The Last Mile Problem: A Critical Assessment of Physics-Based and AI Tools for Small Molecule Binding Prediction in...

Area of Science:

• Computational chemistry
• Drug discovery
• Bioinformatics

Background:

• Docking-based virtual screening (VS) is crucial for identifying drug candidates but suffers from high false-positive rates.
• Alchemical free-energy methods, like absolute binding free energy (ABFE) calculations, can improve VS accuracy but are computationally expensive.

Purpose of the Study:

• To benchmark the effectiveness of alchemical ABFE calculations in virtual screening (VS) workflows.
• To compare ABFE with other physics-based and machine-learning (ML) methods for hit identification.

Main Methods:

• A large-scale benchmark using two datasets: PDBbind (632 complexes) for quantitative accuracy and DUD-E (315 binders/decoys) for VS predictive power.
• Evaluation of alchemical ABFE, end-state physics-based methods, and five ML models (including GNINA and Boltz-2).

Main Results:

• Alchemical ABFE performed well across both benchmarks, showing good quantitative accuracy and predictive power.
• End-state methods excelled at distinguishing actives from decoys in DUD-E but lacked correlation with experimental data in PDBbind.
• ML models generally performed well on PDBbind but struggled with DUD-E, with GNINA and Boltz-2 showing better generalization.

Conclusions:

• A staged VS approach, using Boltz-2 as an initial filter followed by alchemical ABFE, can robustly and cost-efficiently enrich hit lists with true actives.
• This combined strategy optimizes the use of resource-intensive ABFE calculations in drug discovery pipelines.