KANEL: Kolmogorov-Arnold Network Ensemble Learning Enables Early Hit Enrichment in High-Throughput Virtual Screening

The paper introduces KANEL, an ensemble learning workflow that integrates interpretable Kolmogorov-Arnold Networks with traditional machine learning models and diverse molecular descriptors to enhance early hit enrichment in high-throughput virtual screening.

The Gist

Imagine you are a treasure hunter trying to find a single, rare diamond hidden inside a massive warehouse filled with 78 billion rocks. You can't examine every single rock; you only have time to pick the top 128 rocks to take to the lab for testing.

Your goal isn't just to have a "good average" guess at where the diamonds are; your goal is to make sure that the very first 128 rocks you pick are almost certainly diamonds. If you pick 128 rocks and only find one diamond, you've wasted your time and money.

This is the problem scientists face in Virtual Screening (finding new medicines). They have huge libraries of chemical compounds, but they can only test a tiny fraction in the real world.

The Solution: KANEL (The "Super-Team" Approach)

The paper introduces a new method called KANEL. Think of KANEL not as a single genius detective, but as a specialized task force working together.

Here is how it works, broken down into simple concepts:

1. The Problem with "Average" Scores

Traditionally, scientists judged their computer models by how well they ranked the entire warehouse of rocks. They used a score called AUC (Area Under the Curve), which is like asking, "Did you sort the whole warehouse correctly?"

• The Flaw: You could sort the bottom 99% of the warehouse perfectly, but if you missed the diamond in the top 128, you failed the mission.
• The Fix: KANEL focuses entirely on PPV@128. This is a score that asks: "Out of the top 128 rocks you picked, how many were actually diamonds?" This is the only metric that matters when you have a limited budget for lab tests.

2. The Team Members (The Ensemble)

Instead of relying on one super-smart AI, KANEL builds a team of different experts, each looking at the rocks from a different angle:

• The Descriptors (The Eye): Some experts look at the chemical shape (like RDKit). Others look at specific chemical properties (LillyMol).
• The Fingerprints (The Pattern): Others look at the molecular "barcode" (Morgan fingerprints), which is like scanning a unique pattern of lines on the rock.
• The Models (The Brains):
  • XGBoost & Random Forest: These are the "veteran detectives" who have solved thousands of cases. They are reliable and fast.
  • MLP (Neural Networks): These are the "pattern recognizers" that are good at finding complex, non-linear connections.
  • KANs (Kolmogorov-Arnold Networks): These are the new kids on the block. Think of them as "transparent detectives." Unlike other AI models that are "black boxes" (you don't know how they decided), KANs show their work. They can explain why they think a rock is a diamond. They are added to the team for their unique perspective and transparency.

3. The Strategy: "Specialists" vs. "Generalists"

The researchers tested two ways to organize the team:

• The Generalist: Feed all the information (shapes, barcodes, properties) into one giant brain.
• The Specialists: Train one expert on shapes, another on barcodes, and another on properties. Then, let them vote.
• The Result: The Specialists won. Just like a medical team where a cardiologist, a neurologist, and a surgeon all weigh in, the "Specialist Ensemble" found more diamonds than the single "Generalist" model.

4. The Magic of "Weighted Voting"

How does the team decide? They don't just take a simple average. They use a smart system (Optuna) to figure out who is the most trustworthy expert for the specific job.

• If the "Fingerprint Expert" is usually right, the team listens to them more.
• If the "Shape Expert" is usually wrong on this specific dataset, they listen to them less.
• The Outcome: This "Weighted Ensemble" consistently found 9% to 40% more active drugs in the top 128 picks compared to the best single model.

5. Did They Cheat? (Y-Randomization)

To make sure the team wasn't just guessing or memorizing the answers, the researchers scrambled the labels (telling the AI which rocks were diamonds and which were rocks randomly).

• The Result: The team's performance crashed. This proves they were actually learning the real rules of chemistry, not just getting lucky.

The Big Picture Takeaway

KANEL is a new, highly efficient workflow for drug discovery.

• It's Practical: It focuses on the top 128 results, which matches how real-world labs work (using 384-well plates).
• It's Diverse: It combines old-school reliable methods with new, transparent AI (KANs).
• It's Better: By letting different experts specialize and then voting together, it finds more potential medicines faster than any single method could.

In short: If you need to find a needle in a haystack, don't just hire one person to look at the whole haystack. Hire a team of specialists, give them different tools, and let them vote on the best spots to dig. That's what KANEL does for finding new drugs.

Technical Summary

1. Problem Statement

The rapid expansion of ultra-large chemical libraries (e.g., Enamine REAL SPACE with >78 billion molecules) has rendered traditional computational methods like molecular docking impractical for virtual screening (VS). While machine learning (ML) offers a computationally efficient alternative, standard evaluation metrics (e.g., AUC, Balanced Accuracy) are often misaligned with the operational reality of experimental drug discovery.

