This paper describes Felis, a software toolkit for running absolute binding free energy (ABFE) calculations. ABFE is a physics-based method that predicts how strongly a small molecule (a ligand) binds to a protein. Unlike relative methods that compare pairs of similar molecules, ABFE scores each ligand on its own. That independence can be useful in early-stage drug discovery, but ABFE has been hard to use at scale because it needs lots of computer time and careful setup.

The authors built Felis to automate and scale ABFE so it can be used in larger benchmarks and, ultimately, in screening campaigns. They paired Felis with ByteFF25, a data-driven force field for drug-like molecules, and ran it on a diverse set of 43 protein targets with 859 ligands. They also tested a harder case, a KRAS(G12D) dataset that includes very highly charged ligands and a charged cofactor. According to the paper excerpt, Felis produced ligand rankings comparable to state-of-the-art relative binding methods, and it showed robust convergence on the challenging KRAS set. All predictions reported were done in a strict “zero-shot” way, meaning the authors did not tune force-field parameters or the alchemical schedules for the test systems.

At a high level, Felis follows a well-known “double-decoupling” protocol. In that approach the ligand is gradually turned off, or decoupled, in water and in the protein binding site, and the differences give the binding free energy. The toolkit uses geometric restraints between the ligand and protein (distance, angle and dihedral terms) and corrects for those restraints analytically or by simulation. It selects anchor atoms for those restraints by analyzing an initial molecular dynamics run and scoring contacts and backbone stability using DSSP (for protein secondary structure) and ProLIF (for specific protein–ligand interactions). For charged ligands, Felis uses “alchemical waters” — selected solvent molecules that change along with the ligand — to keep the simulated systems neutral. The molecular dynamics engine is OpenMM, with common settings such as a 2 femtosecond time step, the Langevin integrator, TIP3P water, and the AMBER ff14SB force field for proteins.