Patient-level proteomic network prediction by explainable artificial intelligence

The study addresses the challenge of inferring functional, patient-specific protein interaction networks from routine proteomic profiling data to support precision oncology. Traditional functional models (e.g., xenografts, organoids) are slow and technically challenging, and cohort-based network inference methods typically yield only average network features, missing individual heterogeneity. The authors hypothesize that if a neural network can accurately predict a target protein’s expression from other proteins’ expressions, those inputs have functional regulatory relationships to the target. Using explainable AI, specifically layer-wise relevance propagation (LRP), they aim to quantify the contribution (relevance) of each source protein to each target prediction, thereby reconstructing both average and individual (patient-level) interaction networks. This approach seeks to reveal tumor type–specific and patient-specific network characteristics that may underpin phenotypic differences such as metastasis and treatment response, enabling more personalized therapeutic strategies.

Multiple network inference methods from omics data have been proposed, including tree-based approaches like GENIE3 and others suited to homogeneous experimental systems. For clinical, heterogeneous data, the goal shifts from population averages to patient-specific networks. LIONESS was introduced to estimate sample-specific networks by interpolating between cohort-level reconstructions. Prior work showed that LRP can infer average interactions across samples. The authors build on these by using neural networks with LRP to capture nonlinear relationships and infer both cohort-level and individual networks directly from proteomic data, comparing against GENIE3 for average networks and LIONESS for individual networks.

Overview: Two-step pipeline: (1) Train a fully-connected neural network to impute hidden protein abundances from observed proteins; (2) Apply layer-wise relevance propagation (LRP) to attribute each input protein’s contribution to the prediction of each target protein, yielding interaction strengths. Neural network: Fully-connected with ReLU activations; three hidden layers; number of neurons per hidden layer = 10 × input dimension; learning rate 0.03; momentum 0.9; batch size 250; trained for 3600 epochs. Hyperparameters selected via 10-time repeated cross-validation with 50/50 train-test splits, optimizing mean squared error (achieved MSE ≈ 0.48). Training task: for each sample, each protein is hidden with probability p, with p drawn uniformly from [0.01, 0.99] per sample per iteration, producing a model capable of imputing from any subset of observed proteins. Inputs use an expanded form [x, 1−x], with hidden proteins set to [0,0]. Loss is MSE over hidden proteins. LRP: After training, LRP redistributes prediction relevance from outputs to inputs to quantify contributions. Starting from Rout = Yout for a target protein, relevance is propagated layer-wise using LRP-0/ε/γ rules and a generalized symmetrized rule addressing positive/negative inputs/outputs. Hyperparameters: γ = 0.01 (robustness-bias tradeoff), ε = 1e−5 (numerical stability). For sample-wise networks: choose a target protein (always hidden); hide other proteins with p=0.5; predict the target; apply LRP to obtain input relevances; sum the two input neurons of the expanded representation to get a protein-level relevance. Repeat 100 random imputations and average to obtain raw LRP scores LRP_ij. Repeat for all targets to get a full matrix. Define undirected interaction strength as LRP_au = mean(|LRP_AB|, |LRP_BA|). Data: TCPA proteomic/phosphoproteomic data (TCGA-PANCAN19-L4.csv): 5114 samples, 258 proteins. ReactomeFI functional interaction data used for external validation. For phosphoproteins, interactions with their unphosphorylated counterparts and their partners were added. Proteins with fewer than four neighbors were excluded, yielding 147 proteins and 10,731 possible pairs (1838 Reactome-positive, 8893 negative). Data split equally into training and test; normalization (z-score) performed on training and applied to test. LRP interactions computed only on the test set. Evaluation on real data: Compute median LRP_au across samples as robust interaction strength; assess enrichment of top 100 predicted interactions in Reactome via hypergeometric test, and compare to GENIE3 (Bioconductor implementation; undirected score as average of directed GENIE3 scores). Visualize top interactions by tumor type; assess differences across tumor types via Kruskal–Wallis with Holm–Bonferroni correction. For patient-level networks, perform t-SNE on per-sample vectors of 10,731 LRP_au scores to identify clusters; visualize strongest raw LRP (signed) edges and cluster-wise median networks. Synthetic validation: Two datasets generated via a neural network-based simulator with controlled adjacency constraints and Gaussian noise with correlated features. SD1 (homogeneous): all samples share the same block-structured 4×8 group interactions. Compare LRP_au, Pearson’s |r|, and GENIE3 via ROC AUC. SD2 (heterogeneous): each sample belongs to one of four disjoint 8-protein groups with within-group interactions only; evaluate per-sample reconstruction via ROC AUC; compare to LIONESS (R implementation) using Pearson’s r baseline. Additional analyses: boxplots of inter-group interaction strengths; t-SNE of reconstructed interactions.

