blox2 1.1.1

BLOX 2:

BLOX2

BLOX2 is an improved version of BLOX¹. The original implementations are available at https://github.com/tsudalab/BLOX.

Installation (with tutorials)

1. Clone the repository
2. Install uv: https://docs.astral.sh/uv/getting-started/installation/
3. Restart the shell
4. Move to the repository root (e.g., cd BLOX2)
5. Run the following commands:

uv venv --python 3.11.11
source .venv/bin/activate
uv pip install blox2 numpy==2.4.1 pandas==2.3.3 matplotlib==3.10.8 scikit-learn==1.7.2 lightgbm==4.6.0 ipykernel==7.1.0

Installation (package only)

uv pip install blox2 numpy==2.4.1 pandas==2.3.3 matplotlib==3.10.8 scikit-learn==1.7.2 lightgbm==4.6.0

Tutorial

For a quick start, see example_usage.ipynb.

Parameters

SteinNoveltySelector

SteinNoveltySelector selects candidates based on Stein novelty in the predicted objective space.
It can optionally incorporate predictive uncertainty, input-space novelty, distributional predictions, and batch diversity penalties.

Required inputs

observed_features, observed_values, and unobserved_features can be provided as either np.ndarray or pd.DataFrame.

Parameter	Description
observed_features	Feature values of already observed samples
observed_values	Objective/property values of already observed samples.
unobserved_features	Feature values of candidate samples to be selected from.
predictor	Predictor used to estimate objective/property values for unobserved candidates.

Normalization and prediction settings

Parameter	Default	Description
normalize_features	True	Whether to normalize input features before prediction.
value_normalization	"before_pred"	How objective/property values are normalized. Options: "before_pred", "after_pred", "mixed", or "disable".
pred_clip	None	Valid ranges for predicted objective values, given as a list of (min, max) tuples. Cannot be used with value_normalization="before_pred".
sigma	1.0	Gaussian kernel bandwidth used for Stein novelty calculation in objective space.
n_obs_samples	None	If the number of observed samples exceeds this value, a subset of observed samples is used for Stein novelty calculation.

Uncertainty-aware selection

Parameter	Default	Description
use_uncertainty	False	Whether to combine predictive uncertainty with Stein novelty.
uncertainty_ratio	0.5	Weight of the uncertainty score. The final score is (1 - uncertainty_ratio) * standardized Stein novelty + uncertainty_ratio * standardized uncertainty.
uncertainty_aggregation_type	"mean"	How to aggregate uncertainty over objectives/features. Options: "mean", "max", or "l2".
print_uncertainty	False	Whether to print uncertainty information during selection.

Input-space Stein novelty

Parameter	Default	Description
use_input_stein_novelty	False	Whether to also use Stein novelty in the input feature space.
input_stein_novelty_ratio	0.5	Weight of input-space Stein novelty. The final score is (1 - input_stein_novelty_ratio) * output-space Stein novelty + input_stein_novelty_ratio * input-space Stein novelty.
input_stein_pca_dim	None	If set to a positive integer, PCA is applied to the processed input features before input-space Stein novelty or Stein-based batch penalty calculation.
input_stein_sigma	"auto"	Gaussian kernel bandwidth for input-space Stein novelty. If "auto", the mean pairwise distance is used.
input_stein_auto_n_samples	10**5	Maximum number of sample pairs used to estimate the mean pairwise distance when input_stein_sigma="auto".

Batch diversity penalty

Parameter	Default	Description
use_batch_penalty	False	Whether to penalize candidates that are too close to already selected candidates in the same batch.
batch_penalty_ratio	0.5	Weight of the batch penalty. The batch score is (1 - batch_penalty_ratio) * standardized base score - batch_penalty_ratio * standardized batch penalty.
batch_penalty_type	"stein"	Type of batch penalty. Options: "stein", "distance", "simhash", or "simhash_min_hamming".
batch_penalty_cutoff_ratio	0.0	Skips batch-penalty calculation for low-scoring candidates within each chunk.
batch_penalty_simhash_bits	8	Number of SimHash bits used when batch_penalty_type is "simhash" or "simhash_min_hamming". Must be at most 64.

Distributional prediction

Parameter	Default	Description
use_distribution	False	Whether to use predictive distributions from predictor.pred_samples().
distribution_pooling_type	"mean"	How to aggregate Stein novelty scores over predicted samples. Options: "mean" or "max".

Misc

Parameter	Default	Description
chunk_size	256	Number of candidates processed in one chunk during chunked Stein novelty calculation.
verbose_plot_dir	None	If set, saves diagnostic plots of predicted values and selected candidates at each selection step.

Predictor

Although LightGBMPredictor was mainly used in our experiments, BLOX2 is not limited to LightGBM. Any predictor can be used by implementing the Predictor class. The package also includes sklearn-based predictor implementations, such as RandomForestPredictor and SVRPredictor, although these implementations do not support uncertainty estimation. When using these predictors, the lightgbm dependency can be omitted.

License

The BLOX2 source code is licensed under the MIT License.

The files under data/ are provided for reproducibility and are not covered by the MIT License. See data/README.md for the corresponding attribution and license information.

1. Terayama, K. and Sumita, M. and Tamura, R. and Payne, D. T. and Chahal, M. K. and Ishihara, S. and Tsuda, K. (2020). Pushing property limits in materials discovery via boundless objective-free exploration. Chemical Science https://doi.org/10.1039/D0SC00982B ↩