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

arXiv:2603.22274 · physics.comp-ph · Submitted 2026-03-23 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Physics. The day in general physics: instruments, methods and results across the field.

Kai: We're looking at the paper: "Development and large-scale benchmarks of a protein--ligand absolute binding free energy toolkit".

Mira: Here's the gist of the paper. Absolute binding free energy (ABFE) calculations offer a theoretically rigorous approach for predicting protein–ligand binding affinities without the scaffold constraints of relative binding free energy (RBFE) perturbations.

Kai: Here, we present Felis, an open-source, automated, and scalable toolkit designed for high-throughput ABFE calculations <ref:2603.22274#pg2> Paired with ByteFF, a previously developed data-driven molecular mechanics force field for drug-like molecules, Felis achieves ranking performance comparable to state-of-the-art RBFE methods on a diverse dataset comprising 43 protein targets and 857 ligands <ref:2603.22274#pg2> Furthermore, we demonstrate robust convergence and ranking performance of Felis on a more challenging KRAS(G12D) dataset, where some ligands and the cofactor are highly charged <ref:2603.22274#pg2>.

Mira: Crucially, all Felis predictions in this study were generated in a strict zero-shot manner, eschewing custom force-field modifications and alchemical schedule fine-tuning <ref:2603.22274#pg2>.

Kai: Toolkit Development and Methodology Felis was developed as an automated ABFE toolkit to address the high computational costs and lack of large-scale validation hindering the broad adoption of ABFE in high-throughput hit discovery campaigns <ref:2603.22274#pg2> The calculations follow the standard double-decoupling (annihilation) protocol where the ligand is decoupled from explicit solvent or protein environment <ref:2603.22274#pg2> Restraints for protein–ligand alchemical simulations are defined using three protein anchor atoms and three ligand anchor atoms, including distance, angle, and dihedral terms <ref:2603.22274#pg2> The double annihilation protocol involves linearly scaling the partial charge parameters of the ligand to zero while steric and bonded terms remain fully coupled <ref:2603.22274#pg2> Simulation protocols utilize OpenMM with LangevinMiddleIntegrator, maintaining temperature and pressure at 298.15 K and 1.013 bar <ref:2603.22274#pg2>.

Title and authors: Kai: So we're looking at this paper now titled "Development and large-scale benchmarks of a protein--ligand absolute binding free energy toolkit". It’s about making absolute binding free energy calculations possible without the scaffold constraints that usually limit them.

Mira: Right, so the authors are introducing Felis, which they describe as an open source, automated tool designed for high throughput calculations. They paired it with a force field called ByteFF for the ligands and proteins.

Lev: Automating these kinds of calculations is key because they are usually computationally expensive and lack large scale validation. This paper seems to be addressing that directly by creating this toolkit for hit discovery campaigns.

Kai: The main thing I see from the title is that they’re focusing on absolute binding free energy, which sounds really rigorous compared to relative binding free energy methods we've used before.

Mira: Exactly, and the summary explains that Felis achieves ranking performance comparable to state-of-the-art relative binding free energy methods on a dataset of forty-three protein targets and eight hundred fifty-seven ligands <ref:2603.22274#pg1,Felis achieves ranking performance comparable to state-of-the-art>.

Lev: That’s a big number of data points for them to work with, which is important for validating any new method. They also tested it on a really challenging set called KRAS(G12D), where things get complicated because some ligands and the cofactor are highly charged.

Kai: And what I find interesting from the description is that they didn't mess around with custom force field modifications or alchemical schedule fine tuning for these predictions.

Mira: That zero shot approach is a big deal, because it means you don't have to worry about those messy parameter tweaks when you are trying to get a result quickly. They used AMBER ff14SB for proteins and ByteFF25 for the molecules.

Lev: ByteFF25 is interesting because they say its bonded parameters were trained on a more extensive quantum chemistry dataset, which should help with coverage across different chemical spaces. They also found that comparing the AM1-BCC and ABCG2 charge models didn't actually make a statistically significant difference in ranking performance.

