Multi-Agent Drug Discovery from DNA-Encoded Library Screening: An Executable AI4Science Skill

Jin Xu

March 2026

Abstract

We present a fully executable, multi-agent computational pipeline for small-molecule hit identification and compound triage from molecular screening data. Inspired by DNA-Encoded Library (DEL) selection campaigns, this workflow orchestrates four specialized AI agents—Data Engineer, ML Researcher, Computational Chemist, and Paper Writer—under a Chief Scientist coordinator to perform end-to-end virtual drug discovery. Using the MoleculeNet HIV dataset (41,127 compounds, ~3.5% active), our pipeline achieves an AUC-ROC of 0.8095 and an 8.82× enrichment factor in the top-500 predicted actives. After ADMET filtering and multi-objective ranking, we identify 20 drug-like candidates with mean QED of 0.768, mean synthetic accessibility score of 2.83, and 100% Lipinski compliance. Notably, 13 of the top 20 ranked compounds (65%) are confirmed true actives, demonstrating that the composite scoring approach effectively prioritizes genuinely bioactive, drug-like molecules. The entire pipeline is released as a self-contained, reproducible "AI4Science Skill"—a modular computational protocol that can be executed by autonomous agents or human researchers with a single command.

1. Introduction

The discovery of novel small-molecule therapeutics remains one of the most impactful yet resource-intensive endeavors in the life sciences. Traditional high-throughput screening (HTS) campaigns evaluate millions of compounds against biological targets, but suffer from high false-positive rates, limited chemical diversity, and prohibitive costs. DNA-Encoded Library (DEL) technology has emerged as a transformative alternative, enabling the synthesis and selection of libraries containing billions of unique chemical structures through split-and-pool combinatorial chemistry with DNA barcoding (Clark et al., 2009; Goodnow et al., 2017). After affinity selection against a protein target, next-generation sequencing reveals enriched building block combinations, providing a rich dataset for computational hit identification.

However, the analysis of DEL screening data presents significant computational challenges. Raw sequencing counts must be deconvoluted, normalized, and subjected to statistical enrichment analysis. Hit identification requires machine learning models that can learn structure–activity relationships from noisy, imbalanced data. Candidate compounds must then be evaluated for drug-likeness, synthetic accessibility, and ADMET (Absorption, Distribution, Metabolism, Excretion, Toxicity) properties before advancing to synthesis and experimental validation.

In this work, we introduce a multi-agent computational framework that automates the entire post-screening analysis pipeline. Drawing on the paradigm of collaborative AI agents—where each agent possesses domain-specific expertise and communicates through structured interfaces—our system decomposes the drug discovery workflow into four specialized roles:

1. Data Engineer: Processes raw molecular data, validates chemical structures, computes molecular fingerprints, and prepares train/test splits.
2. ML Researcher: Trains and evaluates predictive models for activity classification, identifies top hit candidates with enrichment analysis.
3. Computational Chemist: Computes ADMET properties, applies drug-likeness filters, and ranks compounds using a multi-objective composite score.
4. Paper Writer: Generates structured results reports with publication-ready tables and summary statistics.

A Chief Scientist agent orchestrates the pipeline, enforcing quality gates between phases and making go/no-go decisions based on intermediate results. This architecture mirrors the organizational structure of real-world drug discovery teams, where interdisciplinary collaboration is essential.

We demonstrate this pipeline on the MoleculeNet HIV dataset (Wu et al., 2018), a benchmark collection of 41,127 compounds screened for anti-HIV activity. While our framework is designed for DEL screening data, the MoleculeNet HIV dataset serves as a publicly available, well-characterized surrogate that shares key challenges with DEL data: severe class imbalance (~3.5% active), noisy labels, and the need for multi-property optimization in compound selection.

The complete pipeline is packaged as an executable "AI4Science Skill"—a self-contained computational protocol with defined inputs, outputs, agent roles, and dependencies. This skill can be invoked by autonomous AI systems or human researchers, promoting reproducibility and modularity in computational drug discovery.

