Boltz-2 results are ours: each complex was predicted with Boltz-2 through Revilico’s co-folding engine. FEP+ results are Schrödinger’s published predictions from the same public benchmark; we did not re-run them. Both are scored the same way, by how well each method’s prediction tracks the measured binding free energy within a system.
What the comparison shows
Boltz-2 tracks experiment well in many systems: its correlation is positive in 63 of 72, with an R² above 0.5 in 35. FEP+ is still the more accurate method overall, with a mean R² of 0.59 against 0.44 and the higher correlation in 48 of the 72 systems. The gap is narrowest on the JACS set (0.54 vs 0.58) and the Water set (0.61 vs 0.62), and widest on the Merck systems (0.21 vs 0.49) and the MCS docking sets (0.06 vs 0.60). Boltz-2 does come out ahead in 24 systems, including MCL1 with 42 ligands (0.83 vs 0.49).By benchmark family
Mean R² and Spearman ρ of each method’s predictions against experiment, averaged over the systems in each family.System by system
Every system, sorted by Boltz-2’s correlation. Pick a family to see all of its systems.Where Boltz-2 gains and where it falls behind
The difference in R² for each system, Boltz-2 minus FEP+. Bars to the right are systems where Boltz-2’s predictions correlate better with experiment.How the correlations are distributed
Predicted against experimental, ligand by ligand
Pick a system and a method. Boltz-2’s affinity output is in model units rather than kcal/mol, so its panel shows the trend but has no 1:1 line; the FEP+ panel does.How the benchmark was run
Systems and ligands
Systems and ligands
The ligands come from Schrödinger’s public binding free energy benchmark, which groups protein–ligand series into families (for example the JACS set, the Merck set, macrocycles, fragments and GPCRs). The comparison covers 982 ligands in 77 series. Series with fewer than three ligands, 10 ligands in all, are left out because a correlation over one or two points means nothing, which leaves 72 systems and 972 ligands.
How the numbers are calculated
How the numbers are calculated
For each system, R² is the squared Pearson correlation between a method’s predictions and the experimental binding free energies of its ligands, and ρ is the Spearman rank correlation. Boltz-2’s predicted affinity is used as it comes from the model; FEP+ uses its predicted ΔG. Family averages are the mean over a family’s systems, so every system counts once. No value is shifted or re-fitted.
Reading these results
Correlation, not error
Correlation, not error
Boltz-2 reports affinity in model units, not kcal/mol, so this page compares how well each method ranks and tracks the ligands within a system. It does not give an absolute error for Boltz-2, and it does not say how close its numbers are to a measured free energy.
Different jobs, different costs
Different jobs, different costs
FEP+ is a relative free-energy calculation that needs a prepared perturbation network. Boltz-2 predicts affinity directly from the complex it folds. This page compares accuracy only: the data here has no runtime or cost for either method, so it makes no speed or cost claim.
Small systems are noisy
Small systems are noisy
Many systems have between 3 and 10 ligands. Their correlations swing widely with a single ligand, which is why the family table also reports the systems with 10 or more ligands, and why the biggest gains in the chart are mostly small systems.
The FEP+ numbers are Schrödinger's, not ours
The FEP+ numbers are Schrödinger's, not ours
They come from Schrödinger’s public repository and reflect their protocol, force field and expert preparation of each system, which Boltz-2 does not get.
Sources
- FEP+ predictions and experimental values: Schrödinger, public binding free energy benchmark.
- Boltz-2: the open-source Boltz co-folding model with its affinity prediction head, run through Revilico’s co-folding engine.
- Analysis: per-ligand predictions for all 972 ligands, scored per system as described above.