ISAAC: Auditing Causal Reasoning in Deep Models for Drug-Target Interaction

Technical Summary: ISAAC – Auditing Causal Reasoning in Deep Models for Drug–Target Interaction Prediction

1. Problem Statement

Deep learning models frequently achieve high predictive accuracy on scientific benchmarks, yet this performance does not necessarily imply that the models have learned mechanistically grounded reasoning. Instead, models may rely on spurious correlations, dataset-specific artifacts, or "shortcut" features that correlate with the target in-distribution but lack causal validity. This issue is particularly critical in scientific domains like Drug–Target Interaction (DTI) prediction, where valid outputs must depend on underlying molecular mechanisms (e.g., localized protein–ligand interactions) rather than surface-level statistical patterns.

Current evaluation paradigms rely heavily on predictive accuracy, an observational metric that assesses agreement with labels on a fixed data distribution. Accuracy alone cannot distinguish between causal dependence and spurious association. While existing interpretability methods (e.g., saliency maps) offer correlational explanations, they do not guarantee causal validity, and robustness-based approaches often focus on training objectives rather than auditing the reasoning behavior of frozen, trained models. Consequently, there is a lack of a standard framework to explicitly test whether predictions are driven by structurally meaningful input components.

2. Methodology: The ISAAC Framework

To address this gap, the authors introduce ISAAC (Intervention-based Structural Auditing Approach for Causal Reasoning). ISAAC is a post-hoc structural auditing framework that evaluates a model's reasoning as a behavioral property conditional on its predictive regime, using controlled input-level interventions.

Core Concepts:

• Interventional Behavior: Reasoning is probed by applying deterministic operators $I$ to inputs $x$, creating intervened inputs $x^I$. The model $f_\theta$ is treated as a fixed system; variations in output are induced solely by the intervention.
• Structured Inputs: Inputs are decomposed into identifiable components (e.g., amino acid residues in a protein sequence).
• Matched Intervention Classes: The framework defines two classes of interventions based on an externally specified structural prior (e.g., known binding pockets):
  • Mechanistic Interventions ($I_{mech}$): Perturbations applied to components identified as relevant by the structural prior ($S \subseteq S_{prior}$).
  • Spurious Interventions ($I_{spur}$): Perturbations applied to components outside the prior-defined region ($S' \subseteq S_{comp}$), matched in cardinality and operator to the mechanistic set.
• Interventional Response Difference: Defined as $\delta_\theta(x, I) = f_\theta(x^I) - f_\theta(x)$, capturing the relative change in model output.

Metrics:
ISAAC quantifies reasoning consistency through three metrics:

1. Reasoning Score (RS): A per-input measure defined as $RS(x) = \frac{|m_{mech}(x)|}{|m_{mech}(x)| + |m_{spur}(x)|}$, where $m$ represents the median of signed interventional responses. RS quantifies the relative preference for mechanistically aligned perturbations over misaligned ones.
2. Separation Coefficient ($C_{sep}$): Measures the geometric separation between the distributions of mechanistic and spurious responses, normalized by the intrinsic variability of mechanistic effects.
3. Overlap Rate: The fraction of spurious responses falling within the interquartile range of mechanistic responses.

3. Experimental Setup

The framework was instantiated in the context of Drug–Target Interaction (DTI) prediction using the Davis benchmark (kinase inhibitor interactions).

• Models: Three sequence-based architectures with varying inductive biases were audited: DeepDTA, DeepConvDTI, and TAPB (Target-Aware Prediction Benchmark).
• Intervention Design: Interventions were applied exclusively to the protein target sequence ($t$) while the drug ($d$) remained fixed.
  • Structural Prior: Derived from KLIFS binding-pocket annotations to identify mechanistically relevant residues.
  • Interventions: Included masking and physicochemically constrained substitutions. Mechanistic scopes were sampled from KLIFS pockets; spurious scopes were sampled from non-pocket residues with matched cardinality.
• Protocol: Models were trained and validated using standard procedures. Predictive performance (AUROC) was verified to ensure all models operated within a comparable regime. Auditing was performed post-hoc on a structurally valid subset of the test data (208 targets) without retraining.

4. Key Results

The experiments revealed a significant dissociation between predictive accuracy and reasoning behavior:

• Predictive Performance: All three models achieved nearly indistinguishable AUROC scores on the auditing subset (ranging from 0.876 to 0.907), confirming they operated in a comparable predictive regime.
• Reasoning Scores (RS): Despite similar accuracy, the models exhibited substantial and reproducible differences in their reasoning scores:
  • TAPB: Achieved the highest RS (0.521) and a high Separation Coefficient (0.458), indicating coherent separation between mechanistic and spurious responses.
  • DeepDTA: Showed intermediate RS (0.441) but low separation.
  • DeepConvDTI: Exhibited the lowest RS (0.414) and near-zero separation (0.061), indicating weak sensitivity to mechanistic perturbations and a lack of coherent preference for aligned structures.
• Conclusion on Accuracy: The ~3% difference in AUROC did not correlate with the magnitude of differences in reasoning scores (which spanned 0.414 to 0.521). This demonstrates that accuracy-based evaluation fails to detect differences in the structural organization of model responses.

5. Significance and Claims

The paper claims that predictive accuracy alone is insufficient to characterize model reasoning in scientific machine learning. The primary contributions are:

• ISAAC Framework: A principled, intervention-based auditing protocol that assesses prior-relative structural sensitivity independently of predictive accuracy.
• Behavioral Distinction: The ability to reveal systematic differences in reasoning behavior (e.g., reliance on localized binding sites vs. global sequence statistics) that remain invisible under standard accuracy-based evaluation.
• Model-Agnostic Auditing: The method operates post-hoc on frozen models, requiring no retraining, internal model access, or intervention-level supervision, making it applicable across architectures and domains with structured inputs.

The authors emphasize that their findings highlight a distinction between "predictive success" and the "organization of reasoning." By contrasting responses to structurally aligned versus misaligned perturbations, ISAAC provides a basis for distinguishing models that are indistinguishable under conventional benchmarks, offering a more transparent assessment for scientific applications where mechanistic validity is critical.

Limitations Acknowledged:

• The structural prior (KLIFS) may not capture all affinity-determining residues; thus, metrics reflect prior-relative sensitivity rather than absolute causal validity.
• Spurious controls matched in cardinality but not spatial contiguity, introducing a geometric asymmetry.
• The current study is restricted to target-side interventions on a single dataset and three specific architectures.