Measured Proteins as Concept Labels for Multi-Omics Diagnosis

Cancer models that combine several molecular data types, such as gene expression, DNA methylation, copy number and protein levels, predict outcomes better than models that use one type. They are also opaque. The usual fix is to explain the model afterwards by ranking which genes mattered. That ranking cannot be checked against anything measured, and a clinician cannot act on it. We propose a model whose reasoning passes through a small set of intermediate quantities: the activity of named signaling pathways, such as PI3K-Akt or MAPK. The final prediction is a simple, readable function of those pathway scores. The new ingredient is how the pathway scores are supervised. They are trained against pathway activity computed from protein measurements, in the subset of patients who have such measurements. Those measurements already exist in cancer cohorts but are often missing for individual patients.

This gives the pathway layer two properties that current multi-omics models lack. First, its values can be checked: for a held-out patient whose protein data we hide, the model's predicted pathway activity can be compared with what was actually measured. Second, it can be corrected in a physically meaningful way. A clinician "intervenes" on a concept by ordering the protein assay and substituting the measured value for the predicted one. In a concept bottleneck model, intervention means a human overwriting an intermediate concept with its true value. Our design turns that abstract operation into an actual clinical decision, namely which assay to order. It also folds the chronic missing-data problem of multi-omics cohorts into the same mechanism.

Core Claim

We claim three things. First, a pathway bottleneck whose concepts are partially supervised by measured protein-derived activity will match the accuracy of a strong incomplete-view integrator on TCGA mortality and survival tasks. Second, its pathway scores will agree with held-out protein measurements far better than the pathway nodes of an unsupervised pathway-masked network. Third, revealing measured values for a few pathways at test time will recover most of the accuracy gain that comes from observing the whole protein view. Moreover, the size of that gain can be predicted per patient from the model's uncertainty about each concept. The third claim is what makes the bottleneck clinically meaningful rather than decorative.

How This Differs from Existing Work

Three lines of work each supply part of the design, and none combines the parts. The first concerns intervenable concept layers. Concept Bottleneck Models route predictions through human-readable concepts and show that correcting two concepts on knee X-rays cuts error from over 0.4 to about 0.3. Our test-time protocol comes from this work. However, it needs dense hand-labeled concepts and has only been used on images. Integrating Clinical Knowledge into Concept Bottleneck Models adds a loss that aligns concept importance with expert-ranked tiers and improves out-of-domain F1. This work constrains rather than enables our design, because hand-ranking hundreds of pathways does not scale. We replace expert labels with measurement.

The second line builds biological structure into the network. Predicting Anticancer Drug Response With Deep Learning Constrained by Signaling Pathways wires gene inputs into 46 KEGG pathway nodes and raises Pearson correlation from about 0.74 to about 0.85 on GDSC. We reuse its gene-to-pathway masks. In that model, though, the pathway nodes are never supervised and feed a 256-128-32 hidden-unit network, so nothing guarantees that a node labeled "MAPK" tracks MAPK activity. Interpretable meta-learning of multi-omics data for survival analysis and pathway enrichment gets pathways only after the fact, through attribution scores followed by enrichment analysis. That paper also reports that its proteomics view was noisy on its own. This warns us that protein-derived targets need careful construction and per-pathway reliability filtering.

The third line handles missing data. A Variational Information Bottleneck Approach to Multi-Omics Data Integration, called DeepIMV, handles arbitrary missing views by multiplying per-view Gaussian posteriors (a product of experts). Its authors list pathway-level interpretability as future work. We borrow its missing-view machinery but place it at the concept layer instead of an unnamed latent space. Patient-similarity graph methods, MOGONET and DeepMoIC, reason over patients rather than over molecular mechanisms. They need complete views, so they serve only as accuracy references.

Research Program

The model has three parts. View-specific encoders for mRNA, methylation, miRNA and copy number feed through the KEGG gene-to-pathway masks. Each pathway $k$ receives a Gaussian estimate of its activity from each available view, and these estimates are fused by a product of experts into a mean $\hat c_k$ and variance $\sigma_k^2$. A sparse linear head (Cox for survival, logistic for mortality) maps the vector $\hat c$, together with clinical covariates, to the outcome.

The concept targets $c_k$ are built only for pathways with enough measured proteins mapped to them. Each target is the mean of standardized abundances of member proteins, with phospho-forms signed by their known activating or inhibitory role. Pathways without enough measured proteins remain unsupervised, and we report them separately as such. The training objective is

$$ \mathcal{L} = \mathcal{L}_{task}(y, h(\hat c)) + \lambda \sum_{i \in P}\sum_{k \in S} \left[\frac{(c_{ik}-\hat c_{ik})^2}{2\sigma_{ik}^2} + \tfrac{1}{2}\log \sigma_{ik}^2\right], $$

where $P$ is the set of patients with protein data, $S$ is the set of supervised pathways, $h$ is the linear head, and $\lambda$ trades accuracy against concept fidelity. The Gaussian likelihood term makes $\sigma_{ik}$ a calibrated uncertainty, which drives the assay-selection step.