2. Methods

2.1 Dataset

We use the MoleculeNet HIV dataset, which contains 41,127 compounds tested for their ability to inhibit HIV replication (Wu et al., 2018). Each compound is represented as a SMILES string with a binary activity label (active/inactive). The dataset exhibits severe class imbalance with approximately 3.5% of compounds classified as active (1,443 actives vs. 39,684 inactives), yielding a class imbalance ratio of 27.5:1.

All 41,127 SMILES strings were successfully parsed by RDKit with zero invalid structures removed, indicating high data quality. The dataset was split into training (32,901 compounds, 1,154 actives) and test (8,226 compounds, 289 actives) sets using stratified random sampling to preserve class balance, with a fixed random seed (42) for reproducibility.

2.2 Molecular Featurization

Each valid molecule was converted to a Morgan circular fingerprint (Extended-Connectivity Fingerprint, ECFP) using RDKit (Landrum, 2016). We employed a radius of 2 (equivalent to ECFP4) with 2,048-bit vectors. Morgan fingerprints encode the local chemical environment around each atom as a set of circular substructures, providing a robust and information-rich representation for structure–activity relationship modeling. The resulting feature matrix has dimensionality 41,127 × 2,048.

2.3 Machine Learning Model

We trained a Random Forest classifier with 100 estimators, chosen for its robustness to overfitting, interpretability, and strong performance on molecular fingerprint data. Key hyperparameters were configured as follows:

• n_estimators: 100
• max_depth: 15
• min_samples_split: 10
• min_samples_leaf: 4
• max_features: sqrt (≈45 features per split)
• class_weight: balanced (inverse frequency weighting to address imbalance)

The balanced class weight setting automatically adjusts sample weights inversely proportional to class frequencies, effectively upweighting the minority active class during training. This is critical for imbalanced datasets where naive classifiers tend to predict the majority class.

2.4 Evaluation Metrics

Model performance was assessed using metrics appropriate for imbalanced classification:

• AUC-ROC: Area under the Receiver Operating Characteristic curve, measuring discrimination ability across all thresholds.
• AUC-PR: Area under the Precision-Recall curve, more informative than AUC-ROC for highly imbalanced datasets.
• Enrichment Factor: The ratio of true actives in the top-K predictions to the expected proportion by random selection, quantifying the model's ability to prioritize actives.
• Precision, Recall, F1 Score: Threshold-dependent metrics at the default 0.5 cutoff.

2.5 ADMET Property Computation

For the top-500 predicted actives (ranked by predicted probability), we computed a comprehensive panel of drug-likeness descriptors using RDKit:

• Molecular Weight (MW): Calculated exact molecular weight.
• LogP: Wildman-Crippen partition coefficient, estimating lipophilicity.
• Hydrogen Bond Donors (HBD) and Acceptors (HBA): Counts of H-bond donor/acceptor groups.
• Topological Polar Surface Area (TPSA): Sum of surfaces of polar atoms, correlating with membrane permeability.
• Rotatable Bonds: Count of freely rotatable single bonds, relating to molecular flexibility and oral bioavailability.
• QED (Quantitative Estimate of Drug-likeness): A composite score (0–1) integrating multiple desirability functions across eight molecular properties (Bickerton et al., 2012). Higher values indicate greater similarity to known oral drugs.
• SA Score (Synthetic Accessibility): A heuristic score (1–10) estimating ease of synthesis based on fragment frequencies and structural complexity (Ertl & Schuffenhauer, 2009). Lower values indicate easier synthesis.
• Lipinski Rule-of-5 Violations: Count of violations of Lipinski's rules (MW ≤ 500, LogP ≤ 5, HBD ≤ 5, HBA ≤ 10), a widely used oral bioavailability heuristic.

2.6 Compound Filtering and Ranking

We applied hard filters to remove clearly problematic compounds:

• Molecular weight: 150–600 Da
• LogP: −2 to 6
• Lipinski violations: ≤ 1
• SA Score: ≤ 6.0
• QED: ≥ 0.2

Compounds passing all filters were ranked by a composite drug-likeness score, computed as a weighted combination:

Composite Score = 0.40 × ML_Score + 0.30 × QED + 0.20 × SA_norm + 0.10 × Lipinski_Score

where SA_norm = (10 − SA_Score) / 9 (normalized and inverted so higher is better), and Lipinski_Score = max(0, 1.0 − 0.25 × violations). This weighting balances predicted bioactivity (40%) with drug-likeness (30%), synthetic feasibility (20%), and oral bioavailability compliance (10%).

