Gibergy

About

Gibergy computes free energies for small molecules and protein complexes with models that combine physics and machine learning. Every figure below was measured on data the model was not fitted to, and the calculator prints the one that applies beside each answer. Where a model has no measurement for an input, the answer says so instead of quoting a figure.

Accuracy

Table 1. Typical error (RMSE) on held-out data.

QuestionMeasured againstTypical error
log P, octanol/waterSAMPL, never seen0.49 log units
log P, octanol/waterLipophilicity, never seen0.85 log units
Solvation free energy, CPCM-XCombiSolv, CPCM-X alone2.31 kcal/mol in water, 1.38 in other solvents, rising with size; with one or two O-H, a fitted term and 1.35 in water on unseen solutes
Solvation free energy, SMD blend, ORCACombiSolv, solute class held out1.02 kcal/mol, 1.19 in water; answers only while the ORCA machine is online
Solvation free energy, blend, new solutesFlexiSol, never seen, pre-registered2.67 kcal/mol on 213 rows, against 3.20 for SMD with the trees
Protein-protein bindingSKEMPI and PDBbind, family held out2.28 kcal/mol, r 0.48; 1.90 to 3.02 by type
Mutation ddGSKEMPI, family held out1.45 kcal/mol, r 0.56, single-point mutations
012345012345Measured log PPredicted log PRMSE 0.49, r 0.84, n = 33
Figure 1. Octanol/water log P of the 33 SAMPL6 and SAMPL7 challenge molecules, none of which the model was fitted on. The solid line is parity and the dashed lines are one log unit either side.
0123451-67-1011-1516-2526-35Heavy atoms in the soluteRMSE (kcal/mol)27246830821028All solventsWater
Figure 2. Error of the SMD blend on CombiSolv with each solute class held out, by solute size. Numbers above the bars are solutes in each band; bands with fewer than 10 are left out. The error grows with size, and each answer quotes the band its molecule falls in.
012345Water (116)Octanol (97)Solvent (rows)RMSE (kcal/mol)2.93.12.92.74.23.32.82.7SMDSMD with treesNetworkBlend
Figure 3. FlexiSol solutes that appear in none of the training sets, scored once under a rule written before scoring. The blend averages SMD with trees and the network; over all 213 rows it is 2.67 kcal/mol against 3.20 for SMD with trees.
01234TCR-peptide-MHCn = 1151.90Antibody-antigenn = 6982.10Othern = 20262.29Enzyme-inhibitorn = 1573.02RMSE (kcal/mol)
Figure 4. Protein-protein binding affinity, held out by sequence family over 2996 complexes from SKEMPI and PDBbind. The solid line is the error over all types (2.28 kcal/mol) and the dashed line is the error of always predicting the average (2.59).

Table 2. Mutation ΔΔG on SKEMPI, held out by sequence family. The last column is the error of always predicting the average change.

MutationsRowsComplexesRMSErPredicting the mean
Single-point mutations41433201.450.561.75
Two or more mutations17151492.240.542.65
All58583441.720.572.09

Examples

Table 3. Answers from the calculator for three inputs.

InputCalculationMethodAnswer
Paracetamollog P, octanol/waterfitted transfer model0.44 log units
Paracetamolabsolute G in waterxtb (GFN2), eight conformers-20508.33 kcal/mol
Barnase and barstar, 1BRSbinding affinity measured in a lab (SKEMPI 2.0), shown in place of a prediction, with the model's own beside it-19.09 kcal/mol

Data and licences

The solvation model is fitted on CombiSolv-Exp (Vermeire et al., CC-BY-4.0) and FreeSolv, the log P model on ChEMBL (CC-BY-SA) with every benchmark molecule removed first, and the protein models on SKEMPI and PDBbind under their own terms. MNSol is used only to measure accuracy and is not redistributed. CheMeleon ships under its own licence. The quantum chemistry uses xtb from the Grimme group at the University of Bonn (LGPL-3.0). The tables and figures on this page are read from the model files in the code when the site is built.