Title and authors: Kai: So, when they compare it to state of the art relative binding free energy methods, they find that Felis achieves a ranking performance comparable to FEP plus relative binding free energy with substantially shorter simulations.

Mira: That means you can get good rankings in less time, like using three independent replicas for just three nanoseconds or five nanoseconds instead of twenty nanoseconds. They also showed that the ranking performance of absolute binding free energy is comparable to, though not better than, state-of-the-art relative binding free energy predictions.

Lev: That’s a solid result when you consider the computational cost savings, because for real hardware running on things like qubits or even just large CPU clusters, those simulation times are significant.

Kai: They also touched on the KRAS(G12D) dataset and found that at ten nanoseconds simulation length, Felis-ABFE achieved Spearman’s rho of zero point seven six and Kendall’s tau of zero point six four, which they say demonstrates robust consistency with experimental data for this series.

Mira: But they also noted a caveat there; the predicted binding energies for that KRAS(G12D) series show a systematic offset relative to experiment, which they attribute to the protein reorganization free energy between apo and holo states shifting the absolute baseline.

Lev: That’s a crucial point because it tells us that even with absolute methods, you have to consider those underlying protein changes when interpreting the final numbers. It’s not just about fitting a curve; it’s about understanding that shift.

Kai: So, what they suggest for improvement is focusing on more reliable binding poses and robust protonation state assignment for real world adoption.

Mira: They also plan to integrate an explicit estimate of protein reorganization free energy into Felis so they can assess whether that global shift is actually resolved or if it still exists. That sounds like a logical next step to build on this work.

Lev: I think co optimizing the charge model and van der Waals parameters against hydration free energies and condensed phase liquid properties would also be a good area for future work, because those are things that affect the absolute energy calculation significantly in real conditions.

Title and authors: Kai: So, to wrap up on the "Development and large-scale benchmarks of a protein--ligand absolute binding free energy toolkit," they've built something open source and automated that works well across many targets with good rankings compared to existing methods, but they acknowledge that the absolute baseline shift from protein reorganization is still a factor.

Mira: Basically, it shows you can get accurate rankings without needing those scaffold constraints, which is very useful for hit discovery when you don't have a specific scaffold in mind.

Lev: And for real hardware implementation, the main thing to remember from this paper is that while the scoring method is fast and robust on benchmarks like KRAS(G12D), you still need to account for those protein reorganization energy terms when translating those scores into absolute binding energies.

Kai: That makes sense. It’s a powerful tool for filtering hits quickly, but you need to be aware of the limitations when moving from a benchmark set to an actual biological system.

Mira: Exactly, so this paper gives us a way to get consistent rankings by scoring each ligand independently based on its fit, even if the absolute energy is slightly off due to protein dynamics.

Lev: And for me, from where I sit building error correction hardware, the point is that if we want to run Felis on real hardware efficiently, we need those simulation times to be manageable and our error models have to account for that systematic offset they mentioned.

Kai: So the big picture is a scalable tool that gets us past some of those high computational cost hurdles in protein docking by focusing on absolute energy.

Mira: This paper provides a very useful framework for moving beyond relative methods when you want to assess the true strength of a protein ligand interaction across different chemical classes.

Lev: We’ll see what they do next when they integrate that explicit reorganization energy term, because that's where I think the real test for its utility will be.

The paper's summary: Kai: So, to recap, they’ve built this tool called Felis that calculates absolute binding free energy without needing those complicated relative binding free energy constraints we usually have to deal with in drug discovery.

Mira: Right, so they’re using an open source toolkit and pairing it with a specific force field called ByteFF for the molecules and AMBER ff14SB for the proteins.

Lev: From my side, I see them trying to make this fast enough to run on real hardware without needing those super long simulation times we usually have to settle for in FEP calculations.

Kai: They tested it across forty-three different protein targets and over eight hundred ligands, which is a pretty big dataset for showing off performance.

Mira: And the results show that Felis can rank these compounds similarly to state-of-the-art relative methods, even though the underlying method is absolute binding free energy.

