Integrating DragonFold with FEP: Advancing AI-Driven Drug Discovery
By Jan Domanski, Ph.D.
Summary
In small-molecule drug discovery, lead optimisation (LO) is a critical phase where promising compounds are fine-tuned for potency, selectivity, and safety. Free Energy Perturbation (FEP) has long been a gold standard for predicting ligand-protein binding affinity, but traditional methods are computationally expensive and error-prone. This blog explores how CHARM Therapeutics is revolutionising this process by integrating DragonFold, our proprietary AI-driven protein-ligand prediction platform, with FEP, improving accuracy while eliminating labour-intensive setup steps, paving the way for faster, more scalable drug discovery.
In early-stage small-molecule drug discovery, lead optimisation (LO) is a crucial phase during which the structures of promising drug-like molecules are optimised to improve potency, selectivity, pharmacokinetics, and safety profiles. One of the leading computational models in LO is Free Energy Perturbation (FEP), a pivotal technique in early-stage drug discovery where the goal is to fine-tune compound candidates to maximise their binding affinity to the therapeutic target protein. The ability to predict the binding affinity between a ligand and its target protein is crucial for designing molecules with enhanced potency and selectivity, and FEP has become an invaluable tool for small-molecule drug discovery.
Although FEP has been a gold standard prediction technique for its accuracy in binding energy predictions and its speed and precision in protein structure prediction, the method is computationally expensive, often requiring labour intensive and error prone steps such as ligand-constrained docking. There is also a lack of research, with limited published work exploring its potential, and no successful attempts at using ligand poses predicted by co-folding models such as DragonFold, CHARM Therapeutics’ proprietary deep learning AI enabled protein-ligand prediction platform, directly for FEP.
In a recent manuscript by CHARM Therapeutics titled: Leveraging Alchemical Free Energy Calculations with Accurate Protein Structure Prediction, we proposed a novel combined approach called co-folding-FEP (CF-FEP), which integrates high-quality DragonFold ligand predictions with FEP, with the aim of improving the accuracy of FEP calculations and completely bypassing the need for error-prone FEP setup steps such as ligand-constrained docking.
FEP calculations typically rely on protein-ligand complex structures, which were traditionally obtained through X-ray crystallography, but our data demonstrated that the combined Dragonfold-FEP protocol was shown to perform at least similarly to ‘traditional’ X-ray-based FEP, which opens the way for high-throughput FEP without time-consuming iterative X-ray experiments to enable the technique. This also demonstrates the utility of co-folding-FEP in drug discovery, showing that it is able to compete with state-of-the-art FEP methods.
We believe that this approach will improve the scalability and efficiency of protein-ligand affinity predictions and provide a more rapid and reliable approach to computational LO support, ultimately accelerating the discovery of novel small-molecule therapeutics. By integrating DragonFold with FEP, CHARM Therapeutics is driving a paradigm shift toward faster, more reliable, and scalable early-stage drug discovery. We are already leveraging this innovation across our internal drug discovery programs and remain committed to pushing the boundaries of computational drug design.
Link to pre-print: https://doi.org/10.26434/chemrxiv-2025-wv6z9-v2
