\n The central idea of this approach is to represent protein sequences using\n ProtLearn, a bioinformatics-focused Python library designed for extracting\n features from amino acid sequences.\n
\n\n\n ProtLearn implements a collection of well-established protein descriptors that encode\n sequence composition, local residue patterns, physicochemical properties, and positional\n distributions. Each descriptor family captures a different aspect of protein structure\n and function, allowing raw sequences to be transformed into meaningful numerical vectors.\n
\n\n\n In this notebook, multiple ProtLearn feature groups are computed independently and\n concatenated to form a single 695-dimensional representation per protein.\n These representations serve as the input to a multi-label classifier that\n predicts Gene Ontology (GO) terms.\n
\n\n\n\n ๐ ProtLearn Documentation:\n \n https://protlearn.readthedocs.io/\n
\n ๐ ProtLearn GitHub:\n \n https://github.com/taylor-joren/protlearn\n \n
\n Protein sequences are loaded from FASTA files and converted into a pandas DataFrame\n containing UniProt entry IDs and raw amino acid sequences.\n
\n\n\n ProtLearn only accepts sequences composed of the 20 standard amino acids.\n However, the dataset includes some non-standard or ambiguous residues that must be\n resolved prior to feature extraction.\n
\n\n\n To ensure valid ProtLearn input, the following cleaning rules are applied:\n
\n\n\n After cleaning, all sequences strictly follow the standard amino acid alphabet and are\n fully compatible with ProtLearn feature extraction.\n
\n\n\n Cleaned protein sequences are converted into fixed-length numerical representations\n using ProtLearn, a feature-engineering library for protein sequence analysis.\n
\n\n\n For each protein, multiple complementary feature groups are extracted and concatenated\n into a single vector.\n
\n\n\n All feature groups are horizontally concatenated, resulting in a\n 695-dimensional feature vector per protein.\n
\n\n\n\n\n Precomputed ProtLearn features are loaded from disk and aligned using UniProt\n EntryIDs, while Gene Ontology (GO) annotations are grouped per protein to form\n multi-label targets. To account for the hierarchical nature of GO, each annotated\n term is expanded to include all its ancestor terms using the GO graph, after which\n only the top N most frequent GO terms are retained. These propagated label sets are\n then binarized into fixed-length target vectors and merged with the ProtLearn\n feature matrix to construct the final training dataset,\n
\n\n\n The model is a fully connected multi-layer perceptron (MLP) designed for\n multi-label protein function prediction. It consists of stacked hidden layers\n with batch normalization, ReLU activations, and dropout for regularization.\n Residual connections are added every two layers to facilitate gradient flow\n and improve training stability. The output layer maps to the fixed set of\n selected Gene Ontology terms, producing one score per label for each protein.\n
\n\n\n A custom PyTorch dataset is defined to serve ProtLearn feature vectors and\n multi-label GO targets, returning featureโlabel pairs during training and\n featureโID pairs during inference. The full training set is split into\n training and validation subsets using a fixed random seed to ensure\n reproducibility, and DataLoaders are constructed for efficient mini-batch\n processing with shuffling applied only to the training data.\n
\n\n\n The MLP model is trained using mini-batch gradient descent with the Adam optimizer,\n applying cross-entropy loss for multi-label classification. Training proceeds over\n multiple epochs, with performance evaluated on a validation set at each epoch using\n F1 score. A learning rate scheduler adapts the optimizer based on validation loss, and\n early stopping is implemented to save the model checkpoint corresponding to the best\n validation performance, preventing overfitting and ensuring optimal generalization.\n
\n\n\n Training progress is visualized by plotting both loss and F1 score over epochs\n for the training and validation sets. These plots provide insight into model\n convergence, performance trends, and potential overfitting. \n
\n\n\n The trained MLP model is loaded from the best checkpoint and used to predict\n Gene Ontology term probabilities for all test proteins. Predictions are\n computed in mini-batches, with a sigmoid applied to produce probabilities\n for each GO label. For each protein, only terms with positive probability\n are retained, sorted by confidence, and truncated to a maximum per-protein\n limit. Probabilities are formatted to three significant figures, and the\n resulting predictions are saved in tab-separated format for submission.\n
\n\n