The SAMPL4 host-guest blind prediction challenge: an overview

Area of Science:

• Computational chemistry
• Molecular modeling
• Drug design

Background:

• Accurate prediction of binding affinities is crucial for drug design.
• Supramolecular host-guest systems offer a tractable model for testing computational methods due to their size and flexibility.
• The SAMPL4 challenge aimed to validate computational methods for binding affinity prediction.

Purpose of the Study:

• To prospectively validate computational methods for predicting binding affinities.
• To assess the predictive power of various computational approaches in host-guest systems.
• To establish benchmarks for computational binding affinity prediction in drug discovery.

Main Methods:

• Experimental measurement of binding affinities for 23 guest molecules with cucurbit[7]uril and octa-acid hosts.
• Submission of 35 sets of computational predictions using diverse methods (docking, free energy simulations, quantum mechanics).
• Statistical analysis comparing computational predictions against experimental data and null models.

Main Results:

• Over half of the computational predictions showed better correlation with experimental binding affinities than null models.
• However, most methods exhibited higher root mean squared error and poorer linear regression slopes compared to null models.
• Performance in SAMPL4 was comparable to the previous SAMPL3 challenge, with no single computational approach consistently outperforming others.

Conclusions:

• Current computational methods for binding affinity prediction in host-guest systems show limitations in accuracy, despite improvements in correlation.
• No single method consistently excelled across different hosts, highlighting the need for systematic exploration of energy models and sampling algorithms.
• Further refinement of computational approaches and challenge designs, including addressing salt effects, is necessary for reliable drug design applications.