Create README.md

README.md

@@ -0,0 +1,80 @@
+---
+language:
+- en
+library_name: pyod
+tags:
+- protein-structure
+- anomaly-detection
+- one-class
+- autoencoder
+- protease-inhibitor
+- structural-filtering
+- biology
+- bioinformatics
+- unsupervised-learning
+
+datasets:
+- MEROPS
+- UniProt
+- AlphaFold
+
+model_name: Structural_module-protease_inhibitor
+
+model_description: |
+  Structural_module-protease_inhibitoris  an unsupervised, one-class deep learning model for filtering protein
+  structures that are structurally inconsistent with curated protease inhibitor (PI) like features learned from (RCSBembeddingmodel, github.com/rcsb/rcsb-embedding-model) from protease inhibitor databases. The model learns the structural embedding manifold of known protease
+  inhibitors  and assigns higher reconstruction error to structurally dissimilar inputs.
+user_interface: |
+  Easy-to-use inference interface:
+  "[![Open in Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/drive/1JLhLpvXG4plzPtIliG_CJnYui6P8Pu1J?usp=sharing)"
+training_data_description: |
+  The model was trained on 17,889 curated protease inhibitor structures from the MEROPS
+  database. MEROPS sequences were mapped via similarity search against taxonomy-
+  restricted UniProt datasets (fungi, plants, bacteria), and corresponding structures
+  were obtained from the AlphaFold Protein Structure Database and used for traning the model. 
+
+input_format: |
+  Fixed-length continuous protein structure embeddings derived from three-dimensional
+  structural features (RCSBembedding model, github.com/rcsb/rcsb-embedding-model). Embeddings must be standardized using the provided scaler.pkl
+  before inference.
+
+model_architecture: |
+  Fully connected autoencoder implemented in PyTorch via the PyOD library, featuring
+  a geometrically decreasing encoder, latent bottleneck, symmetric decoder, batch
+  normalization, dropout regularization, and mean squared reconstruction loss.
+
+training_procedure: |
+  The model was trained using the Adam optimizer with weight decay and mini-batch
+  stochastic gradient descent. Hyperparameters were optimized using Bayesian
+  optimization (Optuna, TPE sampler) on an independent 10% validation split
+  (~1,789 structures). The tuning objective was to minimize reconstruction error on
+  unseen but structurally valid protease inhibitor examples. The final model was
+  retrained on the full dataset using the optimal hyperparameters with fixed random
+  seeds.
+
+outputs: |
+  The model outputs a reconstruction-based anomaly score, an outlier probability,
+  and a confidence estimate. Low reconstruction-based anomaly scores indicate structural consistency with known
+  protease inhibitor folds, while high scores indicate structural dissimilarity.
+
+intended_use: |
+  Structural filtering and pre-selection of protease inhibitor–like protein structures
+  in large-scale datasets.
+
+limitations: |
+  Novel PI folds absent from the training data may be incorrectly rejected.
+
+not_intended_use: |
+  Functional annotation, clinical
+  decision-making.
+
+reproducibility: |
+  All preprocessing parameters, training configurations are provided
+  to enable exact reproduction of results.
+
+citation: |
+  Please cite the associated publication and acknowledge the PyOD library, the MEROPS
+  protease inhibitor database, and the AlphaFold Protein Structure Database.
+
+
+---