--alphabet augc --complement uacg for RNA), or pass an empty string to switch it off.
• k-mer size. The number of initial k-mers is len(alphabet) ** k, so a larger alphabet needs a smaller --k. The 27-character alphabet above yields about 20k 3-mers, roughly the same as DNA with --k 7.

Inference has to use the same alphabet as training, since the exported segments are counted against it. Pass the same --alphabet, --case-sensitive and --complement values to pgse-predict; if the segments do not fit the alphabet, PGSE fails with an error rather than predicting on all-zero counts.

Binary mode

By default PGSE predicts a continuous value. --binary 1 (or binary=True on the Python API) predicts a 0/1 label instead: XGBoost is trained with the binary:logistic objective, and a prediction is the probability that the sample is a 1.

pgse-train \
        --label-file "../<path_to>/<your_labels>.csv" \
        --data-dir "../<your_data_dir>/" \
        --binary 1

from pgse import TrainingPipeline

result = TrainingPipeline(
    data_dir='genomes/',
    label_file='labels.csv',
    folds=5,
    binary=True,
).train()

result.score                                # mean AUROC across the folds
probabilities = result.model.predict(['new_1.fna', 'new_2.fna'])   # e.g. [0.02, 0.91]
labels = (probabilities >= 0.5).astype(int)                        # if you want a decision

Every label has to be 0 or 1; anything else fails with an error naming the offending values rather than training a meaningless model. Booleans count as 0 and 1, so a label_file dictionary of True/False works as well as one of 0/1.

The label file is otherwise unchanged:

| labels | files     |
| ------ | --------- |
| 1      | file1.fna |
| 0      | file2.fna |
| 1      | file3.fna |

Binary mode changes the objective, the default metric (auroc instead of essential_agreement) and the meaning of a prediction. Everything else — segment extension, feature selection, folds, alphabets, the saved model — behaves exactly as it does for a continuous target. A saved model records that it is binary, so PGSEModel.load knows its predictions are probabilities:

from pgse import PGSEModel

model = PGSEModel.load('artifacts/resistance')
model.binary       # True
model.predict(['new_1.fna'])   # a probability, not a class

Validation metrics

The score reported for the held-out fold is chosen with --metric (or metric= on the Python API). Essential agreement is the default and keeps the behaviour of earlier versions, but it only makes sense for log2 MIC labels; anything else should pick a metric that matches its label. A binary run defaults to auroc instead.

--metric	What it measures	Best for
essential_agreement	Fraction of predictions within one two-fold dilution of the label	log2 MIC labels (default)
rmse	Root mean squared error, in the units of the label	Continuous labels, penalising large misses
mae	Mean absolute error	Continuous labels with outliers
mape	Mean absolute percentage error over the non-zero labels	Continuous labels compared in relative terms
r2	Fraction of the label's variance explained	Continuous labels, e.g. growth rate
pearson	Linear correlation between label and prediction	Continuous labels, e.g. growth rate
spearman	Rank correlation	Continuous labels where only the ordering matters
accuracy	Fraction of predictions that round to the exact label	Integer labels, e.g. 0/1
mcc	Matthews correlation coefficient, in [-1, 1]	Imbalanced integer labels
auroc	Area under the ROC curve: the chance a positive outranks a negative	Binary labels (default with --binary 1)
auprc	Area under the precision-recall curve, which ignores the true negatives	Binary labels where the positives are rare
log_loss	Mean negative log likelihood, scoring the probabilities themselves	Binary labels where the probability matters, not just the ranking
f1	Harmonic mean of precision and recall	Binary labels, one headline number for the decision
precision	Fraction of the predicted positives that are positive	Binary labels where a false positive is costly
recall	Fraction of the positives that are predicted positive (sensitivity)	Binary labels where a missed positive is costly
specificity	Fraction of the negatives that are predicted negative	Binary labels, reported alongside recall
balanced_accuracy	Mean of recall and specificity	Imbalanced binary labels

The bottom eight are for binary mode. auroc, auprc and log_loss read the predicted probability directly; the rest need a yes/no answer and split the probability at 0.5, as accuracy and mcc already do. log_loss is the only one of them where a smaller score is better, which TrainingResult.best_fold accounts for.

