AEV-PLIG for Rapid Binding Affinity Prediction is Now Available on Vecura

What is AEV-PLIG?

AEV-PLIG (Atomic Environment Vector Protein-Ligand Interaction Graphs) is a deep learning-based scoring function designed to predict protein-ligand binding affinity from pre-docked 3D structures in milliseconds. The model combines ANI-2x Atomic Environment Vectors with a five-layer GATv2 graph attention network ensemble. It helps users rapidly estimate how tightly a small molecule binds to a target protein, delivering predictions in under a second. It is especially useful for high-throughput virtual screening of thousands of compounds to prioritize top candidates before investing in expensive experimental or simulation-based testing.

What can users do with AEV-PLIG on Vecura?

With AEV-PLIG on Vecura, users can:

• Compute fast and precise binding affinities ($pK$): Upload a target protein (PDB format) and a docked ligand (SDF format) to get an immediate prediction of their binding strength.

• Run high-throughput ensemble scoring: Leverage a 10-model GATv2 neural network ensemble to run inference in milliseconds per complex.

• Estimate model confidence and uncertainty: Access individual predictions from all ten ensemble members to evaluate consensus and identify out-of-distribution molecules.

• Bypass heavy infrastructure requirements: Screen thousands of compounds seamlessly without configuring GPU environments, package dependencies, or complex deep learning pipelines.

What the output means

The output provides a predicted binding affinity ($pK = -\log_ K_d$ or $K_i$) and a list of 10 individual ensemble predictions. A higher predicted $pK$ represents a stronger binding affinity (e.g., $pK = 6$ corresponds to approximately $1\ \mu\text$ and $pK = 9$ represents approximately $1\ \text$). The standard deviation among the 10 ensemble predictions serves as an uncertainty score, where a low spread implies high model confidence and a high spread indicates potential out-of-distribution inputs.

This output should be used to support scientific decision making. It does not replace experimental validation.

Why this matters

Predicting protein-ligand binding affinity is a cornerstone of structure-based drug discovery. While accurate physics-based methods like Free Energy Perturbation (FEP) exist, they require days of intensive simulation per compound, making them impractical for large-scale virtual screening. Conversely, fast empirical scoring functions often fail to accurately capture the detailed local chemical environments that govern molecular interactions.

AEV-PLIG successfully addresses these limitations by offering a 400,000-fold speedup over FEP while maintaining competitive predictive accuracy. By utilizing ANI-2x Atomic Environment Vectors, the model maps the intricate local chemical environment around 22 ligand and protein atom types within a 5.1 Å cutoff. This combination of deep representation learning and ensemble voting provides robust and rapid prioritization, accelerating hit-to-lead optimization and reducing the cost of early-stage drug development.

• Developed by: Isak Valsson, Oxford Protein Informatics Group (OPIG), University of Oxford

• Source: Official GitHub Repository

• Reference: Original Paper - Communications Chemistry (2025) (or see the ChemRxiv Preprint)