Lev: That ranking comparison is interesting because it means if you’re filtering hits on a huge library, this tool gives you a reliable ordering without needing to run every single one through a full, expensive simulation.

Kai: They also showed it holds up on some really tough cases, like that KRAS G12D dataset where the ligands and the protein environment are super charged.

Mira: That’s important because handling those complex charge situations shows that this method isn't just for simple systems; it has some robustness in dealing with difficult chemical environments.

Lev: But you have to remember that even on those challenging cases, they found a systematic offset between their numbers and what the experiment actually measures, which points us toward the next kind of problem.

Kai: That offset comes from the protein reorganization free energy, meaning the way the protein shifts when it binds is changing the absolute baseline of their energy calculations.

Mira: So, for someone just looking at this for drug design, it means you can get a quick ranking to narrow down candidates without having to run those incredibly long simulations that are often required for absolute binding energies.

Lev: But the caveat is that you still need a way to account for that protein reorganization energy if you want the absolute numbers to be accurate enough for real biological prediction.

Kai: That’s exactly what they’re planning next, so they want to integrate an explicit estimate of that protein reorganization free energy into Felis so they can check if that shift is even there.

Mira: And they also mentioned co-optimizing the charge models and van der Waals parameters against things like hydration free energies, which should help make those predictions more realistic for condensed phase conditions.

Lev: If you want to run this on real hardware efficiently, you need those simulation times manageable, and if we can get better at modeling that protein reorganization energy, then this tool becomes much more useful for actual experimental guidance.

The paper's improvements: Kai: So, they’re talking about what needs to happen next for this tool to be truly useful in real drug discovery, and they have a few specific directions in mind.

Mira: Right, they are focusing on improving the actual predictions so that we get more reliable absolute binding energies in the future.

Lev: I think the biggest thing is getting better at determining where a ligand actually sits on the protein surface when it binds; if you get that pose wrong, your energy calculation is going to be off.

Kai: They mentioned needing more robust methods for assigning protonation states for those ligands, which is something that often gets messy in simulations but really affects the final energy score.

Mira: That makes sense because if the charge on a ligand is wrong by even a little bit, you’re introducing a big error into the energy calculation.

Lev: And beyond that, they want to integrate an explicit estimate of that protein reorganization free energy we talked about earlier so we can actually see if that baseline shift is real or if it’s just noise.

Kai: That would be a big step forward because it moves us from just calculating a number to understanding the physical process happening in the protein itself.

Mira: Plus, they suggest co-optimizing those charge models and van der Waals parameters against hydration free energies, which means tuning them to how water actually interacts with these molecules in solution.

Lev: That’s where we’d spend a lot of time on the error correction side, because those liquid properties are hard to model accurately and would have a huge overhead if we tried to incorporate that into our hardware models.

Kai: So, for the practical application right now, they see these improvements as necessary for moving from a good ranking tool to something that can reliably predict actual binding affinity numbers for novel compounds.

Mira: It means the next iteration of Felis needs to be smarter about how it handles those complex chemical environments and protein dynamics we talked about.

Lev: If we can nail that explicit reorganization energy term, then running this on real hardware becomes much more meaningful because the error budget for our quantum simulations starts making more sense.

Conclusion: Kai: So we’re wrapping up this look at the paper "Development and large-scale benchmarks of a protein--ligand absolute binding free energy toolkit". Basically, they built an automated system that lets us calculate how strongly a protein and a ligand bind in absolute terms without those confusing relative constraints.

Mira: Exactly, it’s about making that process scalable so researchers can use it for high throughput hit discovery campaigns right away.

Lev: From my side, the main thing is how fast this runs on real hardware; if the simulation times are manageable, then this toolkit actually becomes useful for error correction work.

Kai: They showed good ranking performance compared to state-of-the-art relative methods, and they handled tough cases like that KRAS G12D set pretty well too.

Mira: But we have to remember those caveats about the protein reorganization free energy shift, which means you still need to interpret those absolute numbers with that extra context of how the protein is moving.

