SLAB: Simultaneous Labeling And Binding affinity prediction for protein-ligand structures

Aditya Ranganath, Hyo-Jin Kim, Heesung Shim, Jonathan AllenbioRxiv 2025 · DOI

TL;DR

This paper introduces SLAB, a novel machine learning framework that simultaneously predicts both the binding affinity and the active/inactive status of protein-ligand complexes. By jointly training regression and classification tasks, SLAB effectively utilizes all available data, including inactive compounds, leading to improved prediction accuracy and a more robust drug discovery pipeline.

What Problem Does This Paper Address?

Drug discovery is a critical but resource-intensive process. Accurately predicting how strongly a potential drug molecule (ligand) will bind to a target protein is a key step. Traditional methods are slow, and existing machine learning models often struggle with imbalanced datasets. Specifically, many compounds are screened and found to be "inactive" (don't bind), but these valuable data points are often discarded or treated separately from "active" compounds with measured binding affinities. This leads to underutilization of data and can limit the predictive power of models, especially when dealing with skewed datasets where inactive compounds vastly outnumber active ones.

Methods

The core innovation of SLAB is its joint training framework. Instead of training separate models for predicting binding affinity (regression) and classifying a compound as active or inactive (classification), SLAB combines these objectives into a single, unified model.

  • Network Module: The framework utilizes existing graph neural network architectures like EGNN (Equivariant Graph Neural Networks) or SGNN (Spatial Graph Neural Network) to process the 3D structural data of protein-ligand complexes. These networks are adept at capturing complex molecular interactions.
  • Affinity Predictor & Binding Classifier: The output from the network module is fed into two parallel branches: one predicts the binding affinity (regression), and the other predicts the binary active/inactive label (classification).
  • Composite Multi-task Loss: A custom loss function is employed, which is an unweighted sum of the Mean Squared Error (MSE) for the regression task and the Cross-Entropy (CE) loss for the classification task. This encourages the model to learn shared representations that benefit both tasks simultaneously.
  • Inactive Assignment (IA) and Inactive Threshold (IT): The authors also explore strategies to better handle inactive compounds. This includes assigning specific numerical values (IA) to represent inactivity and defining an "Inactive Threshold" (IT) to categorize compounds with low binding affinity as inactive, even if they show some minimal binding.

Datasets

The SLAB framework was evaluated on three datasets:

  • PDBBind (2020 dataset): A comprehensive benchmark dataset containing experimentally measured binding affinities for protein-ligand complexes. This dataset was used for training and evaluating the general performance of the model.
  • Dengue and Zika Flavivirus Protease Datasets: These datasets were constructed using molecular docking simulations and experimental data. They represent more realistic scenarios with a mix of active and inactive compounds, and also include cases with limited or no experimental data for inactive compounds in the test set.

Key Insights and Ideas

  • Data Utilization is Key: The most significant insight is that jointly training regression and classification tasks dramatically improves model performance by leveraging all available data, including inactive compounds. This is crucial for real-world drug discovery where inactive compounds are abundant.
  • Classification as Regularization: Even when a dataset is primarily composed of active binders (like PDBBind), the classification branch acts as a form of regularization, preventing the model from overestimating binding affinities for weaker binders and improving overall robustness.
  • Handling Inactivity is Crucial: The exploration of Inactive Assignment (IA) and Inactive Threshold (IT) strategies highlights the importance of how inactive compounds are represented and handled. These methods, particularly the IT, can significantly improve regression metrics by focusing the model on more relevant data points.
  • Network Architecture Matters, but SLAB is Adaptable: While EGNN and SGNN proved effective network modules, the paper emphasizes that the SLAB framework is adaptable to different network architectures, suggesting its broad applicability.

Key Results

  • Improved Accuracy Across Datasets: SLAB consistently outperformed baseline regression-only models across all three datasets, showing improvements in metrics like RMSE, MAE, r², Pearson (R), and Spearman (ρ).
  • Significant Gains on Flavivirus Data: The most substantial improvements were observed on the Dengue dataset, where the SLAB approach with EGNN achieved a notable increase in performance compared to the EGNN-only model (e.g., RMSE decreased from 1.6527 to 1.4978).
  • Better Handling of Outliers: The use of an inactive threshold (IT) demonstrated a significant improvement in RMSE, indicating better prediction accuracy, especially for weaker binders.
  • Lightweight Integration: The authors note that the SLAB framework adds a minimal computational overhead, augmenting the network architecture by only about 40 floating-point parameters.

Limitations and Future Work

  • Classification Accuracy Trade-off: While the Inactive Threshold (IT) improved regression metrics, there was a noted drop in classification accuracy in some scenarios. The authors suggest that further tuning of IA values might mitigate this.
  • Dataset Imbalance in Zika: The Zika dataset presented a challenge due to a high imbalance between training and testing sets, with all test set examples being classified as active. This limited the evaluation of classification accuracy on this specific dataset.
  • Future Work: The paper suggests exploring other network architectures beyond EGNN and SGNN, and further investigating the impact of different IA and IT strategies for optimizing both regression and classification performance. The adaptability of the framework to various network backbones is a key avenue for future research.

Summary updated

Abstract

Machine learning models are often used as scoring functions to predict the binding affinity of a protein-ligand complex. These models are trained with limited amounts of data with experimentally measured binding affinity values. A large number of compounds are labeled inactive throughs single-concentration screens without measuring binding affinities. These inactive compounds, along with the active ones, can be used to train binary classification models, while regression models are trained using compounds with binding affinities only. However, the classification and regression tasks are often handled separately, without sharing the learned feature representations. In this paper, we propose a novel model architecture that jointly performs regression and classification objectives, aiming to maximize data utilization and improve predictive performance by leveraging two complementary tasks. In our setup, the regression yields the binding affinity, whereas the classification task yields the label as active or inactive. We demonstrate our method using PDBbind, the standard 3D structure database, as well as a dataset of flavivirus protease compounds with binding affinity data. Our experiments show that the new joint training strategy improves the accuracy of the model, increasing applicability in various practical drug screening scenarios.
BibTeX
@misc{ranganath2025slab,
  author = {Aditya Ranganath and Hyo-Jin Kim and Heesung Shim and Jonathan Allen},
  title = {{SLAB: Simultaneous Labeling And Binding affinity prediction for protein-ligand structures}},
  year = {2025},
  doi = {10.1101/2025.06.03.657720},
  howpublished = {bioRxiv},
}
Figure