For a continuous target such as growth rate, r2 is the usual headline number, with rmse for the error in the target's own units and spearman when only the ranking of the samples matters:

pgse-train \
        --label-file "../<path_to>/<your_labels>.csv" \
        --data-dir "../<your_data_dir>/" \
        --metric r2

from pgse import TrainingPipeline

pipeline = TrainingPipeline(data_dir='...', label_file='...', metric='r2')

PGSE always logs RMSE; --metric selects the extra score reported alongside it once the selected segments are trained.

The metric does not change what PGSE optimises. The objective is set by --binary alone: reg:squarederror by default and binary:logistic with --binary 1. So --metric auroc on a continuous run scores the regressor's output as if it were a ranking, which is legal but rarely what you want — pass --binary 1 to actually train a classifier. accuracy and mcc work either way, on whole-number labels: both round the predictions to the nearest integer before scoring.

The metrics are the static methods of the Metric class in pgse/validation/metrics.py, with the array handling they share in pgse/validation/utils.py. Adding a metric is adding a method — its name becomes the --metric value and the first line of its docstring becomes its entry in --metric's help:

class Metric:
    ...

    @staticmethod
    def max_error(y_true, y_pred):
        """Largest absolute error over the samples.

        Args:
            y_true: True labels.
            y_pred: Predicted values.
        """
        return float(np.max(np.abs(as_array(y_true) - as_array(y_pred))))

A metric that needs extra parameters just declares them as keyword arguments. The pipeline passes all of its validation options to every metric and each receives only the ones its signature names, which is how essential_agreement gets ea_min and ea_max while the others ignore them.

Reducing memory usage

The segment-count matrix (one row per sample, one column per segment) is the largest object PGSE holds in memory. Two optional flags shrink it; they are independent and can be combined. Both default to off, so existing runs are unaffected.

• --uint16 1 stores counts as 16-bit integers instead of 32-bit floats, halving the matrix. Counts are non-negative integers, so this is lossless up to 65535; any larger count is saturated to 65535.
• --sparse 1 stores the matrix as a sparse CSR matrix. This is the big lever for short inputs where most segments are absent from most samples — for example SMILES strings, where each row has only tens of non-zero counts out of thousands or millions of columns, making the matrix >99% zeros. For long, dense inputs such as bacterial genomes the matrix is not sparse, so leave this off.

pgse-train \
        --label-file "../<path_to>/<your_labels>.csv" \
        --data-dir "../<your_data_dir>/" \
        --alphabet "..." \
        --sparse 1 \
        --uint16 1

The same options are available on the Python API as sparse=True and uint16=True:

from pgse import TrainingPipeline

pipeline = TrainingPipeline(data_dir='...', label_file='...', sparse=True, uint16=True)

Important: --sparse must match between training and prediction. In a sparse matrix, unstored zeros are read by XGBoost as missing values, whereas in a dense matrix a count of zero is an explicit 0. Training dense and predicting sparse (or vice versa) shifts the predictions. --uint16 has no such constraint. Pass the same values to pgse-predict (see below).

Distributed computation

To run PGSE on a distributed system, you need to use your environment specific setup. There are multiple examples about running PGSE using Slurm under the slurm-scripts directory.

• job-pgse-array.sh: Run PGSE on a cluster using Slurm with multiple nodes for multiple antibiotics using array jobs. Here -dist is set to 0 as each task is running separately.
• job-pgse-dist.sh: Run PGSE on a cluster using Slurm with multiple nodes for a single antibiotic. Here -dist is set to 1 as the task is running on different nodes.
• job-pgse-single.sh: Run PGSE on a Slurm cluster with a single node for a single antibiotic. Here -dist is set to 0.

Prediction

As a library

PGSEModel is the whole interface: load one, predict with it, keep it around for as many calls as you like.

from pgse import PGSEModel

model = PGSEModel.load('out/ecoli-caz_fold_0')
model.predict(['sample_1.fna', 'sample_2.fna'])

The alphabet and the count settings are stored with the model and restored on load, so there is nothing to keep in step by hand. See As a library above for the rest of what a model can do.