Lev: If we want this for real hardware, we need better ways to model that reorganization energy because if we don't account for it correctly, the overhead lands somewhere else in our error budget.

Kai: It’s a useful tool for filtering hits quickly, but those limitations mean it’s not perfect yet.

Mira: So, the paper sets up a clear path forward by suggesting they need to integrate an explicit estimate of that reorganization energy into Felis and tune the underlying charge models better against things like hydration free energies.

Lev: If they nail that integration, then we could start thinking about how to use this kind of absolute scoring for real hardware constraints, which is a major goal for error correction.

Kai: It’s a solid piece of work because it tackles the computational cost hurdle in protein docking by focusing on that absolute energy calculation.

ByteDance Seed

physics.comp-ph

Submitted: 2026-03-23

Updated: 2026-11-09

Code: https://github.com/schrodinger/public_binding_free_energy_benchmark

License: http://creativecommons.org/licenses/by-nc-nd/4.0/

Importance score: 98/100

The gist: Absolute binding free energy (ABFE) calculations offer a theoretically rigorous approach for predicting protein–ligand binding affinities without the scaffold constraints of relative binding free…

Key concepts

Absolute Binding Free Energy (ABFE)
ABFE is a theoretical calculation that predicts the exact energy required to bind a ligand to a protein. It is more rigorous than relative binding free energy because it calculates the true binding affinity without needing prior reference points, offering direct predictions for drug design.
Double-Decoupling Protocol
This methodology involves decoupling the ligand from the explicit solvent and protein environment during simulation. By linearly scaling the ligand's partial charges to zero while keeping steric and bonded terms coupled, it allows for accurate energy calculations in a computationally efficient manner.
Zero-Shot Calculation
The study emphasizes that Felis performs all predictions without needing custom force-field modifications or fine-tuning of alchemical schedules. This means the toolkit is ready to use immediately on new targets and ligands, significantly reducing the computational burden for researchers.

Terminology

Summary

Absolute binding free energy (ABFE) calculations offer a theoretically rigorous approach for predicting protein–ligand binding affinities without the scaffold constraints of relative binding free energy (RBFE) perturbations. p. 0

Where the paper says it

Development and large-scale benchmarks of a protein–ligand absolute binding free energy toolkit Yu Liu1,∗ , Ailun Wang1,∗ , Yu Xia1 , Zhi Wang1,† , Wen Yan1,† 1ByteDance Seed ∗Equal contribution, †Corresponding authors Abstract Absolute binding free energy (ABFE) calculations offer a theoretically rigorous approach for predicting protein–ligand binding affinities without the scaffold constraints of relative binding free energy (RBFE) perturbations. However, broad adoption of ABFE in high-throughput hit discovery campaigns has been hindered by high computational costs and a lack of large-scale validation.

Here, we present Felis, an open-source, automated, and scalable toolkit designed for high-throughput ABFE calculations <ref:2603. p. 0
Where the paper says it

However, broad adoption of ABFE in high-throughput hit discovery campaigns has been hindered by high computational costs and a lack of large-scale validation. Here, we present Felis, an open-source, automated, and scalable toolkit designed for high-throughput ABFE calculations.

22274#pg2> Paired with ByteFF, a previously developed data-driven molecular mechanics force field for drug-like molecules, Felis achieves ranking performance comparable to state-of-the-art RBFE methods on a diverse dataset comprising 43 protein targets and 857 ligands <ref:2603. p. 0
Where the paper says it

Here, we present Felis, an open-source, automated, and scalable toolkit designed for high-throughput ABFE calculations. Paired with ByteFF, a previously developed data-driven molecular mechanics force field for drug-like molecules, Felis achieves ranking performance comparable to state-of-the-art RBFE methods on a diverse dataset comprising 43 protein targets and 857 ligands.