3. Results

3.1 Model Performance

The Random Forest classifier achieved the following performance on the held-out test set of 8,226 compounds:

The AUC-ROC of 0.8095 indicates good discriminative ability, significantly above the random baseline of 0.5. The AUC-PR of 0.3655 is notably higher than the random baseline of ~0.035 (the prevalence rate), representing a ~10× improvement and confirming that the model effectively prioritizes active compounds despite the severe class imbalance.

3.2 Hit Enrichment

Among the top 500 compounds ranked by predicted activity probability, 155 were confirmed true actives out of 289 total actives in the test set (53.6% recall at top-500). This yields an enrichment factor of 8.82× compared to random selection, meaning the model is nearly nine times more effective than random screening at identifying active compounds in the top-ranked subset.

3.3 ADMET Filtering

Of the 500 ML-predicted hits, all 500 had valid RDKit-parseable structures for ADMET computation. After applying the drug-likeness filters:

The 36.2% attrition rate is consistent with typical drug discovery campaigns, where a substantial fraction of computationally predicted hits fail basic drug-likeness criteria. The remaining 319 compounds represent a tractable set for further analysis and experimental follow-up.

3.4 Top-20 Drug Candidates

The top 20 compounds, ranked by composite score, exhibit excellent drug-like properties:

Summary statistics for the top 20:

The top-20 candidates demonstrate a strong balance between predicted bioactivity and drug-likeness. The mean QED of 0.768 places these compounds in the upper quartile of drug-likeness (typical marketed oral drugs have QED 0.5–0.9). The mean SA Score of 2.83 indicates high synthetic accessibility (scores < 3 suggest straightforward synthesis). All 20 compounds are fully Lipinski compliant with zero violations, and the mean molecular weight of 334.4 Da is well within the optimal range for oral bioavailability.

Most importantly, 13 out of 20 top-ranked compounds (65%) are confirmed true actives in the original MoleculeNet HIV assay. This represents an 18.6× enrichment over the baseline activity rate of 3.5%, demonstrating that the composite scoring approach effectively integrates ML predictions with drug-likeness criteria to prioritize genuinely bioactive, developable molecules.

3.5 Analysis of False Positives in the Top 20

Seven of the top 20 compounds are not labeled as active in the dataset. However, these compounds exhibit exceptional drug-likeness properties (mean QED 0.803 for the false positives vs. 0.749 for the true positives), suggesting they occupy favorable regions of chemical space. In a real DEL screening scenario, such compounds—predicted active with high drug-likeness—would still merit experimental validation, as false negatives in the original assay are not uncommon, and structurally drug-like scaffolds may serve as starting points for medicinal chemistry optimization.

4. Discussion

4.1 Multi-Agent Architecture

The multi-agent design offers several advantages over monolithic pipelines. First, each agent encapsulates domain-specific logic and can be independently tested, updated, or replaced. For example, the ML Researcher agent could be extended with gradient-boosted trees, graph neural networks, or transfer learning approaches without modifying the Data Engineer or Computational Chemist agents. Second, the Chief Scientist enforces quality gates—such as requiring a minimum number of valid molecules before proceeding—that prevent error propagation across pipeline stages. Third, the structured communication between agents (via files and summary dictionaries) creates a natural audit trail for reproducibility.

4.2 Relevance to DEL Screening

While demonstrated on a public benchmark dataset, this pipeline directly applies to DEL screening workflows. In a DEL campaign, the Data Engineer would process sequencing count data and compute fingerprints for enumerated library members. The ML Researcher would train on enrichment scores rather than binary labels. The Computational Chemist's role remains identical—evaluating synthesizability and drug-likeness of predicted hits before committing to off-DNA synthesis and validation.

Our prior work has demonstrated the real-world applicability of ML-guided DEL hit identification, culminating in the discovery of a first-in-class small-molecule ligand for WDR91 (Xu et al., J. Med. Chem. 2023, 66, 5, 3452–3460). The pipeline presented here codifies that workflow into a reproducible, agent-driven protocol.

