Development and large-scale benchmarks of a protein-ligand absolute binding free energy toolkit

The paper introduces Felis, an open-source, automated toolkit that, when paired with the ByteFF force field, enables high-throughput, zero-shot absolute binding free energy calculations with accuracy comparable to state-of-the-art relative binding free energy methods across diverse protein-ligand datasets.

The Gist

Imagine you are a detective trying to solve a mystery: Which key fits which lock?

In the world of drug discovery, the "lock" is a protein inside your body (often a villain causing disease), and the "keys" are potential new drugs. To cure the disease, you need to find the key that fits the lock perfectly and sticks to it tightly.

For a long time, scientists have had two ways to figure out how well a key fits:

1. The "Relative" Method (RBFE): This is like comparing two keys that look almost identical. You ask, "Does Key A stick better than Key B?" It's fast and accurate, but it only works if the keys are already very similar. If you want to test a completely new shape of key, this method gets stuck.
2. The "Absolute" Method (ABFE): This is the "gold standard." It asks, "Exactly how hard does this specific key stick to the lock?" It can test any key, no matter how weird its shape. However, it's like trying to weigh a feather in a hurricane—it's incredibly slow, expensive, and requires massive computing power.

Enter "Felis": The New Detective Tool

The paper you shared introduces a new toolkit called Felis (created by researchers at ByteDance). Think of Felis as a super-automated, high-speed factory built to do the "Absolute" method (ABFE) for thousands of keys at once, without needing a human to tweak the settings for every single one.

Here is how they did it, using some simple analogies:

1. The "Zero-Shot" Magic

Usually, to get a perfect result with these complex simulations, scientists have to "tune" the engine for every specific lock and key. It's like a chef tasting a soup and adding salt, pepper, or sugar for every single bowl they make.

Felis is different. The researchers built a "universal recipe" (called a force field named ByteFF25). They trained this recipe on a massive amount of data before testing it. When they started the benchmark, they didn't change a single ingredient or adjust the heat for the specific locks they were testing. They just turned the machine on and let it run.

• The Analogy: Imagine a robot chef that has learned to cook millions of dishes perfectly. You hand it a new, weird ingredient it's never seen before, and it cooks it perfectly without asking, "Do you want more salt?" That is what "zero-shot" means here.

2. The "Double-Decoupling" Dance

How does Felis measure the stickiness? It uses a clever trick called Double-Decoupling.

• Step 1: Imagine the key is stuck in the lock. Felis gently "un-sticks" it from the lock and moves it into a bucket of water (simulating the body). It measures how much energy that takes.
• Step 2: Then, it takes a fresh key and moves it from the water into the lock. It measures that energy too.
• The Result: By comparing these two steps, it calculates exactly how much the key wants to be in the lock.

3. The "Traffic Jam" Solution

Running these simulations usually requires thousands of computers talking to each other, which causes traffic jams (data bottlenecks).

• Felis's Fix: They designed the system so that each computer (GPU) works on its own little section of the problem without needing to chat with its neighbors constantly. They overlap the sections slightly so the work flows smoothly, but they don't stop to talk.
• The Analogy: Instead of a long line of cars waiting at a single toll booth, Felis built a highway with many lanes where cars can speed along independently, only merging briefly when necessary. This makes the whole process much faster.

The Big Test: Did It Work?

The researchers put Felis to the test in two major ways:

1. The "Marathon" (43 Proteins, 859 Keys): They tested Felis against a massive dataset of 43 different protein locks and 859 different drug keys.

   • The Result: Felis performed just as well as the current "champion" methods (which take much longer to run). It successfully ranked the keys from "best fit" to "worst fit" with high accuracy.
   • The Surprise: It did this in a fraction of the time (3 nanoseconds of simulation vs. the usual 20 nanoseconds) and without any human tweaking.

2. The "Boss Battle" (KRAS Cancer Protein): They tested Felis on a notoriously difficult cancer target called KRAS(G12D). This is a "super-villain" lock that is shallow, messy, and full of electric charges that confuse standard computers.

   • The Result: Felis handled the chaos beautifully. Even with highly charged keys and a tricky environment, it found the right answers, proving it can handle the hardest real-world drug discovery problems.

Why Does This Matter?

For years, the "Absolute" method (testing any key from scratch) was too slow and expensive to use for finding new drugs. Scientists mostly stuck to the "Relative" method, which limits them to tweaking existing drugs.

Felis changes the game. It proves that we can now use the "gold standard" method for high-speed drug discovery. It's like upgrading from a hand-cranked calculator to a supercomputer. This means scientists can now screen thousands of completely new, weird-shaped drug candidates to find cures for diseases that were previously too difficult to tackle.

In short: The paper presents a new, automated, and incredibly fast tool that can predict how well drugs stick to disease-causing proteins, doing it better and faster than ever before, without needing a human to babysit the process.

Technical Summary

1. Problem Statement

While Relative Binding Free Energy (RBFE) calculations are the industry standard for lead optimization due to their high accuracy and efficiency when ligands share a common scaffold, they struggle with scaffold hopping (large structural changes between ligands). Absolute Binding Free Energy (ABFE) calculations offer a theoretically rigorous alternative that scores ligands independently without scaffold constraints, making them ideal for early-stage hit discovery.

However, the widespread adoption of ABFE in high-throughput drug discovery has been hindered by:

• High Computational Costs: ABFE requires extensive sampling to decouple ligands from solvent and couple them to the protein.
• Lack of Large-Scale Validation: Most existing ABFE benchmarks are small (often <10 targets) compared to RBFE benchmarks, making it difficult to assess robustness across diverse chemical spaces.
• Workflow Complexity: Routine production requires automated pipelines, robust diagnostics, and minimal manual intervention (e.g., force-field tuning), which many current tools lack.