22274#pg2> Furthermore, we demonstrate robust convergence and ranking performance of Felis on a more challenging KRAS(G12D) dataset, where some ligands and the cofactor are highly charged <ref:2603. p. 0
Where the paper says it

Paired with ByteFF, a previously developed data-driven molecular mechanics force field for drug-like molecules, Felis achieves ranking performance comparable to state-of-the-art RBFE methods on a diverse dataset comprising 43 protein targets and 857 ligands. Furthermore, we demonstrate robust convergence and ranking performance of Felis on a more challenging KRAS(G12D) dataset, where some ligands and the cofactor are highly charged.

22274#pg2>. Crucially, all Felis predictions in this study were generated in a strict zero-shot manner, eschewing custom force-field modifications and alchemical schedule fine-tuning <ref:2603. p. 0
Where the paper says it

Furthermore, we demonstrate robust convergence and ranking performance of Felis on a more challenging KRAS(G12D) dataset, where some ligands and the cofactor are highly charged. Crucially, all Felis predictions in this study were generated in a strict zero-shot manner, eschewing custom force-field modifications and alchemical schedule fine-tuning.

22274#pg2>.

Toolkit Development and Methodology

Felis was developed as an automated ABFE toolkit to address the high computational costs and lack of large-scale validation hindering the broad adoption of ABFE in high-throughput hit discovery campaigns <ref:2603.22274#pg2> The calculations follow the standard double-decoupling (annihilation) protocol where the ligand is decoupled from explicit solvent or protein environment <ref:2603. p. 2

Where the paper says it

The absolute binding free energy is given by ∆Gbind = ∆Gsolvent + ∆Grestraint + ∆Gprotein, where ∆Gsolvent is the free energy of decoupling the ligand from solvent, ∆Grestraint is the analytical correction for the Boresch-style restraints, and ∆Gprotein is the free energy of coupling the ligand to the protein environment while releasing the Boresch-style restraints. 2 Method The ABFE calculations follow the standard double-decoupling (annihilation) protocol, in which the ligand is decoupled from explicit solvent or the protein environment.

22274#pg2> Restraints for protein–ligand alchemical simulations are defined using three protein anchor atoms and three ligand anchor atoms, including distance, angle, and dihedral terms <ref:2603. p. 2
Where the paper says it

We define these restraints using three protein anchor atoms (P1, P2, P3) and three ligand anchor atoms (L1, L2, L3). The restraint set comprises one distance term (P1-L1), two angle terms (P1-L1-L2 and L1-P1-P2), and three dihedral terms (P2-P1-L1-L2, P1-L1-L2-L3, and L1-P1-P2-P3).

22274#pg2> The double annihilation protocol involves linearly scaling the partial charge parameters of the ligand to zero while steric and bonded terms remain fully coupled <ref:2603. p. 3
Where the paper says it

2 Double annihilation protocol We employ a standard double annihilation protocol for all absolute binding free energy calculations. Along the alchemical pathway, the partial charge parameters of the ligand are linearly scaled to zero, while steric and bonded terms remain fully coupled.

22274#pg2> Simulation protocols utilize OpenMM with LangevinMiddleIntegrator, maintaining temperature and pressure at 298. p. 3
Where the paper says it

Unless otherwise specified, the equations of motion are integrated with the LangevinMiddleIntegrator [55] using a 2 fs time step and a friction coefficient of 0. Temperature and pressure are maintained at 298.

15 K and 1.013 bar <ref:2603. p. 3
Where the paper says it

15 K and 1. 013 25 bar, respectively.

22274#pg2>.

Benchmarking Performance

Felis was benchmarked on 43 protein targets comprising 857 ligands, selected as a subset of a comprehensive RBFE dataset to enable reliable ranking assessment <ref:2603.22274#pg2> Compared with state-of-the-art RBFE results, Felis achieves ranking performance comparable to FEP+ RBFE with substantially shorter simulations (3 ns or 5 ns with three independent replicas versus 20 ns) <ref:2603. p. 8

Where the paper says it

By benchmarking on 43 protein targets and 857 ligands, we showed that Felis delivers ranking performance comparable to state-of-the-art RBFE methods. This performance was achieved with substantially shorter simulations (3 ns × 3 replicas versus 20 ns) and without any forcefield fine-tuning.

