Pathology AI systems read a stained tissue slide and output a diagnosis or risk score. When they are made "interpretable," it is usually by forcing them to first predict a short list of human-named visual features, such as "ulcer" or "spiculated margin," and then diagnose from those. These visual concepts are expensive to annotate. They also describe what a lesion looks like, not why it is diseased, and nothing can check whether the model's internal concept score is true for a given patient. We propose a different kind of intermediate concept, one that is biological and measurable. Spatial transcriptomics datasets pair each small patch of an H&E image (the standard pink-and-purple stained slide) with the gene expression measured at that spot. From that expression we can compute the activity of known biological pathways (cell cycle, immune signaling, hypoxia, and so on). We train an image model to predict these pathway activities and make the diagnosis depend only on them. The result is a concept bottleneck, meaning a model whose diagnosis passes through a small set of named intermediate quantities. Here every concept has a ground-truth value that can be measured independently for any patient who has RNA data.
In two sentences: we let paired image–expression data supply free, dense, mechanistic concept labels for a pathology diagnostic model. Because the Cancer Genome Atlas (TCGA) has both slides and bulk RNA for the same patients, we can then check whether each concept is faithful, and whether swapping in the measured values changes the diagnosis the way the bottleneck says it should. That last test cannot be run on any existing clinical concept bottleneck, because their concepts have no independent measurement.
Core Claim
A histopathology classifier whose only route to the diagnosis runs through predicted pathway-activity scores should satisfy three properties. First, it should come within a small margin of an unconstrained model on TCGA prediction tasks. Second, it should transfer across hospital sites better than that unconstrained model. Third, it should be intervenable: replacing predicted pathway scores with scores computed from the patient's own measured RNA should improve predictions. If that improvement does not happen, the bottleneck is a fiction, which makes this a test that can actually fail. The new ingredients are the source of the concepts (measured transcriptomic pathway activity rather than expert visual descriptors) and the evaluation (concept faithfulness checked against an independent molecular measurement rather than against saliency or annotator agreement).
Where this sits
Clinical concept bottlenecks motivate the design and also show its limits. Concept bottleneck models introduced predicting human concepts before the label. PN-SAMP predicts nine radiologist attributes alongside lung nodule malignancy. SkinCon shows that a post-hoc bottleneck over 48 dermatology concepts can beat a black box out of distribution: AUC 0.639 versus 0.595 for DeepDerm on DDI. SkinCon also shows the annotation ceiling, since only 22 of its concepts appear in at least 50 images. All of these rely on expert visual labels. We replace those labels with concepts computed from measured molecules, which removes the annotation bottleneck.
Two negative results constrain the design. Veltmeijer et al. found that forcing ABC criteria hurt accuracy on common nevi, a "distraction effect." CCBM found that concept sets are incomplete. CCBM closes the gap to within 1–2% of black boxes by adding an unsupervised complement branch whose concepts it cannot name. Taken together, these papers rule out a pure, closed bottleneck as the only design. We therefore include a residual branch, as CCBM does, but we measure and penalize how much of the decision it carries. We also try to name its features against pathways left out of the bottleneck.
The biological side supplies the method. NetActivity encodes expression into gene set activity scores using an autoencoder whose connections are masked by GO/KEGG membership. It reports greater than 90% cross-platform consistency versus about 44% for GSVA, which is why we use it rather than plain enrichment to define concept targets on noisy spot-level data. Its scores have arbitrary signs, so we anchor each score's sign to the mean expression of its member genes. PAST and STimage-1K4M show that morphology predicts expression: PAST reports a mean gene-level Pearson r of 0.619 versus 0.536 for UNI and 0.520 for CONCH. Both use dense embeddings with no bottleneck. We reuse that predictive link and make it the only path to the diagnosis.
