The supplied structure is validated and represented as SMILES, a drawing, numerical structural features or candidate 3D shapes. Unspecified groups and orientations remain recorded as uncertainty.
Preparation workflowSolubility and selected safety-related assays
Hypotheses
Effects, mechanisms and experimental evidence
EFFECT
“Anti-inflammatory” describes a possible reduction in an inflammatory response. Several mechanisms could produce that effect.
MECHANISM
“PTP1B inhibition” proposes that the compound reduces the activity of a protein involved in insulin signaling.
EXPERIMENTAL EVIDENCE
A laboratory measurement on a related compound provides a reason to investigate the same activity in the selected molecule.
Target predictions and experimental evidence
PubCheF suggests biological activities. ChemFIE, PPB3 and PIDGIN suggest proteins or laboratory test systems in which the compound may act. The table compares their predictions about the same protein, alongside SwissTargetPrediction where available.
Measured activity of related compounds
An analogue is a different molecule with a similar structure. These records describe experiments on the pictured reference compounds. Comparing those structures with the food molecule helps identify an activity worth testing. These experimental records are considered alongside the predictions.
Comparison across all molecules
Methods and reliability
Method comparison
Benchmark measures
Target hit rate is the fraction of reference compounds for which a measured target appears within the stated number of predictions. AUROC measures how well scores rank active versus inactive compounds: 0.5 is chance ranking and 1 is perfect ranking on that dataset. For example, AUROC 0.9 means that a randomly chosen active compound would receive a higher score than a randomly chosen inactive compound about 90% of the time, counting ties halfway. MSE is the average squared numerical error; lower is better within the same task and units.
A model can score well when it searches a short list of proteins or recognizes chemistry seen in training. The dataset and experiment therefore determine how far a benchmark result carries over to our unknown compounds.
Comparison table (CSV)Detailed benchmarks & controlsArchived methods & decisions
Scientific coverage and required inputs
Additional methods and availability
Training priorities
Benchmarks and known controls
The larger tests compare how often methods recover measured activities across many compounds. Andrographolide and berberine then provide individual examples whose biology we can examine in detail. Hosted services were evaluated on these two controls; local models also have the larger comparisons below.
Methods and terminology
Each model predicts its defined outcomes from the required inputs. A separate search retrieves measurements on related compounds. Protein sequences and assay descriptions specify the biological question for models that accept them.
Activity measurements can establish the affected process, the concentration required and differences between related compounds. Cellular and exposure measurements are needed to connect a biochemical interaction with effects from food.
Output categories
Glossary
Training models for this study
Selected methods and results
Selected-method tables and study records
Experiment archive
Selection decisions and complete experiment records, including results for both known controls. Ensemble components remain available for checking the combined predictor.