• The Core Challenge: In VS, only a tiny fraction of top-ranked compounds can be experimentally tested. Therefore, a model's ability to rank the entire library correctly is less critical than its ability to enrich the top of the ranked list with true active compounds.
• Metric Misalignment: Global metrics like AUC do not sufficiently capture "early hit enrichment." The authors argue for the use of Positive Predicted Value at N (PPV@N), specifically PPV@128, which aligns with standard experimental screening plate formats (e.g., testing 128 compounds in triplicate on a 384-well plate).

2. Methodology

The authors introduce KANEL, a workflow designed to maximize early hit enrichment through a diverse ensemble approach.

A. Datasets

The study evaluates five highly imbalanced public PubChem BioAssay datasets (AIDs: 485314, 485341, 504466, 624202, 651820), with active fractions ranging from 0.53% to 4.12%.

B. Molecular Representations

Three distinct descriptor families were utilized to ensure feature diversity:

1. LillyMol descriptors: Proprietary molecular descriptors.
2. RDKit-derived descriptors: Standard cheminformatics descriptors.
3. Morgan Fingerprints: Circular fingerprints (256-bit for ensembles, 2048-bit for feature studies).

C. Model Architecture

The ensemble combines four distinct model families trained on the complementary feature sets:

• XGBoost
• Random Forest (RF)
• Multilayer Perceptron (MLP)
• Kolmogorov-Arnold Networks (KANs): Specifically FasterKAN and ReluKAN. KANs are highlighted for their interpretability (learning univariate response functions) and their ability to add diversity to the ensemble.

D. Training and Optimization Protocol

• Splitting: Stratified 80/20 train/test splits repeated five times.
• Hyperparameter Tuning: Optuna was used with 5-fold cross-validation on the training set.
• Optimization Objective: The primary objective was mean PPV@512 during inner-loop optimization, chosen because it positively correlates with early enrichment metrics like PPV@128.
• Ensemble Strategy: The authors compared element-wise product, arithmetic mean, and an Optuna-optimized weighted ensemble applied to predicted class probabilities.

3. Key Contributions

1. KANEL Workflow: A validated ensemble framework that integrates interpretable KANs with established ML models (XGBoost, RF, MLP) and diverse molecular representations.
2. Metric Alignment: A strong argument and demonstration that optimizing for PPV@128 (early enrichment) yields more actionable results for drug discovery than optimizing for global metrics like AUC.
3. Feature vs. Concatenation: Evidence that training specialized models on individual feature sets and fusing them at the prediction level outperforms training a single model on concatenated features.
4. Interpretability Potential: The inclusion of KANs provides a pathway for interpretable ensemble components, a feature often lacking in deep learning ensembles.

4. Key Results

The study demonstrates that the KANEL workflow consistently outperforms the best single model across all five datasets.

• Performance Gains:
  • The Optuna-weighted ensemble improved PPV@128 by 0.06 to 0.12 (absolute) compared to the best single model.
  • This corresponds to a 9% to 40% relative improvement.
  • Example: On AID 624202, PPV@128 increased from 0.36 (best single) to 0.48 (ensemble). On AID 485341, it increased from 0.15 to 0.21 (40% gain).
• Feature Representation:
  • Morgan fingerprints substantially outperformed LillyMol descriptors.
  • Mixing Morgan and LillyMol descriptors yielded modest additional gains over Morgan alone.
• Ensemble vs. Concatenation:
  • On AID 504466, the weighted ensemble of specialized models achieved a PPV@128 of 0.88, whereas an XGBoost model trained on concatenated features achieved 0.83.
• Robustness (Y-Randomization):
  • Y-randomization tests (50% label corruption) resulted in significant performance degradation (e.g., PPV@128 dropping from ~0.82 to ~0.32 for Morgan fingerprints), confirming that the models learned genuine structure-activity relationships rather than spurious correlations.
• Global Metrics:
  • Improvements were not limited to early enrichment; the ensemble also improved global metrics like ROC-AUC and BEDROC.

5. Significance and Future Directions

• Practical Impact: KANEL provides a reliable, interpretable workflow for "hit triage," directly addressing the bottleneck of experimental validation in drug discovery.
• Interpretability: While the current study focuses on predictive performance, the use of KANs suggests a future where ensemble components can be analyzed to understand specific chemical drivers of activity.
• Future Work:
  • Integration of Graph Neural Networks (GNNs) as ensemble components (preliminary tests showed competitive but not superior performance to the current ensemble).
  • Implementation of scaffold-based splits to better assess prospective generalization.
  • Formal interpretability analysis linking learned KAN functions to chemical factors.
  • Prospective testing in experimental drug discovery projects.

In conclusion, KANEL represents a shift from optimizing for general ranking accuracy to optimizing for actionable early enrichment, leveraging the diversity of model architectures and molecular representations to maximize the success rate of experimental screening campaigns.