1. Overview
PPB4 is a target-prediction tool for small molecules. Given a chemical structure, it estimates which protein targets the molecule is most likely to be active against, and which targets it is most likely to be inactive against. Predictions are produced by multitask deep neural networks trained on bioactivity data from ChEMBL 35.
Compared to its predecessor PPB3, PPB4 introduces three changes worth noting up front: (a) it now predicts confirmed inactivity in addition to activity, (b) it is trained on a substantially larger dataset (~5.8 M compound–target interactions), and (c) it adds the MAP4 fingerprint, which is particularly well suited to peptides and other macrocyclic molecules.
2. Your first prediction
Open the homepage and follow these steps:
- Provide a structure. Either draw your molecule in the JSME editor or paste a SMILES string into the Current SMILES box. The drawing and the SMILES box stay synchronised: editing one updates the other.
- Try the example. If you just want to see the tool in
action, click Load example to fill in the test compound
(
O=C(NO)c4ccc(CN3c1cc(C(F)(F)F)ccc1S(=O)(=O)CC2CCCCC23)cc4) and run it through the models to see which targets PPB4 predicts it hits, and which it predicts it does not. - Pick a model. Consensus is the recommended choice: it combines all four fingerprint models. See section 4 below for guidance.
- Choose how many targets to predict. The slider goes from 5 to 50; 20 is usually a good starting point.
- Click Predict targets. A new browser tab opens at
a unique results URL such as
ppb4.gdb.tools/result/<id>and renders the predictions.
Predictions usually finish in under five seconds for a single compound.
3. Reading the results
Each compound you submit is shown as its own card on the results page. Within each card you will see three tabs:
- Merged predictions (default): active and inactive predictions brought together into one ranked list, with one row per target. Coloured pills on each row show which model or models predicted that target. The active and inactive networks are trained separately, so the same target can be predicted by both; when that happens the row carries both pills and both probabilities, shown one under the other, and nothing is discarded. Targets are ranked by the higher of their two scores.
- Active predictions: the top targets ranked by P(active) only.
- Inactive predictions: the top targets ranked by P(inactive) only.
Each row contains the target name (linked to its ChEMBL report card), the ChEMBL ID, the target type and protein class, the organism, and a confidence score between 0 and 1. The confidence is the raw output of the corresponding neural network and should be read as a relative ranking signal rather than as a calibrated probability.
4. Choosing a fingerprint
PPB4 offers four fingerprint-based models. They all map your molecule to a 4096-bit vector that the neural network consumes, but they capture chemistry differently:
| Fingerprint | What it encodes |
|---|---|
| ECFP4 | Circular substructures around each atom, up to a radius of two bonds. |
| AtomPair | Pairs of atoms together with the topological distance between them. |
| Layered | An RDKit path-based fingerprint that captures substructure layering. |
| MAP4 | MinHash-encoded atom pairs with their surrounding circular substructures. |
Which one to use
Start with Consensus. It runs all four models and keeps, for each target, the highest score any of them gave, so a target only needs to be recognised by one fingerprint to surface. Each row on the results page shows which fingerprint produced its score.
Running the same compound through two fingerprints is a good way to gauge robustness: targets that appear in both rankings are more trustworthy.
5. Active vs. inactive predictions
Most target-prediction tools answer one question: which targets is my compound likely to bind? PPB4 also answers a second: which targets is it likely to be inactive against?
The two answers come from two separate networks: one trained on confirmed active interactions, one on confirmed inactive ones. They share the input fingerprint but have independent output layers over different, partly overlapping target sets, so their rankings can disagree on the same target.
6. Filtering by target type
ChEMBL targets cover more than single proteins: cell lines, whole organisms, protein complexes, protein families and protein–protein interactions are all included. The Target type dropdown above each results table lets you focus on one category at a time, and its counts reflect what is actually present in the current top-N list.
7. Similar known compounds
A prediction on its own does not tell you why the model made it. Every row in the results table carries a button that answers that question: it shows the compounds already annotated against that target in the training data, ranked by how similar they are to your query.
The panel has two tabs, Active neighbours and Inactive neighbours, so you can see both sides of the data for the same target: the compounds measured as active on it, and those measured as inactive. Both searches run when the panel opens, so switching between the tabs is immediate, and each tab shows how many compounds it found. A target with no known compounds on one side says so rather than showing an empty panel.
Each search is restricted to that one target and that one side of the data. Similarity is the Tanimoto coefficient on the same fingerprint used for the prediction, and the ten closest compounds are returned.
Each neighbour shows its ChEMBL ID, linked to its compound report card, its structure, and its similarity to your query. A prediction backed by a close neighbour rests on chemistry the model has seen before, and comparing the two tabs shows whether the close analogues around your compound were active, inactive, or a mixture of both. The Export CSV button exports the tab you are viewing.
8. Predicting multiple compounds
Switch to the Paste SMILES tab on the homepage to submit a batch. Enter one SMILES per line: comments, blank lines, and identifiers are not supported, so keep the input clean. Each compound becomes its own card on the results page, each with its own filters and tabs.
10. Running PPB4 locally
The whole application, including its Dockerfile, is available at MaedehDarsaraee/PPB4. The trained networks and the nearest-neighbour data are distributed separately because of their size, and are mounted into the container at run time.
11. Citing PPB4
- PPB4 preprint: in preparation
- PPB3 paper: J. Chem. Inf. Model. 2026, 66, 2466–2473
Questions, bug reports, and collaboration enquiries are welcome via the Contact page.