2. Methodology: The Felis Toolkit

The authors present Felis (Free Energy of Ligand-protein InteractionS), an open-source, automated, and scalable toolkit designed for high-throughput ABFE.

• Force Fields:
  • Proteins: AMBER ff14SB.
  • Ligands/Cofactors: ByteFF25, a data-driven molecular mechanics force field. Unlike previous versions, ByteFF25 retains GAFF2 nonbonded parameters (charges and vdW) but features bonded parameters trained on a larger quantum-chemistry dataset to improve chemical space coverage.
• Thermodynamic Cycle (Double Annihilation):
  • Follows the standard double-decoupling protocol: decoupling the ligand from solvent and coupling it to the protein binding site.
  • Restraints: Uses Boresch-style restraints (distance, angle, and dihedral) between protein and ligand anchor atoms to maintain the binding pose during decoupling. Anchor atoms are automatically selected based on secondary structure stability (via DSSP) and interaction frequency (via ProLIF).
  • Alchemical Transformation:
    • Electrostatics: Partial charges are linearly scaled to zero.
    • Van der Waals: Uses a softcore potential (Beutler form) to avoid singularities.
    • Charge Neutrality: For charged ligands, "alchemical waters" are introduced to offset fractional charges during the transformation.
• Simulation Protocol:
  • Engine: OpenMM with openmmtools.
  • Hardware Efficiency: Implements a no cross-GPU communication design. The alchemical path is partitioned into overlapping segments assigned to individual GPUs. This eliminates synchronization bottlenecks and allows for elastic cloud execution.
  • Sampling: Uses 3 ns or 5 ns protocols with three independent replicas per ligand.
  • Analysis: Free energy differences are estimated using the Multistate Bennett Acceptance Ratio (MBAR).

3. Key Contributions

1. Large-Scale Benchmarking: The first extensive ABFE benchmark covering 43 protein targets and 859 unique ligands, derived from the Schrödinger FEP+ public dataset. This scale is comparable to major RBFE benchmarks.
2. Zero-Shot Validation: All predictions were generated in a strict "zero-shot" manner. No force-field parameters were tuned, and no alchemical schedules were customized for the specific benchmark systems. This demonstrates the generalizability of the toolkit.
3. Performance Parity with RBFE: Demonstrated that ABFE can achieve ranking performance comparable to state-of-the-art RBFE methods (FEP+ with OPLS4), despite the inherent challenges of absolute calculations.
4. Handling of Challenging Systems: Successfully applied the toolkit to a difficult KRAS(G12D) dataset featuring highly charged ligands (dicationic) and complex cofactors (GDP/Mg²⁺), proving robustness in electrostatically demanding environments.

4. Results

• Ranking Performance:
  • Across 43 targets, Felis-ABFE achieved Spearman's $\rho$ and Kendall's $\tau$ values comparable to FEP+ RBFE (20 ns simulations).
  • While FEP+ RBFE showed a slight edge in the cumulative distribution of high-ranking systems, Felis-ABFE (using only 3–5 ns per replica) performed remarkably well, validating that shorter simulations with multiple replicas can suppress statistical noise effectively.
• Accuracy Metrics:
  • Debiased Mean Absolute Error (MAE) and Root Mean Square Error (RMSE) for Felis were close to those of FEP+ RBFE.
  • KRAS(G12D) Case Study: For the highly charged MRTX1133 series, 10 ns simulations yielded a Spearman's $\rho$ of 0.76 and Kendall's $\tau$ of 0.64, demonstrating robust convergence and ranking capability even with complex electrostatics.
• Force Field Sensitivity:
  • Comparisons between AM1-BCC and ABCG2 charge models (both trained on ByteFF25) showed no statistically significant difference in ranking performance, suggesting the bonded parameters and overall protocol are the dominant factors.
• Failure Cases & Analysis:
  • hc_bace_2: Both methods failed due to an extremely narrow experimental dynamic range (0.4 kcal/mol for 8 ligands), making ranking statistically meaningless.
  • scyt_dehyd: Felis initially struggled (poor ranking) due to crystallographic water sampling issues. When crystallographic waters were removed, performance improved significantly ($\rho \approx 0.9$), highlighting the sensitivity of ABFE to water network definitions.

5. Significance and Future Outlook

• Viability for Hit Discovery: The study establishes Felis as a viable, "ready-to-use" engine for computational structure-based drug design, capable of prioritizing hits without scaffold constraints.
• Cost Efficiency: By achieving RBFE-level rankings with significantly shorter simulation times (3–5 ns vs. 20 ns) and automated workflows, Felis lowers the barrier for ABFE adoption.
• Future Directions:
  • Pose Accuracy: The authors note that errors in initial binding poses and protonation states are currently the primary bottleneck, more so than sampling convergence.
  • Force Field Refinement: Future work will focus on optimizing nonbonded parameters (charges and vdW) against hydration free energies and incorporating polarizability or virtual sites.
  • Enhanced Sampling: Integration of techniques like Convergence-Adaptive Roundtrip (CAR) and multiple-time stepping (MTS) is planned to further reduce computational costs and improve convergence.

In conclusion, this paper provides a critical step toward democratizing ABFE, proving that with robust automation and modern force fields, absolute binding free energy calculations can be a practical, high-throughput tool for drug discovery.