At test time, measured protein data, when available or ordered, enters as one more expert with a small fixed variance. That expert dominates the fused estimate, which is exactly a concept intervention. Because the head is linear, we can compute the expected reduction in outcome uncertainty from revealing pathway $k$ in closed form, from the head weight and $\sigma_k$. This yields a ranked list of which pathway assay to order for each patient.

The concept-bottleneck paper warns that adding a residual bypass lets the network ignore the concepts. We therefore build a variant with a small unconstrained side vector of fixed size and measure how much of its predictive signal flows through that vector rather than the concepts.

Baselines

The naive reference is a sparse Cox or logistic model on clinical covariates plus concatenated, standardized omics. This is the setting that Marinka Zitnik et al. describe as prone to overfitting and scale imbalance. It sets the floor that any structured model must clear.

The strongest existing method for our question is DeepIMV (A Variational Information Bottleneck Approach to Multi-Omics Data Integration). It reports a TCGA mortality AUROC of 0.791 with three views and degrades slowly as views go missing. It is the right competitor because it solves the missing-view problem we also face, without interpretable concepts. We will rerun it from the authors' design on our exact cohort and splits, with protein data added as a fourth view. It will get the same encoder widths and the same hyperparameter search budget as our model. We count as beating it if we match its AUROC within the confidence interval while winning clearly on concept fidelity and intervention efficiency. We do not claim higher raw accuracy.

For survival, the reference is the meta-learned multi-omics model of Interpretable meta-learning of multi-omics data for survival analysis and pathway enrichment. It reports C-indices of 0.79, 0.84 and 0.73 on BLCA, HNSC and LUAD. We will reproduce its splits and compare at equal training budget, again aiming for parity rather than superiority.

The key ablation isolates protein supervision. It uses the identical masked architecture and linear head with $\lambda = 0$, which amounts to consDeepSignaling with its hidden network replaced by a readable head. A second ablation replaces the linear head with consDeepSignaling's original multilayer head. This tells us how much accuracy a readable head costs.

Why not just…?

Why not just give DeepIMV the protein view and explain it with attribution scores? This is the main accuracy threat, and DeepIMV may well tie us on AUROC. What it cannot do is let anyone check its latent dimensions against measurements, or say which single assay would most change a given patient's prediction. Experiment two below tests exactly this. We give DeepIMV the full protein view and give our model only the top three predicted-most-useful pathway measurements, then compare how much of DeepIMV's gain we recover. If revealing three pathways recovers little of that gain, the clinical argument for our design fails, and we will say so.

Why not just use a pathway-masked network with a linear head and no protein supervision? This approach is cheaper and already has readable pathway names. However, nothing ties its node activations to biology, and its nodes have no meaningful "true value" to intervene with. The $\lambda = 0$ ablation measures this directly. If its pathway nodes correlate with held-out protein activity as well as ours do, protein supervision is unnecessary.

Why not just impute protein levels from the other omics, then fit a sparse additive model on the imputed pathway scores? This is essentially a sequential concept bottleneck. Combined with a reluctant additive head in the style of A reluctant additive model framework for interpretable nonlinear individualized treatment rules, it is a serious threat. We will run it as a baseline with the same masks and targets. We expect joint training and calibrated per-concept uncertainty to matter for the assay-ranking step, because a two-stage imputer gives no principled uncertainty to rank assays by. If it ties on fidelity, accuracy and intervention efficiency, our contribution reduces to the supervision-and-evaluation protocol, and the paper will be framed that way.

Proposed Experiments

All experiments use TCGA cohorts with mRNA, methylation, miRNA, copy number, protein and clinical data. These are the mortality setting of DeepIMV and the BLCA, HNSC and LUAD survival setting of the meta-learning study. Missing views are left as they occur naturally, and splits are stratified by cancer type.

The first decisive check is concept fidelity. For test patients who have protein data, we hide it and compute the correlation between predicted and measured pathway activity, per pathway. We compare the full model against the $\lambda = 0$ ablation and against the two-stage imputer. The model must clearly beat the unsupervised ablation, or the claim fails.

The second check is intervention efficiency. We reveal measured activity for one, two and three pathways, chosen either by our uncertainty-based ranking or at random. We record AUROC and C-index, and express the result as the fraction recovered of the gain from revealing the full protein view to DeepIMV. The uncertainty-based ranking must beat random choice.

The third check tests robustness to shortcuts using leave-one-cancer-type-out evaluation. This mirrors the out-of-domain tests in Integrating Clinical Knowledge into Concept Bottleneck Models. It asks whether concept-supervised models lose less accuracy than DeepIMV when tissue- and batch-specific correlations shift.

The fourth check is bypass leakage, run on the variant with an unconstrained side vector. We measure accuracy when the side vector is zeroed, and when the concepts are zeroed, to show how much of the prediction actually depends on the named pathways.

We report no prior number for protein-supervised pathway concepts. No reported result was found on Lacuna for concept fidelity of any multi-omics pathway model against measured protein activity.