TPDB
Help/About
Data sources, evaluation protocol, metrics, and using TPDB.
Data Sources
TPDB brings together peptide sequences, activity annotations, and property measurements from curated databases, published datasets, and other named resources. Database Source and Evidence Records preserve provenance and available assay details; source labels do not imply a uniform experimental criterion.
- Advanced Search
- Records include standard sequences, modified peptides, and sequences containing non-standard residues.
- Benchmark
- A standardized evaluation resource with 23 classification tasks and 3 regression tasks. Classification datasets retain only peptides composed of the 20 standard amino acids, with lengths up to 50 aa.
- Unnatural Amino Acids
- A separate collection of individual amino acids, with chemical information, 2D structures, and original source links.
View Tasks and Data Sources
Task sources reported in the supplementary methods include databases and published prediction datasets.
| Task / Resource | Sources |
|---|---|
| Classification | |
| Antibacterial activity | Peptipedia v2.0; AntiBP3; dbAMP 3.0; DRAMP; PepLab |
| Antifungal activity | Peptipedia v2.0; Antifungipept; dbAMP 3.0; AFP-GFuse; AntiFP; DRAMP; PepLab |
| Antiparasitic activity | Peptipedia v2.0; i2APP; dbAMP 3.0; DRAMP |
| Antiviral activity | AI4AVP; Peptipedia v2.0; DRAVP 2.0; dbAMP 3.0; AVPpred-BWR; DRAMP |
| Quorum sensing activity | EnsembleQS; QSPred-FL; Peptipedia v2.0 |
| ACE inhibitory peptides | Peptipedia v2.0; PepLab; FermFooDb |
| Antidiabetic activity | Peptipedia v2.0; PepLab |
| Antihypertensive activity | AHTpin; Peptipedia v2.0; AHTPDB; FermFooDb |
| DPP-IV inhibitory peptides | iDPPIV (predictor datasets); Peptipedia v2.0; PepLab |
| Anticancer activity | Peptipedia v2.0; AntiCP 2.0; iACP-DRLF; dbACP; CancerPPD2; ACP740 / ACP240; SATPdb; PepLab; dbAMP 3.0 |
| Tumor T-cell antigens | ENCAP; LYnet; iTTCA-Hybrid; iTTCA-RF |
| Tumor-homing peptides | StackTHPred; StackTHP; TumorHoPe2; Peptipedia v2.0 |
| Antiinflammatory activity | AIPpred; Peptipedia v2.0 |
| Hemolytic activity | APD (Antimicrobial Peptide Database) subset; HemoPI |
| Neurotoxicity | NTxPred2; Peptipedia v2.0 |
| Toxicity | Peptipedia v2.0; ToxinPred3.0; PepLab |
| Blood-brain barrier penetration | B3Pred; BBPpredict; Peptipedia v2.0 |
| Cell-penetrating peptides | CPPsite3; MLCPP 2.0; PerseuCPP; CellPPD; Peptipedia v2.0; CellPPD-Mod |
| Neuropeptides | Peptipedia v2.0; PepLab |
| Bitter peptides | TastePeptides-Meta / TastePeptidesDB |
| Umami peptides | TastePeptides-Meta / TastePeptidesDB |
| Antioxidant activity | Multi-AOP; AoXPdb; AODB; Peptipedia v2.0; AnOxPePred-1.0; PepLab; FermFooDb |
| Antifouling activity | PeptideBERT dataset |
| Regression | |
| Half-life | PEPlife; THPdb; B3Pdb |
| Caco-2 Permeability | CycPeptMPDB; PepLand |
| PAMPA | CycPeptMPDB; PepLand |
| Chemical Resource | |
| Unnatural Amino Acids | HighPlay2; SwissSidechain; PubChem; linked RCSB Chemical Component Dictionary information |
Evaluation Protocol
The supplementary methods describe fixed, similarity-controlled partitions shared by predictive methods, with five training random seeds.
- Fixed Splits
- Complete similarity clusters are assigned to Train / Valid / Test at approximately 8:1:1. The split stays fixed across models and seeds.
- Training and Selection
- Train fits each model; Valid selects its best training state or stopping point; Test provides the held-out performance. Training seeds are 0, 1, 2, 3, and 4.
- Distribution Evaluation
- Traditional (TR), PU, and Biologically (BI) are evaluated on all three original test schemes. Matching schemes measure same-distribution performance; different schemes measure cross-distribution performance.
