E-GuARD: Expert-Guided Augmentation for Robust Interference Compound Detection

Vincenzo Palmacci, Yasmine Nahal, Matthias Welsch, Ola Engkvist, Samuel Kaski, Johannes KirchmairResearch Square 2025 · DOI

Abstract

Assay interference caused by small organic compounds continues to pose formidable challenges to early drug discovery. Various computational methods have been developed to identify compounds likely to cause assay interference. However, due to the scarcity of data available for model development, the predictive accuracy and applicability of these approaches are limited. In this work, we present E-GuARD (Expert-Guided Augmentation for Robust Interference Compound Detection), a novel framework seeking to address data scarcity and imbalance by integrating self-distillation, active learning, and expert-guided molecular generation. E-GuARD iteratively enriches the training data with interference-relevant molecules, resulting in quantitative structure-interference relationship (QSIR) models with superior performance. We demonstrate the utility of E-GuARD with the examples of four high-quality data sets on thiol reactivity, redox reactivity, nanoluciferase inhibition, and firefly luciferase inhibition. Our models reached MCC values of up to 0.47 for these data sets, with two-fold or higher improvements in enrichment factors compared to models trained without E-GuARD data augmentation. These results highlight the potential of E-GuARD as a scalable solution to mitigating assay interference in early drug discovery.

Introduction

In the early stages of drug discovery, researchers utilize High-Throughput Screening (HTS) to test hundreds of thousands of compounds against biological targets. A persistent challenge in this process is the prevalence of "false positives" caused by assay interference. These "nuisance" compounds can aggregate, react non-specifically with assay components, or interfere with light-based detection methods (such as luciferase inhibition).

While computational models aim to filter these compounds, they are often hindered by a "data desert." High-quality experimental data on interference is scarce, and existing datasets are heavily imbalanced—most compounds are well-behaved, while only a tiny fraction are interferers. This makes it difficult to train robust models that generalize to new chemical spaces.

The E-GuARD Framework

The authors introduce E-GuARD (Expert-Guided Augmentation for Robust Interference Compound Detection), which uses an iterative "teacher-student" loop to expand its own training set through the following steps:

  1. Initial Training: A baseline model (Balanced Random Forest) is trained on existing experimental data.
  2. Goal-Oriented Generation: The framework uses REINVENT4, a reinforcement learning tool, to design new molecules that the current model predicts will cause interference.
  3. Expert-Guided Acquisition: To ensure these new molecules are relevant to drug discovery, the system uses MolSkill. This neural network emulates the preferences of medicinal chemists, prioritizing "drug-like" structures over chemical noise.
  4. Self-Distillation: The most informative and "expert-approved" molecules are added to the training set with "pseudo-labels" from the teacher model. The model is then retrained, and the cycle repeats.

Datasets and Experimental Design

The study utilized four high-quality datasets curated from literature, covering distinct interference mechanisms. As shown in the table below, these datasets initially featured sparse interference rates, ranging from 1.5% to 20%.

Table 1: Composition of the training and test sets for the four interference mechanisms studied.

Key Results and Insights

E-GuARD effectively mitigated class imbalance, reducing the imbalance rate from 0.9 to 0.3 in the most skewed datasets (FI, NI, RR). This created a more balanced training environment, leading to several key performance gains:

  • Performance Boost: The models achieved Matthews Correlation Coefficient (MCC) values up to 0.47, a significant improvement over baseline models.
  • Prioritization Power: The Enrichment Factor (EF)—measuring how much better the model finds interferers than random guessing—saw massive gains. For Firefly Luciferase, the EF increased by 18x; for Nanoluciferase, it increased by 10x.
  • Drug-Likeness: By incorporating MolSkill, the drug-likeness (QED score) of generated interferers for the NI dataset rose from 0.48 to 0.76 over ten iterations, ensuring the synthetic data remained relevant to real-world medicinal chemistry.
  • Mechanism Discovery: The framework successfully "discovered" known interference patterns, such as enriching redox reactivity training sets with PAINS (Pan-Assay Interference Compounds) motifs.

Limitations and Future Work

Despite its success, the authors noted challenges such as generative model mode collapse, where the diversity of generated molecules decreases over time. Additionally, while "pseudo-labeling" is powerful, it can introduce noise that occasionally leads to performance declines in later iterations. Future work may involve more sophisticated handling of label uncertainty or integrating small amounts of real-world experimental validation to "ground" the synthetic data.

Ultimately, E-GuARD demonstrates that well-guided synthetic data can serve as a bridge, filling gaps in chemical space where experimental data is missing and providing a scalable solution for mitigating assay interference.

Summary updated

Abstract

<title>Abstract</title> Assay interference caused by small organic compounds continues to pose formidable challenges to early drug discovery. Various computational methods have been developed to identify compounds likely to cause assay interference. However, due to the scarcity of data available for model development, the predictive accuracy and applicability of these approaches are limited. In this work, we present E-GuARD (Expert-Guided Augmentation for Robust Interference Compound Detection), a novel framework seeking to address data scarcity and imbalance by integrating self-distillation, active learning, and expert-guided molecular generation. E-GuARD iteratively enriches the training data with interference-relevant molecules, resulting in quantitative structure-interference relationship (QSIR) models with superior performance. We demonstrate the utility of E-GuARD with the examples of four high-quality data sets on thiol reactivity, redox reactivity, nanoluciferase inhibition, and firefly luciferase inhibition. Our models reached MCC values of up to 0.47 for these data sets, with two-fold or higher improvements in enrichment factors compared to models trained without E-GuARD data augmentation. These results highlight the potential of E-GuARD as a scalable solution to mitigating assay interference in early drug discovery.
BibTeX
@inproceedings{palmacci2025e,
  author = {Vincenzo Palmacci and Yasmine Nahal and Matthias Welsch and Ola Engkvist and Samuel Kaski and Johannes Kirchmair},
  title = {{E-GuARD: Expert-Guided Augmentation for Robust Interference Compound Detection}},
  booktitle = {Research Square},
  year = {2025},
  doi = {10.21203/rs.3.rs-5740531/v1},
}
Figure