Concept Bottleneck Models

The architecture enforces a strict information bottleneck by separating prediction into two sequential stages. Stage 1 maps raw genomic sequence inputs to a predefined set of human-interpretable biological concepts (e.g., 'promoter presence', 'GC-rich region', 'splice donor site'). Stage 2 uses only these concept activation scores—not the raw nucleotides—to make the final prediction. This constraint guarantees that the model's reasoning is fully expressed in terms of the specified concepts, enabling direct auditing of which biological features drove a classification decision.

Unlike post-hoc attribution methods that discover patterns after training, CBMs require domain experts to define the concept vocabulary upfront. Concepts can include:

• Sequence motifs: TATA box, CpG islands, transcription factor binding sites
• Structural features: predicted nucleosome positioning, DNA bendability scores
• Functional annotations: known enhancer regions, conserved elements, coding potential The model learns to detect these concepts from sequence data during training, creating a shared vocabulary between the algorithm and the biologist interpreting its outputs.

A defining capability of CBMs is that practitioners can directly edit concept predictions at inference time to correct model errors without retraining. If a clinician observes that the model incorrectly predicted 'promoter activity' for a specific genomic region, they can override that concept value, and the final prediction will update accordingly. This enables real-time expert-in-the-loop correction, making CBMs particularly valuable in regulated clinical genomics settings where erroneous predictions carry high risk.

The interpretability guarantee of a CBM is only as strong as its concept set. If critical biological features are missing from the concept vocabulary, the model may learn to encode them indirectly or suffer degraded accuracy. This creates a completeness-accuracy trade-off: a richer concept ontology improves fidelity but increases annotation cost. Research on hybrid CBMs addresses this by allowing a residual connection that passes limited raw sequence information to the final predictor, balancing interpretability with performance on tasks where complete concept enumeration is infeasible.

CBMs naturally integrate with Testing with Concept Activation Vectors (TCAV) for quantitative interpretability analysis. Because the bottleneck layer explicitly represents concept activations, practitioners can measure the sensitivity of the final prediction to each concept by computing directional derivatives. This produces a numerical score quantifying how much a specific biological concept (e.g., 'estrogen receptor binding site') influences a prediction (e.g., 'breast cancer susceptibility locus'), providing regulatory-grade audit trails for clinical decision support systems.

Two primary CBM architectures exist for genomic applications:

• Sequential CBM: Concepts are predicted first, then fed into the final classifier. Simple and fully interpretable, but concept prediction errors propagate forward.
• Independent CBM: Concept predictors and the final classifier are trained separately, with the classifier learning from ground-truth concept labels. This decouples concept learning from task learning, improving robustness to concept prediction errors at the cost of requiring fully annotated concept labels during training. The choice depends on the availability of curated genomic concept annotations.

A method that quantifies a model's sensitivity to user-defined, high-level biological concepts using concept activation vectors. Unlike CBMs, TCAV works post-hoc on already-trained black-box models by measuring how much a concept influences a prediction class. For genomic models, concepts might include GC content, splice site proximity, or chromatin state. TCAV produces a TCAV score indicating directional sensitivity, enabling hypothesis testing without architectural modification.

The general class of methods that assign a relevance score to each input nucleotide or genomic region for a specific neural network prediction. Key approaches include:

• Integrated Gradients: Axiomatic path integral method satisfying completeness
• DeepLIFT: Compares activations to a reference state using rescale rules
• SHAP: Unifies Shapley values for additive feature importance CBMs differ fundamentally by attributing to learned concepts rather than raw input features, providing a higher-level explanation layer.

A systematic perturbation technique that computationally mutates every nucleotide in a sequence to quantify its impact on model predictions. The resulting delta scores reveal which positions are functionally critical. ISM serves as a ground-truth validation method for CBM-learned concepts—if a concept bottleneck identifies a transcription factor binding motif as important, ISM can confirm that mutations within that motif cause the largest prediction shifts.

A method that identifies the minimal set of nucleotide changes required to flip a genomic model's prediction to a different outcome class. In CBM architectures, counterfactuals operate at the concept level—the model identifies which concept activations must change to alter the prediction, then maps those concept changes back to specific sequence edits. This provides actionable interpretability for genome editing applications.

A framework for quantifying the alignment between individual hidden units in a genomic model and human-interpretable biological sequence concepts. Each neuron is scored against a bank of concept masks to determine if it functions as a concept detector. CBMs formalize this idea architecturally—instead of discovering concepts post-hoc, the bottleneck layer is explicitly trained to predict pre-defined concepts, guaranteeing interpretability by construction.

Quantitative measures that evaluate how accurately an explanation reflects the true decision-making logic of a genomic model. Key metrics include:

• ROAR (Remove And Retrain): Iteratively retrains after removing top-attributed features
• AOPC (Area Over the Perturbation Curve): Measures prediction drop as salient positions are perturbed
• Infidelity Measure: Quantifies expected error between input and attribution perturbations CBMs achieve high faithfulness by design since predictions are constrained to use only concept activations.