• Synthetic data (average networks, SD1): LRP_au achieved AUC 0.996 (CI 0.993–0.999), outperforming Pearson’s correlation (AUC 0.755, CI 0.709–0.800) and on par with GENIE3 (AUC 0.988, CI 0.983–0.993).
• Synthetic data (individual networks, SD2): LRP achieved AUC 0.934 (CI 0.933–0.935) per-sample. LIONESS with Pearson’s r achieved AUC 0.893 (CI 0.892–0.894). For interactions that were present in some samples but absent in others, LRP AUC was 0.956 (CI 0.955–0.956) vs LIONESS 0.739 (CI 0.737–0.741).
• Real TCPA data (147 proteins, 10,731 pairs): Top 100 median LRP_au interactions included 56 present in Reactome (p = 1.1×10^−18, hypergeometric); GENIE3 captured 42 (p = 3.8×10^−9). Interactions between unphosphorylated proteins and their phosphorylated variants had highest interaction scores (median 0.47, IQR 0.80) compared to others (median 0.28, IQR 0.31; p < 1e−16, Mann–Whitney U).
• Numerous known interactions recovered with high scores, many in the mTOR, MAPK, apoptosis, and cell-cycle pathways, e.g.: mTOR–Raptor (median 1.0); 4E-BP1–EIF4E (0.74); 4E-BP1–S6 (0.79/0.80); AKT–GSK3 (0.74/0.66/0.8/0.75); AKT–Tuberin (0.68); GSK3–Tuberin (0.83/0.81); NFκB–Tuberin (0.98); Rictor–Tuberin (0.68); β-Catenin–E-cadherin (0.91); EGFR–HER2 (0.96); LCK–SYK (0.77); LCK–PI3K (0.69); EGFR–SHC (0.8); BAD–P38-MAPK (0.77); MEK1–P38-MAPK (0.65); MAPK–MEK1 (1.5, strongest); MAPK–SRC (0.67); BCL2–BIM (0.79); BCL2–p27 (0.66); Caveolin1–Collagen VI (0.67); c-Jun–JNK (0.6); Cyclin B1–FOXM1 (0.88); MEK1–YB1 (0.68); S6–YB1 (1.04/0.67).
• Putative novel or less-established associations highlighted with high LRP_au and shared functional contexts: BID–Stathmin (0.72), BID–N-Cadherin (0.8), Caspase-7–LCK (0.78), Fibronectin–PAI-1 (0.78), Fibronectin–p21 (0.66).
• Cancer-type specificity and heterogeneity: Many interactions varied significantly by tumor type; examples: AKT strong in GBM and UCEC; Cyclin B1–FOXM1 pronounced in UCEC and ovarian cancer; EGFR–HER2 high in GBM, HNSC, LUAD. LCK–SYK showed strong tumor-level differences and was strongest in adenoid cystic carcinoma (n=46, fewer samples).
• Patient-level networks and clustering: t-SNE on per-sample interaction profiles separated tumor groups; clear clusters for KIRC (C3), PRAD (C11), THCA (C6, C7), KIRP (C8), GBM (C4) vs LGG (C10). THCA split across clusters (C6, C7, C9). Some clusters mixed tumor types, indicating tumor type–independent network features. Cluster 2 (stomach, LUAD, PAAD, COAD, READ) showed highly similar networks centered on PARP, Caspase-8, Snail, c-MET, ERCC1, and RB, consistent with prior LUAD findings. In GI, lung, and uterine cancers, interactions among these proteins showed bimodal distributions, suggesting concerted pathway activity in subsets of patients. Interactions among these proteins were highly correlated (Pearson r up to 0.99), implying a common regulatory mechanism.

The findings demonstrate that explainable AI via LRP can infer both cohort-level and patient-level protein interaction networks from proteomic data, capturing nonlinear relationships and providing continuous measures of interaction strength. On synthetic data, LRP matched or exceeded state-of-the-art methods for average networks and clearly outperformed LIONESS for individual network reconstruction, especially for differential interactions across samples. On real TCPA data, LRP predictions were significantly enriched for known Reactome interactions and recovered many canonical signaling relationships (mTOR, MAPK, apoptosis, cell cycle). Importantly, patient-level analysis revealed clusters reflecting both tumor type–specific and cross-tumor shared network architectures, including a conserved pattern involving c-MET, ERCC1, Caspase-8, Snail, PARP, and RB that has been associated with drug resistance and progression in prior LUAD studies. These results support the utility of sample-wise network inference to uncover mechanisms potentially relevant to prognosis and therapy selection, moving beyond static molecular profiling toward functional network characterization in precision oncology.

This work introduces a neural network plus LRP framework to infer protein interaction networks at the level of individual tumor samples. The approach accurately reconstructs known interactions, shows strong performance on synthetic benchmarks for both average and individual networks, and identifies conserved and differential network patterns across major cancer types in TCPA. Notably, a network motif comprising c-MET, PARP, Caspase-8, RB, Snail, and ERCC1 appears in subsets of patients across several cancers, aligning with literature on therapy resistance. The method is generalizable to other molecular data types and could enhance predictive diagnostics by revealing patient-specific oncogenic mechanisms. Future work should investigate inference of causal directionality using asymmetric LRP scores, refine thresholds for actionability, integrate richer clinical annotations, and experimentally validate predicted novel interactions.

• Directionality: Current analysis reports undirected interaction strengths (LRP_au); causal direction was not established. Although two directed LRP scores exist, directionality inference was deferred to future work.
• Thresholding: Interaction strengths are continuous without a principled binary cutoff; regulatory relationships likely span a continuum.
• Validation: Many predicted interactions, especially putative novel ones, require experimental validation; clinical outcome associations were limited by available metadata.
• Data heterogeneity and sample size: Some tumor types (e.g., adenoid cystic carcinoma, n=46) had small sample sizes potentially affecting reliability of tumor-specific inferences.
• Generalization: Training data size is modest for deep learning; although random masking augments effective training cases, performance may depend on data quality and coverage of measured proteins.
• External baselines: LIONESS performance can depend on data distribution and duplicates; comparisons used Pearson’s r implementation as in the original work, which may not reflect all possible baseline configurations.