McMLP
Summary¶
McMLP (Metabolite response predictor using coupled Multilayer Perceptrons) predicts an individual's post-dietary-intervention ("endpoint") metabolite concentrations from their pre-intervention ("baseline") gut microbial composition, baseline metabolome, and the dietary intervention strategy. It is the first deep learning method purpose-built for this task, and outperformed Random Forest and Gradient-Boosting Regressor baselines on both simulated microbial consumer-resource model data and real data from six human dietary intervention studies [1].
Architecture¶
McMLP couples two multilayer perceptrons (ReLU activation, deliberately overparameterized with 6 layers and 2048 hidden units each rather than hyperparameter-tuned) in a two-step design:
- Step 1: baseline microbial composition (centered-log-ratio transformed relative abundances), baseline metabolome (log10-transformed concentrations), and the dietary intervention strategy (binary presence/absence or dose-proportional) predict the endpoint microbial composition.
- Step 2: the predicted (not actual) endpoint microbial composition from step 1, together with the baseline metabolome and intervention strategy, predicts the endpoint metabolomic profile.
Using the predicted rather than actual endpoint microbiome as input to step 2 during training keeps train/test conditions consistent, since at application (test) time no real endpoint data exists yet — the entire point is to predict it. Ablation against a single-MLP one-step baseline (same total capacity) confirmed the two-step decomposition improves prediction accuracy.
Evaluation¶
Three metrics based on per-metabolite Spearman correlation (ρ) between predicted and true concentrations were used throughout: mean ρ̄, the fraction of metabolites with ρ > 0.5, and the mean ρ of the top-5 best-predicted metabolites.
- Synthetic data (Microbial Consumer-Resource Model, 20 food resources/microbes/metabolites, 250 samples, 50 random 80/20 splits): McMLP outperformed Random Forest and Gradient-Boosting Regressor on all three metrics, most clearly when baseline metabolomic profiles were included as input (ρ̄ = 0.595 ± 0.005 vs. lower baseline-method scores). McMLP's advantage was largest at small training-sample sizes (20–50 samples), converging with the baselines as sample size grew past 50 — the opposite of the usual expectation that deep learning underperforms at small n.
- Real data (six dietary intervention studies — avocado, grains, walnut, almond, broccoli, and high-fiber/fermented foods, spanning 18–132 participants and 9–43 measured fecal metabolites): McMLP with baseline metabolomic profiles consistently gave the best performance across studies and metabolite classes (SCFAs, branched-chain fatty acids, bile acids, amino acids, carboxylic acids, lipids, monosaccharides). In the avocado study, including baseline metabolome data improved bile-acid prediction more (ρ̄ 0.182→0.346) than SCFA prediction (ρ̄ 0.260→0.262), attributed to weaker baseline-endpoint correlation for SCFAs.
- McMLP also outperformed mNODE, a prior state-of-the-art method for predicting metabolomic profiles from concurrently-measured microbial composition, which is not designed for the cross-timepoint (baseline→endpoint) prediction task addressed here.
Interpretability¶
Sensitivity analysis of the trained McMLP was used to infer tripartite food–microbe–metabolite interactions, which were validated against ground truth in the synthetic (consumer-resource model) setting and against literature evidence for the real dietary intervention datasets.
Availability¶
Source code and datasets: github.com/wt1005203/McMLP. Shotgun metagenomic sequencing data from the avocado intervention study is deposited at NCBI SRA (PRJNA1198318).