The Bayesian Back End (BayBE) provides a general-purpose toolbox for Bayesian Design of Experiments, focusing on additions that enable real-world experimental campaigns.
Besides functionality to perform a typical recommend-measure loop, BayBE's highlights are:
- ✨ Custom parameter encodings: Improve your campaign with domain knowledge
- 🧪 Built-in chemical encodings: Improve your campaign with chemical knowledge
- 🎯 Single and multiple targets with min, max and match objectives
- ⚙️ Custom surrogate models: For specialized problems or active learning
- 🎭 Hybrid (mixed continuous and discrete) spaces
- 🚀 Transfer learning: Mix data from multiple campaigns and accelerate optimization
- 📈 Comprehensive backtest, simulation and imputation utilities: Benchmark and find your best settings
- 📝 Fully typed and hypothesis-tested: Robust code base
- 🔄 All objects are fully de-/serializable: Useful for storing results in databases or use in wrappers like APIs
Let us consider a simple experiment where we control three parameters and want to
maximize a single target called Yield.
First, install BayBE into your Python environment:
pip install baybe For more information on this step, see our detailed installation instructions.
In BayBE's language, the Yield can be represented as a NumericalTarget,
which we wrap into a SingleTargetObjective:
from baybe.targets import NumericalTarget
from baybe.objectives import SingleTargetObjective
target = NumericalTarget(
name="Yield",
mode="MAX",
)
objective = SingleTargetObjective(target=target)In cases where we are confronted with multiple (potentially conflicting) targets,
the DesirabilityObjective can be used instead. It allows to define additional
settings, such as how these targets should be balanced.
For more details, see the
objective section
of the user guide.
Next, we inform BayBE about the available "control knobs", that is, the underlying system parameters we can tune to optimize our targets. This also involves specifying their values/ranges and other parameter-specific details.
For our example, we assume that we can control three parameters – Granularity,
Pressure[bar], and Solvent – as follows:
from baybe.parameters import (
CategoricalParameter,
NumericalDiscreteParameter,
SubstanceParameter,
)
parameters = [
CategoricalParameter(
name="Granularity",
values=["coarse", "medium", "fine"],
encoding="OHE", # one-hot encoding of categories
),
NumericalDiscreteParameter(
name="Pressure[bar]",
values=[1, 5, 10],
tolerance=0.2, # allows experimental inaccuracies up to 0.2 when reading values
),
SubstanceParameter(
name="Solvent",
data={
"Solvent A": "COC",
"Solvent B": "CCC", # label-SMILES pairs
"Solvent C": "O",
"Solvent D": "CS(=O)C",
},
encoding="MORDRED", # chemical encoding via mordred package
),
]For more parameter types and their details, see the parameters section of the user guide.
Additionally, we can define a set of constraints to further specify allowed ranges and relationships between our parameters. Details can be found in the constraints section of the user guide. In this example, we assume no further constraints.
With the parameter and constraint definitions at hand, we can now create our
SearchSpace based on the Cartesian product of all possible parameter values:
from baybe.searchspace import SearchSpace
searchspace = SearchSpace.from_product(parameters)As an optional step, we can specify details on how the optimization should be conducted. If omitted, BayBE will choose a default setting.
For our example, we combine two recommenders via a so-called meta recommender named
TwoPhaseMetaRecommender:
- In cases where no measurements have been made prior to the interaction with BayBE,
a selection via
initial_recommenderis used. - As soon as the first measurements are available, we switch to
recommender.
For more details on the different recommenders, their underlying algorithmic details, and their configuration settings, see the recommenders section of the user guide.
from baybe.recommenders import (
SequentialGreedyRecommender,
FPSRecommender,
TwoPhaseMetaRecommender,
)
recommender = TwoPhaseMetaRecommender(
initial_recommender=FPSRecommender(), # farthest point sampling
recommender=SequentialGreedyRecommender(), # Bayesian model-based optimization
)We can now construct a campaign object that brings all pieces of the puzzle together:
from baybe import Campaign
campaign = Campaign(searchspace, objective, recommender)With this object at hand, we can start our experimentation cycle. In particular:
- We can ask BayBE to
recommendnew experiments. - We can
add_measurementsfor certain experimental settings to the campaign's database.
