Abstract
While large library docking has discovered potent ligands for multiple targets, as the libraries have grown, the very top of the hit-lists can become populated with artifacts that cheat our scoring functions. Though these cheating molecules are rare, they become ever-more dominant with library growth. Here, we investigate rescoring top-ranked molecules from docking screens with orthogonal methods to identify these artifacts, exploring implicit solvent models and absolute binding free energy perturbation (AB-FEP) as cross-filters. In retrospective studies, this approach deprioritized high-ranking non-binders for nine targets while leaving true ligands relatively unaffected. We tested the method prospectively against results from large library docking AmpC $\beta$-lactamase. From the very top of the docking hit lists, we prioritized 128 molecules for synthesis and experimental testing, a mixture of 39 molecules that rescoring flagged as likely cheaters and another 89 that were plausible true actives. None of the 39 predicted cheating compounds inhibited AmpC up to $200 \mu M$ in enzyme assays, while 57% of the 89 plausible true actives did do so, with 19 of them inhibiting the enzyme with apparent $K_i$ values better than $50 \mu M$. As our libraries continue to grow, a strategy of catching docking artifacts by rescoring with orthogonal methods may find wide use in the field.
Introduction: The Paradox of Ultra-Large Libraries
Structure-based docking is a cornerstone of modern drug discovery, allowing researchers to computationally "fit" millions of molecules into a protein's binding pocket to predict potential drug candidates. However, as virtual chemical libraries have expanded into the billions, a paradoxical problem has emerged: the "best" scores are getting better, but they are also getting "dirtier."
At the very top of these massive hit lists, a small but dominant population of artifacts or cheaters appears. These molecules do not actually bind well but possess unusual physical features or parameterization errors that "trick" specific scoring functions (such as DOCK3.8). While these errors might occur only once in a million, in a library of a billion molecules, they appear a thousand times, potentially burying true ligands under a mountain of computational noise.
Methodology: Orthogonal Cross-Filtering
To address this, researchers proposed a "cross-filtering" strategy. The core idea is that a "cheater" molecule likely exploits a specific flaw in one scoring function that is unlikely to be present in a different, "orthogonal" function based on different physical principles.
The workflow involves:
- Primary Screen: Initial screening of libraries (up to 1.7 billion molecules) using DOCK3.8.
- Orthogonal Rescoring: Top hits are rescored using three distinct methods:
- FACTS and GBMV: Faster, implicit solvent models that calculate the energy cost of moving a molecule from water into a protein environment.
- AB-FEP (Absolute Binding-Free Energy Perturbation): A computationally intensive "gold standard" method using molecular dynamics to simulate the binding process.
- Statistical Filtering: By plotting primary docking scores against rescoring results and applying a bivariate normal distribution, outliers—the predicted cheaters—can be mathematically identified.
Key Results and Prospective Validation
The effectiveness of this approach was demonstrated through both retrospective analysis of nine protein targets and a rigorous prospective experiment on AmpC $\beta$-lactamase.
In the prospective test, 128 molecules from the top of a 1.7-billion-compound screen were selected for experimental testing. These were divided into a "Cheater" group (39 molecules) and a "Plausible" group (89 molecules) based on the rescoring filters.
The experimental results were striking:
- The Cheater Group: 0% of the 39 flagged molecules showed any activity up to $200 \mu M$.
- The Plausible Group: 57% of the 89 molecules were active inhibitors, with 19 showing high potency ($K_i < 50 \mu M$).
The use of AB-FEP as a filter proved particularly precise, correctly identifying 37 out of 39 cheaters without discarding any true inhibitors.
Insights and Future Directions
A critical observation of this study is the "Divergence" Effect: once a library exceeds approximately 300 million molecules, the distribution of scores starts to diverge, with a "thin wedge" of molecules achieving unnaturally favorable scores. This divergence signals the arrival of artifacts.
The study highlights that "orthogonality" is more important than simply having a "better" scoring function. Even simpler models like FACTS can flag molecules that DOCK3.8 overvalues due to missing desolvation penalties.
While this study focused on DOCK3.8, the problem of "cheating" is likely universal across docking programs. Future work will explore applying these high-level physics filters to larger portions of hit lists and integrating orthogonal checks directly into the initial docking process to streamline the discovery of true, potent ligands in the era of billion-molecule libraries.
Summary updated