PepGlider

We present PepGlider, a continuous property regularization framework that enables direct control over their specific values. The method achieves structured latent space with superior disentanglement quality and displays smooth property gradients along regularized dimension. In silico experimental results demonstrate that PepGlider enables independent control of naturally correlated properties, and supports both de novo generation and targeted optimization of existing peptides. In application to antimicrobial peptide design, PepGlider generated candidates with desired antibacterial activity profile and low toxicity profile. Unlike existing approaches, PepGlider provides precise control over continuous property distributions, while maintaining generation quality, thus offering a generalizable solution for peptide design.

Packages requirements

All used packages and their versions are provided in requirements.txt file in repository. You can install them using:

pip install -r requirements.txt

Model training

You can run a model training using train.py file. All default parameters are defined in params_setting.py file. Current configuration enables you to train PepGlider model with MIC E.coli, MIC S.aureus and Nontoxicity regularization. Feel free to change configuration to train your model! Below there is a table with parameters description:

Parameter	Default value	Description
num_heads	4	The number of attention heads in the transformer layers.
num_layers	6	The number of transformer layers in the model's encoder/decoder.
layer_norm	True	Flag specifying whether Layer Normalization should be used.
latent_dim	56	The dimension of the latent space of the VAE model.
encoding	"add"	The method of combining or add input embedding vectors with positional ones in the architecture.
dropout	0.1	The dropout coefficient used for regularization during training.
batch_size	512	The batch size used during training the model.
lr	0.001	The learning rate for the optimizer.
kl_beta_schedule	(0.00001, 0.1, 8000)	The weight growth schedule for the KL-divergence term (e.g., start, target, steps to reach the target).
train_size	None	Optional limit on the size of the training set.
epochs	10000	The number of training epochs.
iwae_samples	10	The number of decoder training iterations used in the IWAE (Implicitly Weighted Autoencoder) model.
model_name	os.getenv(...)	The name of the model for saving files; retrieved from the CLEARML_PROJECT_NAME environment variable.
use_clearml	False	Flag specifying whether to use the ClearML platform for experiment tracking.
task_name	os.getenv(...)	The task name in ClearML; retrieved from the CLEARML_TASK_NAME environment variable.
device	"cuda"	The device for computation (cuda for GPU or cpu).
deeper_eval_every	20	How often (in epochs) to perform a more detailed evaluation.
save_model_every	100	How often (in epochs) to save the model state.
ar_vae_flg	True	Flag enabling AR-VAE or PepGlider model trainig.
reg_dim	[3, 4, 5]	The dimensions (indices) in the latent space that should be subjected to regularization.
gamma_schedule	(0.00001, 20, 8000)	The weight growth schedule for $\gamma$ hyperparameter.
gamma_multiplier	[1, 1, 1, 1, 1, 1]	Multipliers applied to $\gamma$ for individual dimensions.
factor_multiplier	[0.1, 0.1, 0.1, 0.6, 0.6, 0.6]	Multipliers for $\factor$ hyperparameter.
factor_schedule	(1, 1, 8000)	The weight growth schedule for factor regularization.
scale_factor_flg	False	Flag specifying whether to scale the $\factor$ hyperparameter.
mic_flg	True	Flag enabling add MIC values to the train dataset.
toxicity_flg	True	Flag enabling generate non-toxicity scores for the train dataset.
normalize_properties_flg	True	Flag specifying whether to normalize the input data properties.
signum_modification_of_dist_matrix_flg	False	Flag enabling modification of the distance matrix with the signum function.

Model evaluation

All final numeric results, tables and figures were produced using two notebooks:

• figures_notebook_1.ipynb
• figures_notebook_2.ipynb

Feel free to duplicate them and analyse models on this repo! We leave model files to upload in first_working_models directory.