Automated Genomic Interpretation via Concept Bottleneck Models for Medical Robotics

1. Problem Statement

The integration of medical robotics and automation into clinical workflows (e.g., triage, specimen handling, decision support) is hindered by the "black-box" nature of standard deep learning models. In safety-critical genomic medicine, predictions must not only be accurate but also interpretable, auditable, and actionable.

• The Gap: Existing Explainable AI (XAI) methods largely focus on medical imaging. There is a lack of interpretable modeling for DNA sequence data that can bridge the gap between raw genomic inputs and actionable robotic decisions.
• The Challenge: DNA classification tasks (e.g., HIV subtype identification) traditionally rely on alignment-based methods (computationally expensive) or black-box classifiers that lack biological transparency. Furthermore, explanations in current models are often post-hoc visualizations rather than integrated components that drive decision-making policies.

2. Methodology

The authors propose a unified framework that transforms raw DNA sequences into interpretable decisions suitable for medical robots. The pipeline consists of four main stages:

A. Preprocessing: Chaos Game Representation (CGR)

To make DNA sequences compatible with Convolutional Neural Networks (CNNs) without alignment, the system uses Chaos Game Representation (CGR).

• Process: Symbolic DNA sequences ($A, C, G, T$) are mapped to vertices of a unit square. An iterative contraction process generates a 2D trajectory, which is rasterized into a fixed-size image (e.g., $256 \times 256$).
• Benefit: This creates an alignment-free, fixed-dimensional visual representation that preserves genomic signatures and evolutionary distances.

B. Core Architecture: Concept Bottleneck Model (CBM)

The core reasoning mechanism enforces predictions to flow through biologically meaningful intermediate concepts rather than opaque latent features.

• Encoder: A CNN encodes the CGR image into latent features ($z$).
• Concept Layer: A regression head predicts a vector of interpretable biological concepts ($\hat{c}$), such as GC content, CpG density, and k-mer frequencies.
• Classifier: The final class prediction ($\hat{y}$) is derived strictly from the concept vector $\hat{c}$, ensuring the decision is mediated by biological evidence.

C. Regularization and Supervision

To ensure the learned concepts are faithful to biological reality and clinically consistent, the model employs several loss functions:

1. Concept Fidelity Supervision: Aligns predicted concepts with deterministic priors computed directly from the sequence (e.g., actual GC counts).
2. Prior-Consistency Alignment: Enforces monotonicity constraints (e.g., ensuring higher CpG density correlates with specific risk classes) via a hinge-style penalty on classifier weights.
3. KL Distribution Matching: Aligns the distribution of the concept-based classifier with an auxiliary feature-based head (active only during training) to preserve discriminative power lost by the bottleneck.
4. Uncertainty Calibration: Uses temperature scaling and predictive entropy to ensure the model's confidence scores are reliable.

D. Recommendation Layer: Cost-Aware Decision Policy

The final component translates probabilistic outputs into actionable robotic policies.

• Utility Function: Calculates a utility score for actions (e.g., Store, Review, Retest) based on:
  • Expected action cost ($C(a, y)$).
  • Concept-informed risk ($w^T \hat{c}$).
  • Predictive uncertainty (Entropy).
• Optimization: Uses a pairwise ranking loss to align automated recommendations with clinician preferences, balancing accuracy, calibration, and clinical utility (minimizing unnecessary retests).

3. Key Contributions

1. Molecular-Level Interpretability: First integration of CGR with Concept Bottleneck Models for DNA, creating a framework where decisions are explicitly driven by biological concepts (GC, CpG, k-mers).
2. Joint Optimization Strategy: A novel training objective combining classification loss with concept consistency, prior alignment, and distribution matching to ensure high accuracy without sacrificing interpretability.
3. Actionable Robotic Integration: Moving beyond post-hoc explanations to a Recommendation Layer that converts concept-level evidence into executable decision policies (e.g., reducing retest rates), closing the loop for medical automation.
4. Disease-Agnostic Framework: While tested on HIV, the architecture relies on general sequence statistics, making it applicable to other tasks like SARS-CoV-2 variant classification or bacterial typing.

4. Experimental Results

The system was evaluated on two datasets: an in-house HIV gag gene dataset and the public LANL HIV Sequence Database.

• Classification Performance:
  • Achieved 99% Accuracy, 0.85 F1 Score, and 0.93 AUROC on the in-house dataset, outperforming strong baselines (XGBoost, SVM, Logistic Regression, and standard CNNs).
  • Maintained superior performance (0.99 Acc, 0.84 F1) on the diverse LANL dataset.
• Concept Prediction Fidelity:
  • Significantly outperformed Vanilla-CBM and post-hoc regressors in recovering biological signals.
  • GC Content: $R^2 = 0.874$ (vs. 0.512 for Vanilla).
  • CpG Density: $R^2 = 0.615$ (vs. -0.070 for Vanilla), showing a >75% relative improvement on this difficult signal.
  • High Pearson correlations ($r > 0.94$ for GC) and AUROCs ($>0.98$) confirmed the concepts were statistically faithful.
• Faithfulness & Causality:
  • Sufficiency: The model retained 95.8% of full-model accuracy using only the concept vector, proving concepts are sufficient for classification.
  • Necessity: Removing specific concepts (e.g., CpG) caused significant accuracy drops, confirming they are causally tied to decisions.
• Decision Utility:
  • The cost-aware recommendation layer reduced the Retest Rate by ~20% compared to rule-based proxies while maintaining high sensitivity and low Expected Calibration Error (ECE < 0.05).

5. Significance

This work establishes a foundational paradigm for trustworthy medical robotics in genomics. By enforcing a transparent reasoning chain (Sequence $\to$ CGR Image $\to$ Biological Concepts $\to$ Decision), the system addresses the "black-box" trust deficit in AI.

• Clinical Impact: It provides auditable evidence that clinicians can verify against biological priors, facilitating human-in-the-loop adoption.
• Operational Efficiency: The cost-aware layer optimizes resource usage (reducing unnecessary retests), making it suitable for high-throughput automated labs.
• Future Outlook: The framework paves the way for autonomous robotic systems capable of triage, monitoring, and decision support in genomic medicine, moving from simple prediction to interpretable, actionable automation.