- Mixed-source Training
- MS equally combines negatives from the corresponding TR, PU, and BI training and validation sets. It is evaluated on the three original test sets and has no independent MS test set.
How the Result Selections Map to Evaluation
| Website Selection | Training Scheme | Test Scheme |
|---|---|---|
| TR | TR | TR |
| PU | PU | PU |
| BI | BI | BI |
| MS-TR | MS | TR |
| MS-PU | MS | PU |
| MS-BI | MS | BI |
| TR-PU | TR | PU |
| TR-BI | TR | BI |
| PU-TR | PU | TR |
| PU-BI | PU | BI |
| BI-TR | BI | TR |
| BI-PU | BI | PU |
The menu lists all 12 evaluation combinations. TR, PU, and BI denote matched training and test schemes. For hyphenated selections, the first scheme specifies training and the second specifies testing. The six cross-distribution selections follow MS-BI. Compare methods within the same task, metric, and selection. The results use the similarity-controlled partitions summarized by the Overview matrix.
Preprocessing, Similarity Control, and Model Selection
- Classification
- Remove invalid records and exact duplicates, then reduce sequence redundancy with MMseqs2 at 90% identity and 90% coverage. Merge the unique TR, PU, and BI sequences within each task and cluster at 30% identity and 80% coverage before assigning whole clusters to splits. Positive-sample assignments are shared across the three schemes.
- Regression
- Remove invalid records and integrate repeated measurements after outlier handling. Half-life uses MMseqs2 clustering at 30% identity and 80% coverage. Caco-2 and PAMPA use chirality-aware Morgan fingerprints and Butina clustering at Tanimoto similarity 0.60. Label quantile bins help balance the cluster-based splits.
- Checkpoint Selection
- Neural classification checkpoints are selected by validation ROC-AUC improvements greater than 0.001; neural regression checkpoints by improved validation RMSE. XGBoost and LightGBM use validation-loss early stopping; random forests use fixed tree counts.
Metrics
Compare scores within the same task and test scheme. Higher scores are better except for RMSE and MAE, where lower errors are better.
Classification
| Metric | Measures | Better |
|---|---|---|
| ROC-AUC | Ranking of positives above negatives. | ↑ |
| PR-AUC | The precision-recall trade-off. | ↑ |
| MCC | Agreement between predicted and true classes. | ↑ |
| Precision | Correct positives among positive predictions. | ↑ |
| Recall | True positives recovered by the model. | ↑ |
| F1 | Harmonic mean of precision and recall. | ↑ |
| Accuracy | Correct predictions among all samples. | ↑ |
Regression
| Metric | Measures | Better |
|---|---|---|
| RMSE | Prediction error, with more weight on large errors. | ↓ |
| MAE | Average absolute prediction error. | ↓ |
| R2 | Performance relative to a mean-only prediction; it can be negative. | ↑ |
| Pearson | Linear association of predictions and measurements. | ↑ |
| Spearman | Agreement in predicted and measured rankings. | ↑ |
- Mean and error bars: task-method bars show the mean over five seeds, with ± one standard deviation. This is not a confidence interval; hover over a bar for its values.
- Regression scale: targets are log10-transformed: Half-life in seconds and Caco-2 and PAMPA in cm/s. RMSE and MAE are calculated on the corresponding log-transformed target scale; R2, Pearson, and Spearman are dimensionless. Compare errors within a task, not across different properties.
- Robustness summaries: variability across training seeds is distinct from variability across test distributions. Overview heatmaps macro-average task-method means; their displayed error is the mean of the component seed standard deviations.
Using TPDB
- Advanced Search
- Search IDs, sequences, activities, or sources; refine by length and sequence type, or use global sequence alignment. Coordinate-based structure similarity is temporarily unavailable. Open a result to inspect source annotations and available assay details, copy Sequence/FASTA, and view available 3D structures. Exact-sequence experimental PDB matches take priority over predictions; predicted structures are not experimental evidence of activity.
- Benchmark Results
- Choose Activity or Model and one of the 12 training/test evaluation combinations, then switch metrics. For regression, choose Caco-2, Half-life, or PAMPA.
- Benchmark Download
- Retrieve the paper's fixed training, validation, and test sets, plus classification and regression result tables. MS downloads reuse the three original test schemes.
- Unnatural Amino Acids
- Select a chemical category to inspect individual amino acids, 2D structures, and linked source information.
- API
- Use the documented REST endpoints for programmatic retrieval. Preserve source fields when reporting reused or derived data.
For record issues, use Contact us and include the TPDB ID, page URL, and a short description.