Deep learning architecture design for sequence data

deep-learning-architecture-design-for-sequence-data

License: restricted — no clear open-source license detected for the underlying tool; verify licensing before commercial use or redistribution. <!-- asb-license-banner -->

Summary

Design and implement a balanced end-to-end deep neural network architecture that processes raw sequential elution intensity data without manual feature extraction to predict protein–protein interactions from co-fractionation/mass-spectrometry profiles. This skill is essential when raw co-elution data must be directly fed into a learner with class imbalance handling.

When to use

When you have raw co-fractionation/mass-spectrometry elution profiles (normalized intensity vectors across fractions) paired with labeled protein interaction ground truth, and you want to avoid hand-crafted features (e.g., correlation scores, distance metrics) in favor of learning representations end-to-end. Specifically, use this skill when elution data is available as raw intensity profiles (not pre-computed similarity scores) and you need to handle imbalanced positive/negative PPI labels.

When NOT to use

• Input elution data has already been pre-computed into correlation scores or engineered features (use Random Forest or classical classifiers instead, e.g., --feature_selection 11101001 with RF).
• You have only small numbers of labeled PPI examples (<100) and cannot generate sufficient negative samples without severe class imbalance.
• Elution profiles are sparse, high-dimensional, or contain extreme outliers that cannot be normalized effectively, as CNN performance degrades without stable numeric input.

Inputs

• raw co-fractionation/mass-spectrometry elution profiles (normalized intensity vectors, e.g., 2 profiles × 27 fractions per protein pair)
• gold standard PPI labels file (tab-separated: protein_A, protein_B, interaction_label)
• elution profile directory containing input data files

Outputs

• trained CNN model weights
• predicted interaction scores for all protein pairs
• evaluation metrics (precision, recall, AUC, confusion matrix)
• cross-validation fold results (if k-fold enabled)

How to apply

First, preprocess raw elution profiles by normalizing intensity values and handling missing data points to standardize input dimensions (e.g., 27 fractions per profile). Design a convolutional neural network (CNN) that accepts raw, stacked elution intensity arrays without prior feature extraction—SPIFFED uses this approach to eliminate manual feature engineering. Implement class-balancing during training via a weighted loss function (e.g., weighted cross-entropy or focal loss) or by controlling the negative-to-positive PPI ratio (e.g., --POS_NEG_RATIO 5 means 5× negative samples per positive). Train with supervised learning on labeled pairs, optionally using k-fold cross-validation to assess generalization (--K_D_TRAIN k --FOLD_NUM 5). Evaluate on held-out test data (default 30% split) by computing binary classification metrics (precision, recall, AUC) and interaction scores for each protein pair. Save trained model weights and predicted interaction scores to disk.

Related tools

• SPIFFED (Reference implementation of balanced end-to-end CNN for CF-MS interactome prediction; provides CNN, Label Spreading, and Random Forest classifiers with class-balancing strategies and raw elution profile input handling.) — https://github.com/bio-it-station/SPIFFED
• EPIC (Parent tool from which SPIFFED was derived; uses manual feature engineering rather than end-to-end learning.) — https://github.com/BaderLab/EPIC
• TensorFlow 1.13.1 / Keras 2.2.4 (Deep learning framework used by SPIFFED to define and train CNN layers.)
• scikit-learn (Provides preprocessing (normalization), cross-validation splitting, and evaluation metrics (AUC, precision, recall).)

Examples

python ./main.py -s 000000001 /path/to/elution/profiles -c /path/to/gold_standard.tsv /path/to/output -o out -M CNN -n 10 --LEARNING_SELECTION sl --K_D_TRAIN k --FOLD_NUM 5 --TRAIN_TEST_RATIO 0.3 --POS_NEG_RATIO 5 --NUM_FRC 27

Evaluation signals

• Input elution profiles are successfully loaded, have consistent dimensionality (e.g., all pairs have 2 profiles × 27 fractions), and pass normalization without NaN or Inf values.
• CNN model trains without divergence (loss decreases monotonically or via early stopping) and converges within expected epoch count; validation loss plateaus or improves on held-out fold data.
• Class balance is achieved during training: positive and negative samples are weighted or resampled so that the model does not collapse to trivial predictions (e.g., accuracy > 60%, AUC > 0.6 on test set for a balanced binary problem).
• Predicted interaction scores are in a bounded range (e.g., [0, 1] for sigmoid output) and show ranking power (AUC ≥ 0.65 or higher on test set, depending on signal-to-noise in the CF-MS data).
• Cross-validation fold results are consistent across folds (mean test AUC ± std < 0.1), indicating the model generalizes and is not overfitting to one fold.

Limitations

• SPIFFED requires Python 2.7 and older dependencies (TensorFlow 1.13.1, Keras 2.2.4); modern environments may face compatibility issues.
• Performance depends critically on gold standard label quality and completeness; sparse or noisy ground truth will degrade training signal.
• CNN architecture details (layer depth, filter counts, activation functions) are not fully exposed in the README, limiting reproducibility and hyperparameter tuning.
• Imbalanced negative PPI ratios (e.g., --POS_NEG_RATIO 5) artificially inflate negative samples, increasing training time and memory; very high ratios may harm generalization.
• Raw elution profiles must be normalized and aligned by fraction number; missing fractions or inconsistent preprocessing can cause shape mismatches and training failures.

Evidence

• [readme] SPIFFED differs from EPIC in that it uses a convolutional neural network to analyze raw co-elution data, thereby eliminating the need for manual feature engineering.: "SPIFFED differs from EPIC in that it uses a convolutional neural network to analyze raw co-elution data, thereby eliminating the need for manual feature engineering."
• [intro] A balanced end-to-end deep learning model for interactome prediction from co-fractionation/mass-spectrometry (CF-MS) data: "A balanced end-to-end deep learning model for interactome prediction from co-fractionation/mass-spectrometry (CF-MS) data"
• [readme] If you want to run Convolutional Neural Network (CNN) or Label Spreading (LS), you must set this parameter to "-s 000000001": "If you want to run Convolutional Neural Network (CNN) or Label Spreading (LS), you must set this parameter to "-s 000000001""
• [readme] This parameter stores the ratio of negative PPIs to positive PPIs. (default: 1): "This parameter stores the ratio of negative PPIs to positive PPIs. (default: 1)"
• [other] Preprocess elution data by normalizing intensity values and handling missing data points. Design a balanced deep neural network architecture with feature-extraction-free layers to accept raw elution profiles. Implement class-balancing strategy during training to handle imbalanced positive/negative PPI labels.: "Preprocess elution data by normalizing intensity values and handling missing data points. Design a balanced deep neural network architecture with feature-extraction-free layers to accept raw elution"