22274#pg2> Both methods show that the ranking performance of ABFE is comparable to, although not better than, state-of-the-art RBFE predictions <ref:2603. p. 5
Where the paper says it

The debiased errors of Felis-ABFE are close to those of FEP+ RBFE. To our best knowledge, this is the first demonstration that the ranking performance of ABFE is comparable to, although not better than, state-of-the-art RBFE predictions, despite the dataset having been designed for RBFE benchmarking, where the dynamic range of ligand affinities is rather limited for each target protein.

22274#pg2>.

Force Field and Parameterization

The toolkit paired Felis with the AMBER ff14SB force field for proteins and ByteFF25 for ligands and cofactors <ref:2603.22274#pg2> ByteFF25 retains the same nonbonded parameters as GAFF2 while bonded parameters are trained on a more extensive quantum-chemistry dataset to improve coverage across chemical space <ref:2603. p. 1

Where the paper says it

Throughout, we paired Felis with the AMBER ff14SB force field for proteins and ByteFF25 for ligands and cofactors. Compared with ByteFF24 [45], ByteFF25 retains the same nonbonded parameters (charges and van der Waals, vdW) as GAFF2, while the bonded parameters are trained on a more extensive quantum-chemistry dataset to improve coverage across chemical space.

22274#pg2> All results were generated in a strict “zero-shot” setting, eschewing custom force-field modifications and alchemical schedule fine-tuning <ref:2603. p. 0
Where the paper says it

Furthermore, we demonstrate robust convergence and ranking performance of Felis on a more challenging KRAS(G12D) dataset, where some ligands and the cofactor are highly charged. Crucially, all Felis predictions in this study were generated in a strict zero-shot manner, eschewing custom force-field modifications and alchemical schedule fine-tuning.

22274#pg2>. The comparison between AM1-BCC and ABCG2 charge models showed no statistically significant difference in ranking performance <ref:2603. p. 6
Where the paper says it

Using the same training data and protocol, we trained ByteFF25-ABCG2 and compared the performance of the two charge models. We did not observe a statistically significant difference between the two charge models (Fig.

22274#pg2>.

Challenging Case Study Insights

The evaluation on the KRAS(G12D) dataset, featuring highly charged ligands, demonstrated the ability of Felis to handle complex binding scenarios <ref:2603.22274#pg2> For this series, at 10 ns simulation length, Felis-ABFE achieves Spearman’s ρ = 0. p. 8

Where the paper says it

3ns 5ns 8ns 10ns MD length −29 −28 −27 −26 −25 −24 −23 Predicted ∆ G (kcal/mol) A Convergence tested with MRTX-1133 Spearman’s ρ Kendall’s τ Ranking metrics 0. 0 Rank correlation B Ranking performance for 10 ligands 3 ns×3 5 ns×3 10 ns×3 Figure 4 Case study on KRAS(G12D) MRTX1133 series.

76 and Kendall’s τ = 0. p. 5
Where the paper says it

In contrast, Felis-ABFE used the same variant-level postprocessing to obtain ligand-level predictions, but did not apply CCC. To quantify performance within each target, we compared four metrics: Spearman’s ρ, Kendall’s τ , and debiased mean absolute error (MAE) and debiased root mean square error (RMSE).

64, demonstrating robust consistency with experimental data <ref:2603. p. 8
Where the paper says it

76 and Kendall’s τ = 0. 64, demonstrating robust consistency with experimental data.

22274#pg2>. The predicted binding energies for the KRAS(G12D) series exhibit a systematic offset relative to experiment, consistent with the interpretation that protein reorganization free energy between apo and holo states shifts the absolute baseline <ref:2603. p. 8
Where the paper says it

Detailed results from the benchmark runs at 3 ns, 5 ns and 10 ns are shown in Fig. Predicted ∆G values for the KRAS(G12D) series exhibit a systematic offset relative to experiment, consistent with the interpretation that protein reorganization free energy between apo and holo states shifts the absolute baseline [44].

