Gibby

Free energies for small molecules and protein complexes

How it works

You give it a molecule as SMILES, or a protein complex as a PDB file or a four-character PDB code, and it returns a free energy. Each answer comes with the thermodynamic cycle behind it, the terms that add up to it, the assumptions it makes, and an error bar measured on data the model was not fitted to.

This host answers within one request and keeps nothing afterwards. MM/GBSA and the queue for long calculations run in the service on your own machine.

What runs on this host

This host runs what finishes within one request: protein-protein affinity, mutation ddG, log P, absolute G for molecules up to 20 heavy atoms, and solvation by CPCM-X up to 35, the last two with xtb. Each answers in a function that keeps nothing once it has replied, so nothing is queued or stored here. It sends a request elsewhere in two cases: to fetch a structure from the RCSB when you give it a PDB code, and, for SMD solvation and the blend, to queue the job on the ORCA machine, which keeps it until you delete it (privacy). The pages load no third-party code or cookies beyond sign-in, and the calculator's 2-D preview is drawn by RDKit on the host. MM/GBSA and the queue for long calculations are in the service you run on your own machine.

Examples

Each answer is the calculator's own for that input.

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

Accuracy

Each figure is the typical error on data held out from fitting. The server prints the one that applies beside every answer.

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
Protein-protein bindingSKEMPI and PDBbind, family held out1.90 to 3.02 kcal/mol by type
Mutation ddGSKEMPI, family held out1.45 kcal/mol

Frequently asked questions

What can it answer?

Absolute Gibbs free energies and solvation free energies for small molecules, log P between two solvents, MM/GBSA binding energies for a protein-ligand or protein-protein complex, the effect of a mutation on binding, and protein-protein affinity from the interface.

What can it not answer?

log D at a stated pH, because there is no pKa model here, so every log P is for the neutral species. There is no fast route for large ions. MM/GBSA leaves out the entropy of binding, so its absolute value is not meaningful, and the server says so next to every MM/GBSA number. A salt, hydrate or mixture written as one SMILES with a dot, because every error bar was measured on single molecules: ask about each molecule, or the parent of a salt, on its own.

Why does it show the decomposition?

A single number gives you no way to judge whether to trust it. With the cycle, the terms, the assumptions and the largest likely error printed beside it, you can usually see where a bad answer went wrong.

Is my structure uploaded anywhere?

Yes. A structure you give the calculator is sent to this host for the function to answer, so it does leave your machine. The host does not keep it: nothing is written that outlives the request, signed in or not. The one exception is SMD solvation and the blend: that SMILES is queued on the ORCA machine and kept until you delete it from the account page, as the privacy page explains. If you run the service yourself, the file stays on your machine under the hash of its bytes in ~/.cache/gibbs/, in a directory only your account can read.

How is it secured?

This host has no database or file store and keeps nothing between requests. When sign-in is on, it goes through Google, so no password exists here; the session is a signed HttpOnly cookie, and the routes that compute refuse a browser without one. The pages are served with a Content Security Policy that allows scripts only from this host, and every response carries Strict-Transport-Security. The service you run yourself adds a queue and a structure store readable only by your account, and expects a proxy in front of it to handle TLS.

The API

The calculator is a page on top of a small HTTP API. Each route takes JSON and answers in the response, and a GET on any route describes what it takes, with an example request.

RouteEngineTakes
POST /api/affinityC, protein-protein affinitypdb_id or structure, groups
POST /api/ddgB, mutation ddG pdb_id or structure, groups, mutations
POST /api/logpD, log P against water smiles, solvent
POST /api/gibbsD, absolute G with xtb smiles, solvent
POST /api/solvationD, solvation with xtb's CPCM-Xsmiles, solvent, model
POST /api/orcaD, SMD or the blend, queued on the ORCA machinesmiles, solvent, model
POST /api/orca/jobthat job's state and, when done, its answerid
POST /api/chainsthe chains in a structure, to name two partnerspdb_id or structure
POST /api/detectthe complex type, read off the sequencespdb_id or structure, groups
POST /api/moleculethe SMILES in a .pdb, .cif, .sdf, .mol, .mol2 or .xyzstructure, filename
POST /api/depicta 2-D drawing, as SVG smiles
GET /api/healthwhich engines are here, and why the others are not-
DELETE /api/datadelete what is held for the signed-in person, and sign out-
GET /api/gatewhether sign-in is on, and whether this browser is signed in-
SMD, sign-in and size limits

SMD solvation and the blend built on it need ORCA, which this host cannot run. POST /api/orca queues them on a separate ORCA machine and POST /api/orca/job reads the answer back, with the error bar measured for that model. That works only while the ORCA machine is online, and /api/health says whether it is; when it is not, the request is refused rather than answered by CPCM-X. MM/GBSA needs OpenMM and is not available here. POST /api/solvation answers with CPCM-X, says so in every answer, and quotes the error bar CPCM-X measured at that molecule's size, or on solutes sharing one of its elements where those scored clearly worse.

When sign-in is on, the POST routes answer only a signed-in browser. The GET descriptions stay open to anyone.

Request bodies are limited to 4 MB, because Vercel rejects anything larger before the function runs, and structures to 50,000 atoms. For a bigger structure, give its PDB code or run the service on your own machine.