GeneDisco ICLR-22 Challenge Starter Repository

Python version
Library version

The starter repository for submissions to the GeneDisco challenge for optimized experimental design in genetic perturbation experiments.

GeneDisco (to be published at ICLR-22) is a benchmark suite for evaluating active
learning algorithms for experimental design in drug discovery.
GeneDisco contains a curated set of multiple publicly available experimental data sets as well as open-source
implementations of state-of-the-art active learning policies for experimental design and exploration.


pip install -r requirements.txt



  • Create a cache directory. This will hold any preprocessed and downloaded datasets for faster future invocation.
    • $ mkdir /path/to/genedisco_cache
    • Replace the above with your desired cache directory location.
  • Create an output directory. This will hold all program outputs and results.
    • $ mkdir /path/to/genedisco_output
    • Replace the above with your desired output directory location.

How to Run the Full Benchmark Suite?

Experiments (all baselines, acquisition functions, input and target datasets, multiple seeds) included in GeneDisco can be executed sequentially for e.g. acquired batch size 64, 8 cycles and a bayesian_mlp model using:

run_experiments \
  --cache_directory=/path/to/genedisco_cache  \
  --output_directory=/path/to/genedisco_output  \
  --acquisition_batch_size=64  \
  --num_active_learning_cycles=8  \

Results are written to the folder at /path/to/genedisco_cache, and processed datasets will be cached at /path/to/genedisco_cache (please replace both with your desired paths) for faster startup in future invocations.

Note that due to the number of experiments being run by the above command, we recommend execution on a compute cluster.
The GeneDisco codebase also supports execution on slurm compute clusters (the slurm command must be available on the executing node) using the following command and using dependencies in a Python virtualenv available at /path/to/your/virtualenv (please replace with your own virtualenv path):

run_experiments \
  --cache_directory=/path/to/genedisco_cache  \
  --output_directory=/path/to/genedisco_output  \
  --acquisition_batch_size=64  \
  --num_active_learning_cycles=8  \
  --schedule_on_slurm \
  --schedule_children_on_slurm \

Other scheduling systems are currently not supported by default.

How to Run A Single Isolated Experiment (One Learning Cycle)?

To run one active learning loop cycle, for example, with the "topuncertain" acquisition function, the "achilles" feature set and
the "schmidt_2021_ifng" task, execute the following command:

active_learning_loop  \
    --cache_directory=/path/to/genedisco/genedisco_cache \
    --output_directory=/path/to/genedisco/genedisco_output \
    --model_name="bayesian_mlp" \
    --acquisition_function_name="topuncertain" \
    --acquisition_batch_size=64 \
    --num_active_learning_cycles=8 \
    --feature_set_name="achilles" \

How to Evaluate a Custom Acquisition Function?

To run a custom acquisition function, set --acquisition_function_name="custom" and --acquisition_function_path to the file path that contains your custom acquisition function (e.g. in this repo).

active_learning_loop  \
    --cache_directory=/path/to/genedisco/genedisco_cache \
    --output_directory=/path/to/genedisco/genedisco_output \
    --model_name="bayesian_mlp" \
    --acquisition_function_name="custom" \
    --acquisition_function_path=/path/to/src/ \
    --acquisition_batch_size=64 \
    --num_active_learning_cycles=8 \
    --feature_set_name="achilles" \

…where "/path/to/" contains code for your custom acquisition function corresponding to the BaseBatchAcquisitionFunction interface, e.g.:

import numpy as np
from typing import AnyStr, List
from slingpy import AbstractDataSource
from slingpy.models.abstract_base_model import AbstractBaseModel
from genedisco.active_learning_methods.acquisition_functions.base_acquisition_function import \

class RandomBatchAcquisitionFunction(BaseBatchAcquisitionFunction):
    def __call__(self,
                 dataset_x: AbstractDataSource,
                 batch_size: int,
                 available_indices: List[AnyStr], 
                 last_selected_indices: List[AnyStr] = None, 
                 model: AbstractBaseModel = None,
                 temperature: float = 0.9,
                 ) -> List:
        selected = np.random.choice(available_indices, size=batch_size, replace=False)
        return selected

Note that the last class implementing BaseBatchAcquisitionFunction is loaded by GeneDisco if there are multiple valid acquisition functions present in the loaded file.

Submission instructions

For submission, you will need two things:

Please note that all your submitted code must either be loaded via a dependency in requirements.txt or be present in the src/
directory in this starter repository for the submission to succeed.

Once you have set up your submission environment, you will need to create a lightweight container image that contains your acquisition function.

Submission steps

  • Navigate to the directory to which you have cloned this repo to.
    • $ cd /path/to/genedisco-starter
  • Ensure you have ONE acquisition function (inheriting from BaseBatchAcquisitionFunction) in
    • This is your pre-defined program entry point.
  • Build your container image
    • $ docker build -t submission:latest .
  • Save your image name to a shell variable
    • $ IMAGE="submission:latest"
  • Use the EvalAI-CLI command to submit your image
    • Run the following command to submit your container image:
      • $ evalai push $IMAGE --phase gsk-genedisco-challenge-1528
      • Please note that you have a maximum number of submissions that any submission will be counted against.

That’s it! Our pipeline will take your image and test your function.

If you have any questions or concerns, please reach out to us at [email protected]


Please consider citing, if you reference or use our methodology, code or results in your work:

    title={{GeneDisco: A Benchmark for Experimental Design in Drug Discovery}},
    author={Mehrjou, Arash and Soleymani, Ashkan and Jesson, Andrew and Notin, Pascal and Gal, Yarin and Bauer, Stefan and Schwab, Patrick},
    booktitle={{International Conference on Learning Representations (ICLR)}},




Arash Mehrjou, GlaxoSmithKline plc
Jacob A. Sackett-Sanders, GlaxoSmithKline plc
Patrick Schwab, GlaxoSmithKline plc


PS, JSS and AM are employees and shareholders of GlaxoSmithKline plc.