22274#pg2>.

Future Directions

Real-world adoption of ABFE will likely require further improvements to enhance predictive accuracy, prioritizing more reliable binding poses and robust protonation-state assignment <ref:2603.22274#pg2> Future work plans include integrating an explicit estimate of protein reorganization free energy into Felis ABFE so that we can assess whether the global shift is resolved or mitigated <ref:2603. p. 9

Where the paper says it

In future work, we plan to integrate an explicit estimate of protein reorganization free energy into Felis ABFE so that we can assess whether the global shift is resolved or mitigated. Last but not least, integrating efficient enhanced sampling techniques[81, 82] is almost always beneficial to improve the convergence and reliability of thermodynamic sampling.

22274#pg2>. Other improvements could involve co-optimizing charge model and van der Waals parameters against hydration free energies and condensed-phase liquid properties <ref:2603. p. 9
Where the paper says it

This focus is particularly crucial because the bonded terms in ByteFF already provide a strong description of the intramolecular potential energy surface, as evidenced by the benchmarks reported in the ByteFF24 study [45]. Pragmatically, this entails co-optimizing the charge model and van der Waals parameters against hydration free energies and condensed-phase liquid properties, following strategies developed in the OPLS and OpenFF efforts [13, 77, 78].

22274#pg2>.

Acknowledgements

We thank Chang Han, Xingyuan Xu, Tianze Zheng for beneficial discussions and training of ByteFF25 <ref:2603.22274#pg2>.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Automated High-Throughput ABFE Calculations: The Felis toolkit can perform absolute binding free energy calculations in a strict zero-shot manner, eschewing custom force-field modifications and alchemical schedule fine-tuning. This makes it a ready-to-use tool for computational structure-based drug design.

  2. Force Field Robustness: The system benefits from paired force fields, specifically AMBER ff14SB for proteins and ByteFF25 for ligands and cofactors. ByteFF25 retains the same nonbonded parameters as GAFF2 while having bonded parameters trained on a more extensive quantum-chemistry dataset to improve coverage across chemical space.

  3. Ranking Performance Comparison: The improved system can achieve ranking performance comparable to state-of-the-art RBFE methods on diverse datasets, as demonstrated by Felis achieving ranking statistics close to FEP+ RBFE in both metrics.

  4. Challenging System Handling: The toolkit is capable of handling complex scenarios, such as the KRAS(G12D) dataset, which involves highly charged ligands and a complex charged environment, allowing for robust testing of the protocol.

  5. Error Mitigation through Post-Processing: The system can generate debiased results by applying post-processing steps to align predicted binding free energy values with experimental means, ensuring that P i (DeltaGfelis i - DeltaGExp i) = 0 relationship holds for each ligand series on a given target.

Abstract

Absolute binding free energy (ABFE) calculations offer a theoretically rigorous approach for predicting protein--ligand binding affinities without the scaffold constraints of relative binding free energy (RBFE) perturbations. However, broad adoption of ABFE in high-throughput hit discovery campaigns has been hindered by high computational costs and a lack of large-scale validation. Here, we present Felis, a source-available, automated, and scalable ABFE toolkit that integrates RelocationMC to enhance binding-site water sampling. Using the ByteFF25-AM1-BCC force field, ABFE calculations with Felis yielded similar average rank correlations to FEP+ RBFE across 43 benchmark systems comprising 857 ligand entries. The offset-adjusted errors were larger than those of FEP+ RBFE. In scytalone dehydratase, RelocationMC rapidly hydrated an initially dry cavity and reduced the dependence of binding free energy offsets on the initial water configuration. Across the benchmark, it reduced absolute offsets in 27 of 43 systems without a loss of aggregate predictive performance. We also evaluated sampling-length dependence and ligand ranking for a challenging KRAS(G12D) dataset containing highly charged ligands and a charged cofactor. Neither force-field parameters nor alchemical schedules were adjusted using experimental affinities from the benchmarks. These results support the use of Felis for ligand prioritization and highlight remaining sampling limitations.

Sources

Related papers