4.3 The AI4Science Skill Paradigm

We introduce the concept of an "AI4Science Skill"—a self-contained computational protocol that packages domain knowledge, code, dependencies, and agent roles into a single executable unit. Skills are designed to be:

• Reproducible: Fixed random seeds, explicit dependencies, and versioned code ensure consistent results.
• Modular: Each agent can be independently upgraded or replaced.
• Composable: Skills can be chained together (e.g., a DEL library design skill → a screening analysis skill → a lead optimization skill).
• Executable by AI or humans: The pipeline can be invoked programmatically by autonomous agents or run manually by researchers via a single command.

This paradigm aligns with the broader vision of AI-driven scientific workflows, where complex research tasks are decomposed into well-defined, reusable components that can be orchestrated by both human and artificial intelligence.

4.4 Limitations

Several limitations should be noted. First, our ML model uses Morgan fingerprints, which, while effective, do not capture three-dimensional molecular information. Graph neural networks operating on molecular graphs could potentially improve performance. Second, the ADMET properties computed here are descriptor-based predictions rather than experimental measurements; actual ADMET outcomes may differ. Third, the composite scoring weights (0.4/0.3/0.2/0.1) were selected based on domain heuristics rather than optimization, and different applications may benefit from adjusted weightings. Finally, the MoleculeNet HIV dataset, while valuable as a benchmark, does not fully recapitulate the noise characteristics and combinatorial structure of DEL screening data.

5. Conclusion

We have presented a multi-agent drug discovery pipeline that automates hit identification and compound triage from molecular screening data. The pipeline achieves strong predictive performance (AUC-ROC 0.8095, 8.82× enrichment) and produces a curated set of 20 drug-like candidates with 65% confirmed activity, 100% Lipinski compliance, and favorable synthetic accessibility. Released as an executable AI4Science Skill, this work demonstrates that collaborative AI agents can effectively replicate and automate the interdisciplinary workflows central to early-stage drug discovery. We anticipate that such modular, agent-driven computational protocols will become increasingly important as AI systems take on larger roles in scientific research.

6. Data and Code Availability

The complete pipeline, including all agent scripts, configuration, and the SKILL.md specification, is available as a self-contained AI4Science Skill. The MoleculeNet HIV dataset is publicly available from the DeepChem data repository. All code is implemented in Python using RDKit (Landrum, 2016), scikit-learn (Pedregosa et al., 2011), pandas, and NumPy.

Requirements: rdkit-pypi, scikit-learn, pandas, numpy

Execution: python run_pipeline.py

References

1. Bickerton, G. R., Paolini, G. V., Besnard, J., Muresan, S., & Hopkins, A. L. (2012). Quantifying the chemical beauty of drugs. Nature Chemistry, 4(2), 90–98.

2. Clark, M. A., et al. (2009). Design, synthesis and selection of DNA-encoded small-molecule libraries. Nature Chemical Biology, 5(9), 647–654.

3. Ertl, P., & Schuffenhauer, A. (2009). Estimation of synthetic accessibility score of drug-like molecules based on molecular complexity and fragment contributions. Journal of Cheminformatics, 1(1), 8.

4. Goodnow, R. A., Dumelin, C. E., & Keefe, A. D. (2017). DNA-encoded chemistry: enabling the deeper sampling of chemical space. Nature Reviews Drug Discovery, 16(2), 131–147.

5. Landrum, G. (2016). RDKit: Open-source cheminformatics. http://www.rdkit.org

6. Pedregosa, F., et al. (2011). Scikit-learn: Machine learning in Python. Journal of Machine Learning Research, 12, 2825–2830.

7. Wu, Z., et al. (2018). MoleculeNet: a benchmark for molecular machine learning. Chemical Science, 9(2), 513–530.

8. Xu, J., et al. (2023). Discovery of a First-in-Class Small-Molecule Ligand for WDR91 Using DNA-Encoded Chemical Library Selection and Machine Learning. Journal of Medicinal Chemistry, 66(5), 3452–3460.