InferencePipeline is the older interface, kept for existing callers: it takes the model and segment files separately, and the alphabet and sparse setting have to be passed again by hand to match the ones used for training. main-pgse-inf.py is a worked example.

from pgse import InferencePipeline

pipeline = InferencePipeline(
    '../volatile/var/result-k6-CAZ-perf_fold_0.json',
    '../volatile/var/result-k6-CAZ-perf_fold_0_segs.csv',
    workers=8
)
print(pipeline.run(['../volatile/cgr/Sample_002-MOLMIC_B2.scaffolds.fna']))

As a standalone program

To run prediction as a standalone program, install the package and use the following command as an example:

pgse-predict \
        --model-file "../<path_to_model>.json" \
        --segments-file "../<path_to_segments>.csv" \
        --data-dir "../<you_data_dir>/" \
        --workers 8

If the model was trained on a non-DNA alphabet, pass the same --alphabet, --case-sensitive and --complement values that were used for training. See Alphabets. If training used --sparse 1, prediction must use it too — see Reducing memory usage.

Logging

Logging goes to the console only. Set PGSE_LOG_FILE, pass --log-file on the command line, or call pgse.log.logger.add_file_handler(path), to also append it to a file.

R package

To use PGSE through the R package, consult the package documentation.

For Development

PGSE uses uv for packaging and dependency management. All project metadata and dependencies live in pyproject.toml.

Prerequisite: a Rust toolchain. Building from source compiles the native counting kernel (the Rust crate under native/, built into the package as pgse._native), so you need cargo/rustc on your PATH. Install them from rustup.rs:

curl --proto '=https' --tlsv1.2 -sSf https://sh.rustup.rs | sh
# rustup only updates PATH for *new* shells, so load it into the current one
# (or open a new shell) before installing:
. "$HOME/.cargo/env"

The build fails loudly if the toolchain is missing rather than silently shipping an install without the extension, so every install has the fast path. (A pure-Python counter still exists as a runtime safety net, and logs a warning if it is ever used.)

Clone the repository and create a synced environment. This installs the runtime and dev dependencies and performs an editable install of pgse, compiling the Rust kernel as part of it:

uv sync

To run an editable install on its own:

uv pip install -e .

To build the distribution artifacts (the wheel bundles the compiled kernel):

uv build

Run the tests with:

uv run pytest

Releases to PyPI are automated by GitHub Actions: when the version in pyproject.toml changes on main, the workflow uses cibuildwheel to build one abi3 wheel per platform (Linux, macOS, Windows; each valid for CPython 3.10+), builds an sdist, and publishes them.

Acknowledgements

This work was funded, in part, by UKRI and the Wellcome trust.

This work was undertaken on Barkla, part of the High Performance Computing facilities at the Univeristy of Liverpool, UK.

Common Issues

XGBoost training is only using one core.

Some linux distributions need an environment variable OMP_NUM_THREADS=<num threads> to be set to allow XGBoost to use multiple cores.

Q & A

Why do we perform feature partitioning?

There are four reasons why feature partitioning is crucial in PGSE. First, feature partitioning is used as a memory reduction technique. The model is trained on a subset of the features at a time, therefore, the memory consumption is reduced while maintained a relatively stable RAM usage regardless of the number of total features. Second, feature partitioning helps to parallelise the training process. Each partition can be trained on a different worker across different nodes. This is particularly useful as XGBoost training consumes most of the time in the training process. Third, from the experiments we have conducted, we found that feature dimensionality affects the model's optimal hyperparameters. For example, higher feature dimensionality requires a shallower tree depth in general. PGSE is a dynamic system that and the total number of features can be different in each round. Therefore, partitioning the features into similarly-sized sub-features can help to minimise the impact of the feature dimensionality on the model's hyperparameters. Finally, feature partitioning helps to preserve the feature importance information from XGBoost. Likely due to the pruning process, more feature importance information will be lost (become 0) if the dimensionality increases.

The size of each partition is set with --partition-size-target (or partition_size_target= on TrainingPipeline), which defaults to 5000 features.

Why do we eliminate features?

If segment A is extended into segment B, A becomes a subsequences of B. For pairs like A and B, we only need to keep the ones with higher feature importance. Extension and elimination are two crucial parts of the PGSE system, which grows the genome segments longer and the elimination process guarantees that the growth will stop eventually. Additionally, elimination guarantees the convergence of the system as the feature dimensionality will start decreasing at some point till all features stop growing.