Skip to content

Limitations

This section outlines the conceptual, statistical, and computational limitations of CETSAx–NADPH. These are not implementation flaws; they reflect the assumptions and constraints of the underlying methodology.

Understanding these limitations is necessary for correct interpretation of results.

1. Model Assumptions

Logistic response constraint

The dose–response model assumes a sigmoidal relationship between concentration and stability.

  • Proteins with non-monotonic or multi-phase behavior are not well captured.
  • Complex binding mechanisms (e.g. multi-site, allosteric effects) are reduced to a single effective EC50.

As a result, some biologically valid responses may be filtered out or mischaracterized.

Monotonic smoothing

Isotonic regression enforces monotonicity.

  • This removes noise but also removes genuine non-monotonic structure.
  • Subtle transitions or intermediate states may be flattened.

2. Sensitivity Scoring

Heuristic weighting

The NADPH Sensitivity Score (NSS) is a weighted combination of features.

  • Weights are biologically motivated but not learned from data.
  • Different weighting schemes may change protein rankings.

The score should therefore be treated as a relative ranking, not an absolute quantity.

Dependence on fit quality

Sensitivity depends on fitted parameters.

  • Poor fits propagate into NSS.
  • Filtering reduces noise but may remove borderline cases.

3. Data Limitations

CETSA-specific constraints

CETSA measures thermal stability, not binding directly.

  • Stabilization may reflect:

    • ligand binding
    • complex formation
    • indirect metabolic effects

  • Destabilization may arise from:

    • competition
    • conformational shifts
    • degradation or aggregation

Therefore, CETSA signals are context-dependent proxies, not direct mechanistic measurements.

Replicate variability

  • Limited replicates reduce reliability of EC50 estimation.
  • High noise at low concentrations can bias fits.

Coverage bias

  • Not all proteins are equally detected.
  • Abundant or stable proteins are overrepresented.

This introduces systematic bias in downstream analyses.

4. Pathway and Network Analysis

Correlation-based networks

Co-stabilization networks are based on correlation.

  • Correlation does not imply causation.
  • Shared patterns may arise from indirect effects or experimental artifacts.

Annotation dependence

Pathway enrichment depends on annotation quality.

  • Missing or incorrect annotations affect results.
  • Broad pathway definitions dilute specificity.

5. Latent Representation

Linear assumptions (PCA)

  • PCA captures only linear structure.
  • Nonlinear relationships may be missed.

Factor interpretation

  • Latent factors are not uniquely identifiable.
  • Biological interpretation depends on feature loadings and context.

6. Sequence Modeling

Data size constraints

  • Performance depends on the number and quality of labeled proteins.
  • Small datasets limit generalization.

Label noise

Labels are derived from experimental fits.

  • Errors in EC50 or classification propagate into training.
  • The model may learn dataset-specific artifacts.

Model bias

  • ESM embeddings encode evolutionary information, not experimental context.
  • Predictions reflect both sequence signal and pretrained biases.

Limited structural awareness

  • The model operates on sequence, not explicit 3D structure.
  • Structural context is inferred implicitly, not modeled directly.

7. Explainability

Gradient-based attribution limits

  • Saliency and Integrated Gradients depend on model gradients.
  • These methods highlight sensitivity, not causality.

High importance does not guarantee biological relevance.

Resolution vs interpretation

  • Residue-level signals can be noisy.
  • Aggregation across proteins is required for robust conclusions.

8. Computational Constraints

GPU requirements

  • Sequence modeling is resource-intensive.
  • Limited GPU memory restricts batch size and model complexity.

Runtime scaling

  • Curve fitting scales with number of proteins × conditions.
  • Large datasets increase runtime significantly.

9. Interpretation Risks

Overinterpretation of single proteins

  • Individual hits can be misleading.
  • Reliable conclusions require consistent patterns across:
    • replicates
    • pathways
    • multiple analysis layers

Black-box misuse

  • The sequence model can produce confident predictions even when data is weak.
  • Interpretability tools mitigate this but do not eliminate the risk.

10. Summary

CETSAx–NADPH provides a structured and interpretable framework, but it operates under:

  • simplified biophysical assumptions
  • noisy experimental inputs
  • statistical and computational constraints

Results should be interpreted as:

  • hypothesis-generating, not definitive proof
  • strongest when supported across multiple layers of analysis

Careful validation with orthogonal experiments is recommended.