Note that these two steps can be performed in any order. In particular, available measurements can be submitted at any time and also several times before querying the next recommendations.
df = campaign.recommend(batch_size=3)
print(df) Granularity Pressure[bar] Solvent
15 medium 1.0 Solvent D
10 coarse 10.0 Solvent C
29 fine 5.0 Solvent B
Note that the specific recommendations will depend on both the data already fed to the campaign and the random number generator seed that is used.
After having conducted the corresponding experiments, we can add our measured targets to the table and feed it back to the campaign:
df["Yield"] = [79.8, 54.1, 59.4]
campaign.add_measurements(df)With the newly arrived data, BayBE can produce a refined design for the next iteration. This loop would typically continue until a desired target value has been achieved in the experiment.
BayBE has several modules to go beyond traditional approaches. One such example is the use of custom encodings for categorical parameters. Chemical encodings for substances are a special built-in case of this that comes with BayBE.
In the following picture you can see
the outcome for treating the solvent, base and ligand in a direct arylation reaction
optimization (from Shields, B.J. et al.) with
chemical encodings compared to one-hot and a random baseline:
(installation)=
The easiest way to install BayBE is via PyPI:
pip install baybeA certain released version of the package can be installed by specifying the
corresponding version tag in the form baybe==x.y.z.
If you need finer control and would like to install a specific commit that has not been released under a certain version tag, you can do so by installing BayBE directly from GitHub via git and specifying the corresponding git ref.
For instance, to install the latest commit of the main branch, run:
pip install git+https://github.com/emdgroup/baybe.git@mainAlternatively, you can install the package from your own local copy. First, clone the repository, navigate to the repository root folder, check out the desired commit, and run:
pip install .A developer would typically also install the package in editable mode ('-e'), which ensures that changes to the code do not require a reinstallation.
pip install -e .If you need to add additional dependencies, make sure to use the correct syntax
including '':
pip install -e '.[dev]'There are several dependency groups that can be selected during pip installation, like
pip install 'baybe[test,lint]' # will install baybe with additional dependency groups `test` and `lint`To get the most out of baybe, we recommend to install at least
pip install 'baybe[chem,simulation]'The available groups are:
chem: Cheminformatics utilities (e.g. for theSubstanceParameter).docs: Required for creating the documentation.examples: Required for running the examples/streamlit.lint: Required for linting and formatting.mypy: Required for static type checking.onnx: Required for using custom surrogate models in ONNX format.simulation: Enabling the simulation module.test: Required for running the tests.dev: All of the above plustoxandpip-audit. For code contributors.
BayBE collects anonymous usage statistics only for employees of Merck KGaA, Darmstadt, Germany and/or its affiliates. The recording of metrics is turned off for all other users and is impossible due to a VPN block. In any case, the usage statistics do not involve logging of recorded measurements, targets/parameters or their names or any project information that would allow for reconstruction of details. The user and host machine names are irreversibly anonymized.
- You can verify the above statements by studying the open-source code in the
telemetrymodule. - You can always deactivate all telemetry by setting the environment variable
BAYBE_TELEMETRY_ENABLEDtofalseoroff. For details please consult this page.
- Martin Fitzner (Merck KGaA, Darmstadt, Germany), Contact, Github
- Adrian Šošić (Merck Life Science KGaA, Darmstadt, Germany), Contact, Github
- Alexander Hopp (Merck KGaA, Darmstadt, Germany) Contact, Github
- Alex Lee (EMD Electronics, Tempe, Arizona, USA) Contact, Github
A list of know issues can be found here.
Copyright 2022-2024 Merck KGaA, Darmstadt, Germany and/or its affiliates. All rights reserved.
Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License.
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent