Metabolite annotation mlp gnn ensemble

metabolite-annotation-mlp-gnn-ensemble

Summary

Ensemble spectral prediction (ESP) combines MLP and GNN models enhanced with multi-tasking on LDA spectral topic labels and attention mechanisms to improve metabolite annotation rank performance on ESI/LC-MS spectra. This skill reproduces the 23.7% average rank improvement of ESP over the MLP baseline and generates comparative performance metrics.

When to use

Use this skill when you have ESI/LC-MS test spectra requiring metabolite annotation and need to compare ensemble-based neural network predictions (ESP) against a baseline MLP model to quantify performance gains. Specifically apply this skill when your goal is to measure and report average rank and Rank@K metrics to validate that the combined MLP–GNN ensemble with multi-tasking and attention outperforms single-model baselines by a known target margin (≥23.7% improvement).

When NOT to use

• Input spectra are EI/GC-MS data (ESP shows improvements only on ESI/LC-MS; README states 'not EI/GC-MS data in NEIMS')
• Pre-trained models are unavailable and retraining is not feasible (ensemble models require prior MLP and GNN checkpoints)
• Spectral candidate set size differs substantially from the full NIST set (average rank metric is sensitive to candidate pool composition and generalization to smaller pools is not documented)

Inputs

• Pre-trained ESP ensemble model (PyTorch .pt file, e.g., ESP_can.pt)
• Pre-trained MLP baseline model (PyTorch .pt file, e.g., best_model_mlp_can.pt)
• Pre-trained GNN model (PyTorch .pt file, e.g., best_model_gnn_can.pt)
• ESI/LC-MS test spectral dataset (PyTorch geometric .pkl file, e.g., torch_tecand_1000bin_te_cand100_torchgeometric.pkl)
• LDA spectral topic labels (derived from pos_train.csv or equivalent training data)

Outputs

• Average rank metric for MLP baseline predictions (numerical value ± std dev)
• Average rank metric for ESP ensemble predictions (numerical value ± std dev)
• Rank@K curves for both models (K=1 to 20, cumulative fraction correct)
• Percentage performance improvement: ((ESP_rank − MLP_rank) / MLP_rank) × 100
• Comparison table with baseline vs. ensemble metrics and confidence intervals

How to apply

Load pre-trained ESP, MLP, and GNN models from the HassounLab/ESP repository. Load the ESI/LC-MS test spectral dataset (e.g., NPLIB1 or NIST-20 preprocessed data). Generate predictions on test spectra using the MLP baseline model and the ESP ensemble model (which internally combines weighted predictions from MLP and GNN components enhanced with LDA-based multi-tasking on spectral topic labels and attention mechanisms to capture peak dependencies). Compute average rank and Rank@K metrics for both models on the full NIST candidate set. Calculate percentage performance gain using ((ESP_rank − MLP_rank) / MLP_rank) × 100. Verify that the ESP average rank improvement meets or exceeds the 23.7% threshold; report results in a comparison table with both baseline and ensemble metrics, including standard deviations.

Related tools

• ESP (Ensembled Spectral Prediction) (Primary ensemble model that combines MLP and GNN predictions with multi-tasking on LDA spectral topics and attention mechanisms to generate ranked metabolite annotations) — https://github.com/HassounLab/ESP
• MLP baseline (NEIMS implementation) (Single-model baseline for comparison; generates spectral predictions via feed-forward neural network) — https://github.com/HassounLab/ESP
• GNN baseline model (Single-model baseline using graph neural networks to capture molecular graph structure; ensembled with MLP in ESP) — https://github.com/HassounLab/ESP
• LDA (Latent Dirichlet Allocation) (Generates spectral topic labels used as additional multi-tasking supervision to enhance MLP and GNN model representations)

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

• ESP average rank is lower (better) than MLP baseline average rank by ≥23.7% on the full NIST candidate set
• Rank@K curves for ESP are above (higher cumulative fraction) those of MLP across K=1 to 20
• Standard deviations reported for both metrics are non-zero and overlapping error bars confirm statistical stability
• Reported average rank ± std dev values match expected output ranges (MLP ~339 ± 1264, GNN ~241 ± 939, ESP ~279 ± 1170 on NPLIB1 test set)
• Output table includes both average rank and Rank@K metrics with consistent formatting and all 819 test spectra evaluated

Limitations

• Models and reported metrics are trained and evaluated on NPLIB1 data only; NIST-20 models cannot be published due to NIST license restrictions, limiting reproducibility on industry-standard datasets
• Improvement margin (23.7% MLP over GNN shown; 37.2% GNN over baseline) varies by model pairing and dataset; generalization to other ESI/LC-MS datasets or alternative candidate sets is not documented
• Attention mechanism and LDA multi-tasking enhance the model but require careful hyperparameter tuning (e.g., mt_lda_weight=0.01, correlation_mix_residual_weight=0.7); sensitivity to these parameters is not explored in the README
• Computational cost of ensemble inference (combining MLP and GNN predictions with attention) is not reported; evaluation on large-scale real-world spectra may require GPU acceleration (optional but recommended)

Evidence

• [other] ESP model achieves a 23.7% increase in average rank performance over the MLP model on ESI/LC-MS data: "We have shown improvements with ESP over the MLP model (implementation of NEIMS model (Wei et al., 2019) with a generalized dataset ESI/LC-MS but not EI/GC-MS data in NEIMS), in terms of a 23.7%"
• [readme] MLP and GNN models are enhanced by multi-tasking on LDA spectral topic labels and attention mechanisms: "First, 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"
• [readme] Ensemble model trained on ranking tasks to generate weighted average of MLP and GNN predictions: "Next, we create an Ensembled Spectral Prediction (ESP) model that is trained on ranking tasks to generate the average weighted MLP and GNN spectral predictions"
• [readme] Performance metrics are average rank and Rank@K on test spectra: "Our results, measured in average rank and Rank@K for the test spectra, show remarkable performance gain over existing neural network approaches"
• [readme] Pre-trained models on NPLIB1 dataset are available; NIST-20 models cannot be published: "The pretrained MLP, GNN, and ESP models on the NPLIB1 dataset are /pretrained_models/best_model_mlp_can.pt, /pretrained_models/best_model_gnn_can.pt, and /pretrained_models/ESP_can.pt,"