TL;DR
This paper introduces Convex-PL-R, a novel physics-informed machine learning scoring function for predicting protein-ligand binding affinities. It addresses a common bias in existing methods towards larger binding interfaces and demonstrates improved performance across multiple benchmarks, making it a valuable tool for accelerating drug discovery.
What Problem Does This Paper Address?
Drug discovery pipelines heavily rely on virtual screening to identify promising drug candidates. A crucial step in this process is accurately predicting how strongly a small molecule (ligand) will bind to a target protein. This is often done using "scoring functions" that estimate the binding free energy. However, many existing scoring functions, especially those that are "knowledge-based" (derived from statistical analysis of known protein-ligand complexes), suffer from a significant bias: they tend to favor ligands that interact with larger surface areas of the protein. This can lead to false positives, where larger, less specific interactions are incorrectly prioritized over more subtle, but potentially more potent, binding events. Furthermore, the underlying heuristics in many of these functions are not easily interpretable, making it difficult to understand why a particular prediction is made. This paper aims to overcome this bias and improve the generalizability and interpretability of scoring functions.
Methods
The authors introduce Convex-PL-R, an evolution of their previous Convex-PL model. The core idea is to build a physics-informed scoring function that incorporates thermodynamic principles.
- Physics-Informed Foundation: The model starts from a derivation of the standard binding free energy equation, which includes terms for potential energy, solvation, and entropy.
- Addressing the Bias: To counteract the bias towards larger interfaces, they introduce additional physics-based terms, particularly focusing on entropic contributions.
- Regression Model: While the original Convex-PL was trained for pose prediction (classifying correct vs. incorrect ligand poses), Convex-PL-R is trained using a regression approach. This allows it to predict absolute binding affinities by learning from available experimental binding data.
- Feature Engineering: The model utilizes descriptors related to solvent-accessible surface areas and distance distributions between grid points representing solvent and complex atoms.
- Model Retraining and Augmentation: Convex-PL-R is retrained on a more diverse dataset that includes cofactors and modified residues, and a smaller cutoff radius for pairwise interactions is used to reduce variance.
- Docking Protocol: They also developed a novel docking protocol by integrating the Convex-PL scoring function into AutoDock Vina (named VinaCPL) for more efficient and accurate docking.
Datasets
The authors utilized several datasets for training and evaluation:
- Training Data:
- PDBBind 2019
- Binding MOAD 2020
- A custom set of complexes from PDB containing cofactors or modified residues.
- A subset of PDBBind 2019 specifically chosen for regression training.
- Evaluation Benchmarks:
- CASF Benchmarks (2013 and 2016): Standard benchmarks for evaluating scoring functions, comprising tests for pose prediction, scoring, ranking, and virtual screening.
- D3R Grand Challenges (2, 3, and 4): Datasets from challenges focused on blind prediction of protein-ligand poses and affinities.
- Directory of Useful Decoys (DUD): A dataset specifically designed for virtual screening benchmarking, containing targets with active compounds and their corresponding decoys.
- LIT-PCBA: A virtual screening benchmark with a large number of active and inactive compounds.
Key Insights and Ideas
- Bias is a Major Hurdle: The analysis of existing scoring functions on CASF benchmarks clearly showed a pervasive bias towards larger binding interfaces, significantly impacting their real-world utility in virtual screening.
- Entropy Matters: Incorporating entropic terms, derived from thermodynamic principles, was crucial for penalizing this bias and improving performance.
- Regression for Affinity: Shifting from a classification-based approach (pose prediction) to a regression-based approach (affinity prediction) using experimental data was key to improving the prediction of absolute binding affinities.
- Dataset Diversity is Crucial: Retraining on a more diverse dataset, including cofactors and modified residues, enhanced the model's ability to handle a wider range of protein-ligand interactions.
- Smaller Interaction Cutoffs Help: Reducing the cutoff radius for pairwise interactions in the scoring function also contributed to mitigating the bias.
Key Results
- Improved Performance on CASF: Convex-PL-R demonstrated good results on CASF benchmarks, often ranking among the top performers, and showing substantial improvement over its predecessor, Convex-PL.
- D3R Challenges: Convex-PL-R and its variant Convex-PLcof outperformed their predecessor in pose prediction on D3R Grand Challenges 3 and 4.
- DUD and LIT-PCBA: On the DUD dataset, Convex-PL-R and Convex-PLcof showed considerable improvement over Convex-PL and outperformed AutoDock Vina and Vinardo in early enrichment metrics. In the LIT-PCBA benchmark, Convex-PL-R outperformed AutoDock Vina.
- Ranking: Convex-PL-R achieved top-2 to top-3 rankings in the majority of CASF tests, and top-2 in DUD and D3R challenges.
- Bias Mitigation: Histograms showed that Convex-PL-R, compared to earlier versions, better predicted ligands with SASA and atom counts closer to those of true binders, indicating a successful reduction in the bias towards larger interfaces.
Limitations and Future Work
- Metal Ion Support: While the dataset included cofactors, the model's support for metal ions (like zinc) is still not fully robust, as indicated by outliers in the CASF benchmarks.
- Outliers and Complex Cases: Certain challenging complexes, particularly those involving domain swapping or unusual binding site reconstructions, still presented difficulties for Convex-PL-R and other scoring functions.
- DUD Benchmark Limitations: The authors noted potential issues with the DUD benchmark itself, including unproven inactivity of decoys and biases in decoy generation, which might affect performance evaluations.
- LIT-PCBA Performance: While Convex-PL-R outperformed AutoDock Vina on LIT-PCBA, the overall ROC AUC and enrichment factors were generally low, possibly due to specific target characteristics within the benchmark.
- Future Work: The authors suggest further development in handling metal ions and exploring more sophisticated solvation models. The integration of Convex-PL-R into docking pipelines like VinaCPL is a key step towards its practical application.
Summary updated