Sparse autoencoder work is the closest neighbor, and it shapes our ablations. A sparse autoencoder (SAE) re-expresses a model's activations as a large dictionary of rarely active features. Pedrocchi et al. find biological and batch features in single-cell model SAEs and steer batch integration from 0.50 to 0.69, but about a third of their features stay uninterpretable. CB-SAE shows that unsupervised dictionaries miss 27–45% of the concepts users want, and adds a supervised concept module, but only on general vision-language benchmarks with human-named concepts. Our concept module is closer in spirit to CB-SAE. It differs in two ways: the supervision is a molecular measurement, and faithfulness is verified per patient against held-out RNA.
Research Program
The first stage builds concept targets. We train or reuse NetActivity-style pathway encoders on bulk expression and apply them to spatial spots from STimage-1K4M and the Xenium-derived PAST data. Because cell-level Xenium panels are targeted, pathway coverage there is restricted to well-covered gene sets. We keep only pathways whose scores are reproducible across adjacent sections or replicate slides. The test-retest correlation of measured scores sets the ceiling for any image predictor, so a pathway with unreliable measurements is dropped rather than learned.
The second stage trains the concept predictor. A small head sits on a frozen pathology encoder (UNI or CONCH features) and predicts the sign-anchored pathway scores per tile. The predicted scores are pooled to the slide by a learned attention weighting. Pooling is restricted so that a slide-level concept stays a weighted average of tile-level concepts, which keeps it in the same units as bulk RNA.
The third stage trains the diagnostic head on TCGA slides. Its input is the pooled concept vector plus a low-dimensional residual vector. An $L_1$ penalty on the residual pathway's contribution trades accuracy against how much of the decision is unexplained. We report the full trade-off curve rather than a single point.
The fourth stage is the audit. On TCGA patients with bulk RNA, we compute measured pathway scores with the same encoder and run three tests. The first measures concept fidelity: the correlation between the slide-level predicted and measured score for each pathway. The second is intervention: we replace predicted concepts with measured ones and measure the change in task performance. The third is directional consistency: for each concept, we check whether moving it by its prediction error shifts the output in the direction the head's weights imply. Residual features are then decomposed with a small SAE and correlated with measured scores of pathways outside the bottleneck, which gives the residual names where it can be named.
Baselines
The naive reference is a black box. An attention-pooled slide classifier is trained on the same frozen UNI/CONCH tile features, with the same splits, epochs and hyperparameter search budget, and no concept layer. This sets the accuracy the bottleneck must approach. It also sets the cross-site robustness the bottleneck must beat. Beating it means matching its in-site metric within the margin CCBM reports for its own setting (1–2%) while showing a smaller drop on held-out TCGA tissue-source sites.
The strongest existing competitor is a PAST-style dense morpho-molecular embedding fed to the same diagnostic head. PAST's reported numbers are a gene-expression Pearson r of 0.619 and a 0.91 H-score correlation with pathologists. For TCGA-BRCA survival it reports group separation, but no hazard ratio or log-rank p-value appears in the file, so no reported result was found on Lacuna for PAST's TCGA survival metric, and we will rerun it ourselves. If PAST weights are unavailable, we will train an equivalent contrastive image–expression encoder on the same spatial data with matched compute. Matching this model's accuracy while adding intervenability counts as success. Exceeding it out of site counts as a strong result.
The key ablation is a predicted-genes pipeline. It predicts individual gene expression from the image, as PAST does, then computes GSVA pathway scores on the predicted genes, then fits the same head. It uses identical data and differs only in whether pathway structure is the learning target. This isolates whether learning pathway activity directly gives more faithful, more robust concepts than composing noisy per-gene predictions.
A second ablation trains an unsupervised SAE on the frozen encoder with a linear head over its features, with features labeled post hoc by correlation with measured pathways. This tests whether supervision is needed at all.
For clinical concept bottlenecks, we cite SkinCon's PCBM AUC of 0.639 on DDI and Pang et al.'s out-of-domain F1 gain from 43.64% to 49.75% on RaabinWBC. Both serve as evidence that concept routing can help under shift. They are not directly comparable, because no expert visual concept set exists for TCGA slides.
Why not just…?
