Ensemble model weight learning for spectral prediction

Ensemble Model Weight Learning for Spectral Prediction

Summary

Train an ensemble weighting layer on ranking tasks to learn optimal weights for combining MLP and GNN spectral predictions into a single weighted-average prediction for metabolite annotation. This approach improves annotation accuracy by leveraging complementary strengths of both neural network architectures through task-specific optimization.

When to use

You have pre-trained MLP and GNN models that generate different spectral predictions for the same metabolite candidates, and you want to combine them to improve ranking performance (average rank, Rank@K metrics) without retraining the base models. Use this skill when you have a ranking task dataset with true metabolite rank labels and aim to learn adaptive weights for each model's contribution.

When NOT to use

• Base MLP and GNN models are not yet trained or enhanced with multi-task LDA and attention mechanisms—train them first using task_id=task_001 and task_id=task_002.
• You lack a ranking task dataset with true metabolite rank labels; ensemble training requires supervised ranking objectives.
• Your goal is only to predict spectral properties (e.g., fragment intensities) rather than to rank metabolite candidates—use individual MLP or GNN models directly.

Inputs

• Pre-trained MLP model (PyTorch .pt file with multi-task LDA and attention enhancements)
• Pre-trained GNN model (PyTorch .pt file with multi-task LDA and attention enhancements)
• Ranking task dataset with spectra and ground-truth metabolite rank labels (e.g., pos_train.csv formatted data)
• Test candidate set with candidate size and resolution parameters (e.g., 100 candidates per spectrum, 1000 bins)

Outputs

• Trained ensemble weighting parameters (PyTorch .pt model file, e.g., ESP_can.pt)
• Weighted-average spectral predictions for test spectra
• Ranking metrics on test set (average rank ± std dev, Rank@1 through Rank@20)

How to apply

Load the pre-trained MLP and GNN models enhanced with multi-task learning on LDA spectral topic labels and attention mechanisms. Prepare a ranking task dataset containing spectra with ground-truth metabolite rank labels. Initialize an ensemble weighting layer and define a ranking loss function (listwise or pairwise ranking objective). Train the ensemble on the ranking dataset using gradient descent to learn optimal weights that combine MLP and GNN outputs as: weighted_prediction = w_mlp * mlp_output + w_gnn * gnn_output. Evaluate the ensemble on a held-out validation set using ranking metrics (average rank, Rank@K) to verify that the learned weights improve upon individual model baselines.

Related tools

• PyTorch (Deep learning framework for defining, training, and saving ensemble weighting layers and ranking loss functions)
• DGL (Deep Graph Library) (Framework for managing GNN models and graph-based spectral data processing)
• PyTorch Geometric (Geometric deep learning library for loading and handling GNN candidate ranking datasets)
• LDA (Latent Dirichlet Allocation) (Generates spectral topic labels used as multi-task learning targets to enhance base MLP and GNN models)
• ESP (Full implementation of ensemble model, training loop, and ranking evaluation metrics) — https://github.com/HassounLab/ESP

Examples

python ens_train_canopus.py --cuda 0 --disable_two_step_pred --disable_fingerprint --disable_mt_fingerprint --disable_mt_ontology --correlation_mat_rank 100 --full_dataset --mode 'canopus'

Evaluation signals

• Ensemble average rank on test set is lower (better) than both baseline MLP and baseline GNN average ranks—expected improvement ≥23.7% over MLP baseline (e.g., MLP: 339.35 → ESP: 184.8 or better).
• Rank@K metrics (Rank@1–Rank@20) for the ensemble are monotonically increasing and exceed both individual model curves across all K.
• Learned ensemble weights (w_mlp, w_gnn) are non-negative and sum approximately to 1.0, indicating a valid weighted combination.
• Training loss (ranking objective) on the training set decreases monotonically and plateaus; validation ranking metrics do not show consistent degradation (no severe overfitting).
• Ensemble predictions on held-out test set are reproducible: same model file and hyperparameters yield identical average rank and Rank@K values.

Limitations

• Ensemble weight learning depends critically on the quality and diversity of base MLP and GNN models; if both models are poorly trained or highly correlated, ensemble improvement is limited.
• Ranking task dataset must have sufficient size and representative metabolite rank distribution; small or biased ranking datasets may lead to overfitted or poorly generalized weights.
• The method is evaluated on ESI/LC-MS data and NPLIB1 dataset; performance on EI/GC-MS data or other spectral databases is not reported and may differ.
• Pre-trained models are published only for NPLIB1 data; NIST-20 models cannot be published due to license restrictions, limiting reproducibility on proprietary datasets.

Evidence

• [intro] Ensemble model trained on ranking tasks to generate weighted average MLP and GNN predictions: "Ensembled Spectral Prediction (ESP) model that is trained on ranking tasks to generate the average weighted MLP and GNN spectral predictions"
• [intro] Multi-task learning on spectral topic labels and attention mechanisms enhance base models: "the MLP and GNN are enhanced by: 1) multi-tasking on additional data (spectral topic labels obtained using LDA (Latent Dirichlet Allocation), and 2) attention mechanism to capture dependencies among"
• [intro] ESP ensemble shows 23.7% average rank improvement over MLP baseline: "23.7% increase in average rank performance over MLP model on ESI/LC-MS data"
• [readme] Training procedure for ensemble on ranking tasks using pre-trained models: "Before you train a new ESP model, you must have pretrained MLP and GNN models. To train a new ESP model, set --te_cand_dataset_suffix to an empty string or don't call this argument."
• [readme] Example ranking metric evaluation (Average rank ± std dev and Rank@K): "Average rank 279.557 +- 1170.300 Rank at 1 0.187 Rank at 2 0.277 Rank at 3 0.328"