SynCED incorporates several computational predictors to extend coverage of chemicals lacking experimental data. This page describes each predictor, its performance, applicability domain, and where to find additional information or source code.
Target Predictions
Ligand-based protein target predictions in SynCED are generated using a method conceptually based on the Similarity Ensemble Approach (SEA). ECFP4 molecular fingerprints (radius 2, 2,048 bits; RDKit version 2025.09.4) and Tanimoto similarity are used to compare SynCED chemicals against a curated set of reference ligands from ChEMBL. Candidate targets are ranked by the mean of the five highest Tanimoto similarities to reference ligands annotated for each target.
Confidence & Applicability Domain Metadata
Each predicted chemical–target association in SynCED is accompanied by the following metadata, displayed alongside the prediction:
| Field | Description |
|---|---|
| Mean top-k Tanimoto similarity | Prediction confidence score. Reflects the mean of the five highest Tanimoto similarities between the query compound and reference ligands for that target. |
| Max Tanimoto similarity | Applicability domain indicator. Predictions where this value is below 0.3 are flagged as low confidence, indicating the query compound falls outside the well-characterized chemical space of the training data. |
| Empirical p-value | Statistical significance score derived from a null distribution of 100,000 random Tanimoto scores across all tested targets. |
| Number of reference compounds | The number of ChEMBL reference ligands supporting the prediction, indicating the breadth of evidence underlying each association. |
Training Data & Chemical Space
Reference ligands were drawn from ChEMBL records with valid canonical SMILES, single-protein target annotations, and a reported IC50, EC50, Ki, or Kd below 10 mM. Because ChEMBL bioactivity data are enriched for drug-like compounds, predictive performance for structurally dissimilar environmental chemicals — such as industrial additives, pesticides, or persistent pollutants — may be lower than benchmark figures suggest.
The UMAP projection below shows the distribution of ChEMBL training compounds (grey) relative to SynCED chemicals (blue), illustrating the overlap and gaps in chemical space coverage.
UMAP projection of ECFP4 fingerprints. Grey: ChEMBL reference ligands. Blue: SynCED chemicals.
Example Predictions
To illustrate predictive utility, target predictions for P,P'-DDD (CED0000016) — a persistent organochlorine pesticide and DDT metabolite — included the sodium-dependent dopamine, serotonin, and noradrenaline transporters, aromatase (CYP19A1), CYP2D6, and the D(2) and D(3) dopamine receptors. These predictions are biologically plausible given the compound's known neurotoxicity and endocrine-disrupting activity, and complement its experimentally characterized interactions with androgen and estrogen receptors, GABA-A receptor subunits, voltage-gated sodium channels, and Na⁺/K⁺-ATPase.
SynCED explicitly distinguishes predicted targets from experimentally supported or literature-derived interactions. Predictions are intended to support hypothesis generation, mechanistic interpretation, and the prioritization of chemicals for further investigation.
Spectral & Chromatographic Predictors
Spectra and chromatographic properties were predicted for SynCED chemicals lacking experimental data using several locally developed tools. Performance benchmarks, applicability domain information, and links to source code or web servers for each tool are provided below.
| Tool | Output | Performance | Applicability Domain / Notes | Reference / Link |
|---|---|---|---|---|
| FraGNNet | Positive- and negative-ion ESI–MS/MS (QTOF and Orbitrap) | Cosine similarity >0.85 between predicted and observed spectra | Best suited to compounds representable as molecular graphs with standard organic functional groups. Interpret with caution for inorganics, polymers, and highly complex structures. | Young et al. (2024) |
| PredEI | EI-MS spectra | Average cosine similarity 0.755 ± 0.099 (NIST20 test set); 0.768 ± 0.102 (NIST23 test set). 80% of compounds identified as the top hit; 87% within the top two; 100% within the top ten when EI-MS and RI data are combined. | Validated on underivatized compounds and TMS/BSTFA-derivatized compounds. Best suited for volatile and semi-volatile organics within the mass range of the NIST training data. Predictions for thermally labile, very large, or highly polar compounds should be treated with caution. | Wakoli et al. (2024) |
| CASPRE / PROSPRE | 1H and 13C NMR spectra | PROSPRE 1H MAE 0.10 ppm (water, HMDB holdout); 0.19 ppm (organic solvents, NP-MRD holdout). PROSPRE outperforms MNOVA, NMRShiftDB2, and CASCADE in both conditions. | PROSPRE was trained primarily on water-soluble biological compounds (metabolites, amino acids, nucleotides). Performance is best for small, water-soluble organics (<600 Da). Reduced accuracy for structurally diverse natural products and compounds dissolved in organic solvents. Predictions supported for water, methanol, chloroform, and DMSO. | Sajed et al. (2024) |
| RT-Pred | HPLC retention times | Mean R² of 0.95 across 20 chromatographic methods | Validated on small organic molecules across standard reversed-phase HPLC methods. Predictions for highly polar, ionic, or very large compounds may be less reliable. | Zakir et al. (2025) |
| RIpred | GC retention indices | Mean absolute percentage error within 3% for most GC column models evaluated | Validated on volatile and semi-volatile organic compounds. Less suitable for thermally labile compounds or those with very high molecular weights. | Anjum et al. (2023) |
UMAP projections comparing the training chemical space of each spectral predictor to the SynCED chemical inventory are shown below. These visualizations help users assess the degree of overlap between the predictor's training domain and the compound they are querying.
UMAP projections for each spectral predictor. Grey: training set compounds. Blue: SynCED chemicals.
Predictions for highly complex structures, inorganic compounds, polymers, and compounds containing unusual functional groups should be interpreted with greater caution, as these may fall outside the training domain of one or more predictors. Consult the primary publications linked above for full applicability domain descriptions.
Structure
