🧬 biology

An Explainable AI Framework Integrating Survival Analysis and Machine Learning for COX Inhibitor Evaluation

This paper presents an explainable AI framework that integrates machine learning algorithms with classical survival analysis to enhance the predictive accuracy and interpretability of COX inhibitor evaluation using benzimidazole triazole analogs.

Original authors: Maryam Hamidi, Parvin Azhdari, Kianoush Fathi Vajargah

Published 2026-06-24

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Maryam Hamidi, Parvin Azhdari, Kianoush Fathi Vajargah

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

The Big Picture: A "Taste Test" for New Medicine

Imagine you are a chef trying to create a new, super-delicious soup (a new drug) to cure a specific illness. You have a few secret ingredients (chemical compounds) and you want to know: Will this soup actually work, and how long will it keep people healthy?

Usually, chefs (scientists) have to cook the soup, serve it to a few people, and wait to see what happens. This takes a long time, costs a lot of money, and sometimes the results are confusing because the people eating the soup are all different.

This paper proposes a new way to test these "soups" before they even reach the kitchen. The authors built a digital simulation framework that acts like a super-fast, super-smart taste-tester. It combines two different ways of thinking about the problem:

The Statistician's Rulebook: A traditional, reliable way of calculating risks (like a strict recipe).
The AI's Intuition: A flexible, pattern-spotting computer brain that can find hidden connections (like a genius chef who tastes a pinch of salt and knows exactly what spice is missing).

The Three "Kitchens" (Datasets) Used

To prove their new framework works, the researchers didn't just use one set of data. They cooked in three different "kitchens":

The Small Real Kitchen (n=90): This was a real experiment done in a lab using human cancer cells. They tested a new chemical compound (a benzimidazole-triazole hybrid) on three different types of cells. It's like a small, real-world taste test with 90 tasters.
The Perfect Simulation Kitchen (n=1,000): Since real experiments are small and messy, they created a fake but mathematically perfect dataset of 1,000 virtual patients. This is like a video game simulation where they control every variable to see if their computer models can handle a large crowd without getting confused.
The Hybrid Kitchen (n=312): They mixed the real data with the fake data to create a "middle ground" group. This helped them test if their models could handle a medium-sized crowd fairly.

The Contenders: Who is the Best Predictor?

The researchers pitted six different "predictors" against each other to see who could best guess how long the cells would survive after taking the drug.

The Old Guard (Cox Proportional Hazards & Logistic Regression): These are like experienced, rule-following accountants. They are very good at explaining why something happened (e.g., "The dose was too high, so survival dropped"). They are easy to understand, but they sometimes miss complex, hidden patterns.
The New Guard (Random Forest, SVM, Deep Learning): These are like super-observant detectives. They look at thousands of clues at once and find complex, non-linear patterns that accountants might miss. They are great at guessing the outcome, but sometimes they are "black boxes" where you don't know exactly how they reached the conclusion.
The Team Captain (Ensemble): This is a group where all the accountants and detectives vote together.

The Results: Who Won?

The paper found that different tools are best for different jobs:

For pure guessing power (Accuracy): The Random Forest (the detective) won. It was the best at predicting who would survive and who wouldn't, achieving a high score of 89% accuracy. It was great at spotting the tricky, non-linear relationships between the drug dose and the cell's reaction.
For explaining the "Why" (Interpretability): The Cox Model (the accountant) won. While it wasn't the absolute best at guessing, it gave clear, stable numbers (Hazard Ratios) that tell doctors exactly how much a specific dose changes the risk. This is crucial for doctors who need to understand the mechanism, not just the guess.
The Winner: The study concludes that you need both. You need the detective's high accuracy to find the best candidates, and the accountant's clear explanation to understand why they are good candidates.

The "Explainable AI" Magic (SHAP)

One of the biggest problems with AI is that it's often a "black box"—it gives an answer, but you don't know why. This paper added a special layer called SHAP (Shapley Additive Explanations).

Think of SHAP as a magnifying glass that the AI holds up to its own brain. It allows the researchers to see exactly which ingredients mattered most.

The Finding: The magnifying glass showed that the Dosage (how much drug was given) and Covariate 1 (a specific biological factor) were the most important ingredients.
The Stability Check: They ran the test 100 times to make sure the AI wasn't just getting lucky. They found that the AI consistently pointed to the same ingredients every time, proving the results were stable and reliable.

The Bottom Line

This paper doesn't claim to have cured cancer or released a new drug to the market. Instead, it built a better blueprint for testing drugs.

It shows that if you combine the strict logic of traditional statistics with the pattern-spotting power of modern AI, and then use a "magnifying glass" (XAI) to make sure the AI is telling the truth, you get a much more reliable way to screen new drugs.

In short: They built a digital lab that helps scientists decide which new chemical compounds are worth the money to test in real life, saving time and resources by filtering out the bad ones early using a mix of math and machine learning.

Technical Summary: An Explainable AI Framework Integrating Survival Analysis and Machine Learning for COX Inhibitor Evaluation

Problem Statement
The evaluation of Cyclooxygenase (COX) inhibitors, particularly for inflammation and pain modulation, faces a methodological bottleneck characterized by fragmented analytical approaches and reduced model interpretability. While machine learning (ML) offers superior pattern recognition for complex biochemical datasets, it often lacks the transparency required for clinical decision-making. Conversely, classical survival analysis, specifically the Cox Proportional Hazards (Cox PH) model, provides interpretable hazard ratios but may fail to capture non-linear relationships and complex interactions inherent in biochemical covariates. There is a lack of unified frameworks that jointly benchmark classical survival models against modern ML algorithms within a consistent, explainable pipeline for preclinical drug evaluation.