Why not just run Grad-CAM or SHAP on the black box? Saliency shows where the model looks, not what quantity it is using. Soelistyo et al. show that equally accurate models can rely on different internal logic, and that only active perturbation separates these cases. Saliency also offers no handle for correcting a prediction. The intervention test in stage four is exactly what saliency cannot provide, so we report it for the bottleneck. For the black box, the closest analogue is to check whether its heatmaps track the measured pathway maps on held-out spatial slides.
Why not just predict genes from H&E and run enrichment, as PAST-style models already allow? This is the most serious threat, and it is a main ablation, not a strawman. NetActivity's 90% versus 44% cross-platform consistency gap with GSVA suggests that composing gene-level predictions through an unstable projection will give noisier concepts. That evidence comes from measured bulk data, not predicted data, so it may not carry over. If the predicted-genes pipeline matches our concept fidelity and intervention gains, then the contribution narrows. It would then consist of the bottleneck-plus-audit protocol, not the pathway-level target, and we would report it that way.
Why not just train an unsupervised SAE on the pathology encoder and label its features afterward? It is cheaper and needs no retraining. However, CB-SAE reports that such dictionaries miss 27–45% of desired concepts and that only about 19% of neurons are both interpretable and steerable. Pedrocchi et al. leave about a third of features uncharacterized. The SAE ablation measures how many of our retained pathways have any SAE feature correlating above the fidelity of our supervised predictor.
Why not just add expression prediction as an auxiliary task without a bottleneck? This plausibly gives the best accuracy, and Veltmeijer et al.'s distraction effect argues that forcing concepts can cost accuracy. What a multitask model cannot offer is a decision that provably depends on the concepts, so intervening on its auxiliary outputs does nothing. We will run it as an upper-accuracy reference. If it is also more robust out of site, we will say so plainly. The case for the bottleneck would then rest on auditability alone.
Proposed Experiments
Concept fidelity is measured on held-out spatial slides from STimage-1K4M and PAST-20M, split by patient and by source study. The metric is per-pathway Pearson correlation between predicted and measured scores, normalized by the measured scores' test-retest reliability. It is compared against the predicted-genes-plus-GSVA ablation and the SAE ablation.
Slide-level transfer then uses TCGA-BRCA and TCGA-PRAD. These two cohorts are chosen because PAST and NetActivity respectively already used them. We measure correlation between pooled predicted and bulk-measured pathway scores. The downstream tasks are survival (concordance index) and available clinical endpoints (AUC). Evaluation holds out tissue-source sites, so that stain and scanner shift resemble the out-of-domain settings where Pang et al. and SkinCon found concept routing helpful.
The decisive checks are three. First, if replacing predicted concepts with measured ones does not improve the TCGA metric, the bottleneck is not doing what it claims, and we stop calling it interpretable. Second, if the residual branch must carry most of the decision to reach black-box accuracy at our chosen penalty, the pathway vocabulary is the wrong concept set for these tasks. Third, if the bottleneck's cross-site drop is no smaller than the black box's, the robustness argument fails, and only the audit contribution stands. Each outcome is informative, and the third is the one we consider most uncertain.
References
- Concept Bottleneck Models
- Joint learning for pulmonary nodule segmentation, attributes and malignancy prediction
- SkinCon: A skin disease dataset densely annotated by domain experts for fine-grained debugging and analysis
- Integrating Clinically-Relevant Features into Skin Lesion Classification
- Concept Complement Bottleneck Model for Interpretable Medical Image Diagnosis
- NetActivity enhances transcriptional signals by combining gene expression into robust gene set activity scores through interpretable autoencoders
- PAST: A multimodal single-cell foundation model for histopathology and spatial transcriptomics in cancer
- STimage-1K4M: A histopathology image-gene expression dataset for spatial transcriptomics
- Sparse Autoencoders Reveal Interpretable Features in Single-Cell Foundation Models
- Interpretable and Steerable Concept Bottleneck Sparse Autoencoders
- Integrating Clinical Knowledge into Concept Bottleneck Models
- Virtual perturbations to assess explainability of deep-learning based cell fate predictors