This paper introduces Koffee Unbinding Kinetics, a new computational method that rapidly predicts how long drug molecules stay attached to proteins. This approach is significantly faster than current techniques, enabling high-throughput screening for improved drug efficacy and safety in drug discovery.
This paper addresses the challenge of scalable ligand-protein residence time (RT) prediction, which is crucial for mitigating drug discovery program failures. It introduces Koffee™ Unbinding Kinetics, a computational method that reforms the ligand unbinding problem into an energy-resolved optimization process. This approach enables rapid, high-throughput RT simulations across a diverse dataset of ligand-protein complexes.
This paper addresses how to accurately model ligand unbinding by reformulating the problem from a time-resolved simulation into an energy-resolved optimization process. The core aim is to identify low-energy unbinding pathways to describe the ligand dissociation process.
This study performed an internal validation of an unbinding kinetics methodology. It utilized a diverse dataset of 332 ligand-protein complexes across 14 targets, incorporating publicly available sets such as eIF4E, FAK, HSP90, M3, Thermolysin, and TTK, alongside proprietary data. The core method involved Koffee Unbinding Kinetics simulations, executed without manual setup or parameterization. Protein structures were prepared using PDBFixer for missing elements and Protoss for assigning protonation states, while ligand similarity was assessed with iSIM using RDKit fingerprints. Experimental unbinding kinetics data was primarily obtained through surface plasmon resonance experiments.
The study developed the Koffee Unbinding Kinetics simulation algorithm for fast and reliable residence time (RT) predictions. This methodology achieved AUC-ROC values of 0.88 and 0.86 for long-acting and non-transient binder classification tasks, respectively. The method demonstrated strong predictive ranking performance on non-transient binders across three sets (X-ray, Docking, Prospective) with Spearman's rho (ρ) values between 0.54 and 0.67. For accurately identifying non-transient binders, AUC-ROC scores were 0.98 (X-ray), 0.88 (Docking), and 0.83 (Prospective). The unbinding kinetics method showed superior performance compared to ligand molecular weight as a baseline descriptor in five out of six datasets. The simulation methodology is applicable across diverse ligand-protein complexes and is insensitive to ligand molecular weight.
Proprietary Software and Code Availability
The "Koffee™ Unbinding Kinetics" software used in the study is proprietary. Why it matters: This limits reproducibility, as researchers outside the developing company cannot independently verify the results or build upon the specific implementation without access. It also restricts widespread adoption and collaborative development within the broader scientific community.
Limited Public Access to Detailed Data
While some chemical structures and specific dataset compositions are mentioned as "Supporting Information Available", the complete details of all amalgamated FAK and TTK datasets are provided in supporting information that is not directly accessible within the provided text. Why it matters: Incomplete public access to the precise datasets (e.g., specific experimental off-rates used, full ligand structures beyond initial validation) could hinder full validation, independent replication, or comparative studies by other research groups.
Reliance on a Single Simulation per Result
The approach does not require multiple replicas to converge, with each reported result derived from a single simulation. Why it matters: While this accelerates computation, relying on single simulations might not fully capture the stochastic nature and inherent variability of molecular dynamics. This could potentially affect the robustness and statistical confidence of the predicted kinetics, especially for systems with complex energy landscapes or rare events, making it difficult to quantify uncertainty accurately.
This paper introduces a new simulation methodology, Koffee Unbinding Kinetics, which enables rapid and reliable predictions of ligand residence times. This approach offers significantly faster calculations and powerful predictive performance compared to existing molecular dynamics-based techniques. Researchers can leverage this method for high-throughput computational drug discovery workflows.
Many authors are employees or consultants of Kvantify ApS, which developed the commercial software "Koffee™ Unbinding Kinetics" discussed in the paper. Kvantify ApS has also filed for patents related to this technology.