Methodology
The study proposes an integrated multi-model framework that combines survival analysis with classification-based machine learning, evaluated through a rigorous nested cross-validation (NCV) protocol.

Datasets: The research utilizes three complementary data sources to ensure robustness and methodological assessment:
1. Experimental Biochemical Dataset ( $n=90$ ): Derived from in-house MTT viability assays on three human cancer cell lines (HeLa, HCT-116, HepG2) treated with synthesized benzimidazole-triazole hybrids at varying concentrations and time points.
2. Standardized Reference Dataset ( $n=1000$ ): A synthetically generated dataset (JPDR baseline) based on Weibull baseline hazards and predefined statistical assumptions. This dataset is used exclusively for methodological benchmarking and stability analysis, not biological inference.
3. Harmonized Cohort ( $n=312$ ): A semi-synthetic cohort created by integrating the experimental data with a statistically matched subset of the reference dataset to balance biological realism with statistical power.
Models Evaluated: Six predictive models were systematically benchmarked:
- Survival Models: Cox Proportional Hazards (Cox PH), Cox Elastic-Net (Cox-ML), and Deep Survival Network (Deep-S).
- Classification Models: Logistic Regression, Random Forest (RF), and Support Vector Machine (SVM).
- Ensemble: A weighted average of the base models.
Evaluation Protocol:
- Nested Cross-Validation: A 5-fold outer loop for performance estimation and an inner 5-fold loop for hyperparameter tuning (using grid or random search) to prevent overfitting.
- Metrics: Performance was measured using the Concordance Index (C-index) for survival discrimination, Area Under the ROC Curve (AUC) and Accuracy for classification, and Integrated Brier Score (IBS) for calibration.
- Explainability (XAI): The framework incorporates SHAP (Shapley Additive Explanations) values to quantify feature contributions. A novel SHAP Stability Index (SI) was introduced to assess the consistency of feature importance rankings across 100 bootstrap replicates.
- Diagnostics: Cox PH assumptions were validated using Schoenfeld residuals and Martingale plots. Time-dependent concordance statistics (Uno's C) were used for survival models.

Key Contributions

Unified Benchmarking Framework: The study establishes a coherent pipeline for comparing classical survival models and flexible ML classifiers on biochemical data, addressing the gap in prior literature that often treats these paradigms separately.
Multi-Cohort Evaluation Strategy: By utilizing experimental, simulated, and harmonized datasets, the framework allows for controlled methodological assessment while maintaining a link to biological reality.
Explainable AI Integration: The paper introduces a SHAP Stability Index (SI) to quantify the repeatability of feature importance under resampling, filling a gap in explainable survival analysis. It also proposes a residual-interpretation layer to map model residuals to biochemical covariates.
Cross-Task Harmonization: The authors demonstrate a method to convert time-to-event survival predictions into fixed-horizon (365-day) risk probabilities, enabling direct comparison between survival and classification metrics (AUC, Accuracy) on a unified scale.

Results

Predictive Performance:
- Random Forest (RF) consistently achieved the highest discrimination and accuracy across classification tasks, with an AUC of 0.89 and accuracy of 0.85 on the harmonized cohort. It demonstrated superior ability to capture non-linear dose-response relationships.
- Deep-S (Deep Survival) achieved the highest time-dependent discrimination among survival models with a C-index of 0.83 and the lowest Integrated Brier Score (0.142), indicating excellent calibration.
- Cox PH provided a C-index of 0.78 (on the harmonized cohort) and remained the most interpretable model, offering stable hazard ratio estimates.
- Ensemble models slightly improved performance metrics over individual top models, suggesting that aggregation reduces variance.
Feature Importance and Stability:
- SHAP analysis identified Dose and Covariate 1 as the dominant predictors of COX inhibitor efficacy.
- These features exhibited high stability (SI > 0.85) across bootstrap replicates, confirming their robust predictive influence.
- Permutation importance tests corroborated that Dose had the most significant impact on model performance.
Biological Validation:
- Kaplan-Meier curves derived from the experimental dataset showed significant survival differences between treatment groups (log-rank $p < 0.001$ ), particularly for Compound 5 at higher concentrations.
- The model-derived risk stratifications aligned with observed biological separation, validating the framework's ability to capture meaningful outcome structures.

Significance and Claims
The paper claims that its primary significance lies in providing a transparent, reproducible, and explainable pipeline for the early-stage evaluation of drug candidates. By balancing the predictive power of flexible ML algorithms (like Random Forest and Deep-S) with the mechanistic interpretability of classical survival models (Cox PH), the framework supports evidence-based decisions in drug discovery.

The authors emphasize that while ML models offer superior discrimination, the Cox PH model remains indispensable for generating hazard ratios that provide direct mechanistic insights for clinical and pharmacological decision-making. The study positions its framework not as a replacement for experimental validation, but as a computational tool to prioritize candidates and optimize resource allocation in preclinical research. The work concludes that integrating these paradigms accelerates translational research by offering a rigorous method to assess compound efficacy while maintaining the interpretability required for biomedical applications.

Note: The authors explicitly state that the results are hypothesis-generating based on in silico simulations and initial viability tests, and that pharmacological efficacy requires further validation through orthogonal experiments (e.g., enzyme inhibition, in vivo confirmation).