Ibex

This is a series of 72 Keras3-based amino acid encoders for Heavy and Light chains of the B cell Receptor (BCR). These are the individual trained models for larger Ibex package that will be available after review from BioConductor. The package itself has the functionality to employ these models and interact with single-cell data, but due to the size of the trained models, the keras objects are stored here.

Sequence inputs 

• Human Heavy: 10000000
• Human Light: 5000000
• Human Heavy-Expanded: 5000000
• Human Light-Expanded: 2500000
• Mouse Heavy: 5000000
• Mouse Heavy-Expanded: 5000000

Model details

• Architecture: 512-256-128-256-512
• Parameters: 
  • Batch Size = 128
  • Latent Dimensions = 128
  • Epochs = 128
  • Loss = Mean Squared Error (CNN) & KL Divergence (VAE)
  • Activation = ReLU
  • Learning rate = 1e-6
  • Optimizers: Adam
• Early stopping was set to patients of 10 for minimal validation loss and restoration of best weights
• CNN autoencoders have batch normalization layers between the dense layers. 

More information available at the GitHub Repo