G: Gibbs free energy is the thermodynamic potential (G = H - TS) that, at constant temperature and pressure, represents the maximum non-expansion work a system can perform.

Δ_rG: Reaction Gibbs free energy at given physiological metabolite concentrations, which determines thermodynamic feasibility.

Δ_rG°: Reaction Gibbs free energy in standard conditions, which reflects the difference in thermodynamic stability of the reactants and products.

The actual and standard reaction free energies are connected by ΔrG = ΔrG° + RT ln Q, where R is the universal gas constant, T is the absolute temperature, and Q is the reaction quotient determined by reactant and product concentrations.

dGbyG web is a web-based platform that uses graph neural networks to predict standard Gibbs free energy change (Δ_rG°) for biochemical reactions. The name stands for "ΔG predicted by Graph neural networks." Our platform provides two main services:

• GEM ΔG Search: Search through 100+ genome-scale metabolic models (GEMs) with pre-calculated Δ_rG° values
• Reaction Predictor: Calculate Δ_rG° values for custom chemical reactions using our AI model

dGbyG is a graph neural network-based tool for predicting Δ_rG° of metabolic reactions.

1. Molecular graphs: each metabolite is encoded as an atom-bond graph with rich chemical features.
2. GNN prediction: the network learns thermodynamic representations from these graphs and computes Δ_fG° for metabolites. Δ_rG° was then computed from the weighted sum of the Δ_fG° of the substrates and products with the stoichiometric coefficients as the weights.
3. Methodological advantage: Unlike group-contribution approaches that rely on predefined chemical groups, dGbyG directly models atoms and bonds, ensuring robustness and generalization even with limited training data or unseen reaction mechanisms.
4. Condition parameters: users can set pH, ionic strength (I), pMg, and electrical potential (V) to reflect physiological environments.
5. Output: predicted Δ_rG° together with an uncertainty estimate.

Compared with the best existing approaches, dGbyG reduces the median prediction error on the validation set from 5.33 kJ/mol to 4.11 kJ/mol, while simultaneously increasing the coverage of reactions in the human genome-scale metabolic network model Recon3D from 64.14% to 71.22%.

If you would like to learn more about the principles behind dGbyG, please refer to the paper unraveling principles of thermodynamics for genome-scale metabolic networks using graph neural networks.

SD is the model's predictive uncertainty for that reaction (kJ/mol).

• How we compute it: For each training reaction we estimate a variability index α from its experimental SEM and sample size; we then create 100 perturbed datasets by adding Gaussian label noise N(0, α), train 100 models, and take the SD of their predictions for each reaction. (We also use α as a sample weight in the loss.)
• How to interpret: SD measures the spread of the 100 ensemble predictions. Smaller SD means higher agreement among models (higher confidence).

Our GEM library currently includes 115 GEMs:

• BiGG Models (107 GEMs, excluding Recon3D): sourced from BiGG Models.
• Metabolic Atlas (7 GEMs): sourced from Metabolic Atlas GEM repository. (for these 7 models, we only retain reactions and metabolites annotated with BiGG IDs).
• Recon3D (1 GEM): the human Recon3D model from the latest release by Elizabeth Brunk et al. (SBRG/Recon3D).

You can search for:

• Models: Search by model name (e.g., "Mouse-GEM", "iML1515", "Recon3D")
• Reactions: Search by reaction BiGG ID or name (e.g., "HEX1", "Hexokinase")
• Metabolites: Search by metabolite BiGG ID (with compartment, e.g., "g6p_c") or BiGG Universal ID (without compartment, e.g., "g6p")
• Genes: Search by gene ID or name (e.g., "3948_AT2", "LDHC")

For each model compartment (e.g., cytosol, mitochondrion, lysosome), we set the parameters used in Δ_rG° calculations—pH, ionic strength (I), pMg, and electrical potential (E)—as follows:

1. Literature: Values reported for the given species/compartment.
2. BioNumbers: If not available in the literature, use typical compartment values from BioNumbers.
3. Closest-species: If neither source provides a value, adopt conservative values from closely related species.

The chosen parameters are shown on the model page under “Compartment Conditions”

We support multiple compound identifier formats:

• SMILES: Simplified Molecular Input Line Entry System (e.g., CCO for ethanol)
• InChI: International Chemical Identifier (e.g., InChI=1S/C2H6O/c1-2-3/h3H,2H2,1H3 for ethanol)
• InChI Key: Hashed version of InChI (e.g., LFQSCWFLJHTTHZ-UHFFFAOYSA-N for ethanol)
• KEGG ID: KEGG compound identifiers (e.g., C00031 for glucose)
• MetaNetX ID: MetaNetX database identifiers (e.g., MNXM7381 for glucose)
• ChEBI ID: Chemical Entities of Biological Interest identifiers (e.g., 17234 for glucose)
• PubChem CID: PubChem Compound IDs (e.g., 702 for ethanol)
• HMDB ID: Human Metabolome Database identifiers (e.g., HMDB0000122 for glucose)
• Compound Name: Common chemical names (e.g., phosphate, acetate)
• BiGG Universal ID: BiGG database identifiers without compartment suffixes (e.g., succ not succ_c)

Recommendation: Use SMILES or InChI for best accuracy and reliability.

Rules:

• Use = (with spaces) to separate reactants and products.
• Use + between compounds; put coefficients in front (e.g., 2 H2O).

Examples:

You can either choose a preset for common compartments or set your own conditions.

Presets: Standard, Cytosol, Mitochondria, Lysosome, Golgi, etc. (each preset fills pH, I, pMg, E, T)

Custom conditions:

• pH (0-14)
• Ionic strength (0-1 M)
• pMg (0-14)
• Electrical potential (-1 to 1 V)

Yes. Batch mode processes up to 100 reactions in one run.

• Paste one equation per line; optionally set global pH / I / pMg / E.
• Download results as a CSV.

Δ_rG° values are reported at 25 °C (298.15 K).

• Most experimental thermodynamic data and our model training are standardized at 298 K.
• Reliable temperature adjustment would require reaction-specific ΔH and ΔS (ΔG = ΔH - TΔS), and often heat-capacity terms, which are not available at the genome scale.