Planet V2.0 Achieves Enhanced Protein-Ligand Affinity Prediction for Faster Drug Discovery

Predicting how strongly a drug binds to its target protein remains a significant challenge in modern pharmaceutical research, a process that is both lengthy and expensive. Haotian Gao, Xiangying Zhang, Jingyuan Li, and colleagues at Fudan University have addressed this problem with an upgraded version of their protein-ligand affinity prediction model, named PLANET v2.0. This new iteration overcomes limitations in representing protein-ligand interactions by incorporating a Mixture Density Network to more accurately predict binding modes. The researchers trained the model using a multi-objective strategy, effectively modelling the relationship between interaction distance and energy to improve affinity predictions. Benchmarked against established methods like Glide SP and validated on a large commercial dataset, PLANET v2.0 demonstrates enhanced scoring, ranking, and docking capabilities, offering a potentially valuable tool for accelerating drug discovery pipelines and is freely available for use.

Improving Protein-Ligand Contact Map Prediction with PLANET v2.0

Drug discovery is a time-consuming and financially intensive process, and virtual screening offers a means of acceleration. Scoring functions, a key tool within virtual screening, exhibit a precision closely tied to screening efficiency. Previous research resulted in the development of a graph neural network model, PLANET (Protein-Ligand Affinity prediction NETwork), but this model demonstrated limitations in representing protein-ligand contact maps. Accurate prediction of the protein-ligand contact map is crucial for improving inaccurate binding modes and, consequently, poor affinity predictions. This study proposes PLANET v2.0 as an upgraded version of the original model, designed to address these shortcomings and contribute to more efficient and accurate virtual screening workflows.

Protein-Ligand Affinity Prediction via Probability Distributions

Researchers increasingly rely on virtual screening to accelerate drug discovery, a lengthy and expensive undertaking. The team engineered PLANET v2.0, an upgraded version of their previous protein-ligand affinity prediction model, PLANET, to address limitations in representing protein-ligand contact maps. This approach enables the model to better understand the complex relationship between protein and ligand binding, modelling non-covalent interactions with probability density distributions and employing a Gaussian mixture model to describe the relationship between distance and energy.

Experiments employed the CASF-2016 benchmark to rigorously evaluate PLANET v2.0’s performance, assessing its scoring, ranking, and docking power. The system delivers demonstrably improved screening power compared to both the original PLANET and Glide SP, validating its effectiveness on a large-scale commercial dataset. The research harnessed the power of deep learning to create a high-dimensional representation of structure-affinity pairs, moving beyond the limitations of traditional scoring functions. PLANET v2.0 was designed for generalizability, aiming to achieve strong performance across scoring, ranking, docking, and screening power, achieved through careful optimisation of interaction representation. This model distinguishes itself by predicting absolute binding affinities, unlike some contemporary models that only allow for relative comparisons, revealing a significant advancement over existing methods with approximately a 10% improvement in scoring power. By addressing the challenges of limited training data and task preference, the study provides a practical tool freely available at https://www.pdbbind-plus.org.cn/planetv2, poised to enhance virtual screening workflows.

PLANET v2.0 Improves Protein-Ligand Affinity Prediction

Scientists have developed PLANET v2.0, an upgraded protein-ligand affinity prediction model designed to accelerate drug discovery processes. The research addresses limitations in previous models by improving the representation of protein-ligand contact maps, crucial for accurate affinity predictions. Experiments reveal that PLANET v2.0 accurately models the relationship between interaction distances and binding energies using modified Gaussian mixture models, allowing for both favorable and unfavorable interaction representation.

The team measured performance on the CASF-2016 benchmark, demonstrating excellent scoring, ranking, and docking power. Results demonstrate a notable improvement in screening power when compared to both the original PLANET model and Glide SP, a widely used scoring function. Specifically, the model was tested against an ultra-large-scale commercial dataset to validate its robustness and efficiency in a realistic drug discovery context. Data shows that PLANET v2.0 achieves state-of-the-art performance across all benchmarks, indicating its versatility as a comprehensive scoring function. Researchers innovatively replaced unbalanced binary labels of interactions with probability density distributions, significantly enhancing the model’s docking power.

The study details how PLANET v2.0 integrates distance likelihood with empirical scoring functions within a unified MDN framework, enabling comprehensive end-to-end affinity prediction. Measurements confirm that PLANET v2.0’s parameters for the energy function align with those of the distance density function, streamlining the model’s architecture and improving interpretability. This breakthrough delivers a tool capable of predicting protein-ligand affinity, potentially reducing the time and cost associated with identifying promising drug candidates, and is freely available to researchers.

PLANET v2.0 Improves Affinity and Docking Performance

Researchers have developed PLANET v2.0, an updated protein-ligand affinity prediction model building upon their previous work with PLANET v1.0. This new model employs a multi-objective training strategy and utilises Gaussian mixture models to predict both binding modes and the relationship between interaction distance and energy. By directly extracting features from protein and ligand structures, PLANET v2.0 predicts parameters used to define probability density functions, offering prior knowledge for more accurate affinity prediction. Evaluations using the CASF-2016 benchmark and a large-scale commercial dataset demonstrate PLANET v2.0’s improved scoring, ranking, and docking capabilities compared to existing methods like Glide SP and the original PLANET.

The model also shows promise in lead optimisation and virtual screening scenarios, with a notable ability to predict binding conformations with strong interpretability. While performance decreased when tested on the most recent PDBbind v2024 dataset, the authors acknowledge this may be due to differences in the protein structures used, and suggest future models should incorporate the latest data available. PLANET v2.0 is freely available for use and represents a significant advancement towards practical tools for accelerating drug discovery.