Frequently asked questions

PPB4 (Polypharmacology Browser 4) is a web tool that predicts which protein targets a small molecule is likely to be active or inactive against. It uses multitask deep neural networks trained on bioactivity data from ChEMBL.

PPB4 covers roughly 8,000 ChEMBL targets and is developed by the Reymond Group at the University of Bern.

Three changes are worth noting:

• Confirmed-inactive prediction. PPB4 introduces a dedicated model that predicts inactivity, not just activity.
• Larger training set. PPB4 is trained on ~5.8 M compound–target interactions from ChEMBL 35, substantially more than PPB3.
• MAP4 fingerprint support. A MinHash atom-pair fingerprint better suited to peptides and macrocycles is now available alongside ECFP4, AtomPair, and Layered.

Yes. PPB4 is free for everyone to use through this web interface.

• PPB4 preprint: in preparation
• PPB3 paper: J. Chem. Inf. Model. 2026, 66, 2466–2473

Each fingerprint has its own multitask deep neural network. The network takes a 4096-bit molecular fingerprint as input and outputs one score per target. All targets are predicted simultaneously, which is why we call it multitask.

For each fingerprint type (ECFP4, AtomPair, Layered, MAP4), PPB4 actually runs two networks: one trained to predict confirmed activity, and one trained to predict confirmed inactivity. The two share the input but have separate output layers covering different target subsets.

The active model in PPB4 covers 7,552 targets; the inactive model covers 7,178. 5,905 targets appear in both, for 8,825 distinct targets overall.

The network produces one score for every target it covers, and the targets are ranked by that score from highest to lowest. Only the top N you asked for are shown.

The score is the raw output of the network's sigmoid layer, with values in the range 0 to 1. It should be interpreted as a relative ranking signal: higher means more likely, and the value reflects how strongly the model associates your compound with that target.

In the Merged predictions tab each target appears once, ranked by the higher of its two scores. Because the active and the inactive networks are trained separately, a target can be predicted by both; its row then carries both labels and both probabilities rather than only the stronger one. The two values come from different networks, so they are not on a common scale and should be compared with that in mind.

PPB4 is trained on bioactivity data from ChEMBL 35. After standardisation and filtering, the dataset contains roughly 5.8 M molecule–target interactions: 2.5 M actives across 7,551 targets and 3.3 M inactives across 7,177 targets, for 1,720,285 molecules and 8,824 targets in total. Molecules are limited to 80 heavy atoms, and a target is kept only if at least five molecules are annotated against it.

A molecule is active on a target when the reported measurement is at least 50 % effect, or a concentration of 10 µM or better (IC₅₀, K_i, K_d, EC₅₀ and the other standard ChEMBL activity types). Anything measured but weaker than that is inactive.

This is the main change from PPB3, which discarded the weak measurements. PPB4 keeps them and trains a second model on them, which is where the inactive predictions come from. A single threshold is used rather than the measured value, because PPB4 is meant to suggest targets worth testing, not to predict potency.

Only measured molecule–target pairs are used. A pair that was never tested is absent from both sets, not treated as inactive.

A target only appears in PPB4 if it had at least 5 unique compounds in the relevant class (active or inactive) after filtering. Targets below that threshold are dropped because there is not enough signal to learn from.

Every structure goes through the same three steps, in the training data and in your query alike: only the largest fragment is kept, so salts and counterions are removed; charges are neutralised; and the result is written as a non-isomeric canonical SMILES, so stereochemistry is not used by the models.

The training data is filtered further: molecules are limited to 80 heavy atoms, and a target is kept only if at least five molecules are annotated against it.

Draw a structure in the editor on the homepage or paste one or more SMILES, choose a model and how many targets to return, then click Predict targets. The Tutorial walks through each step and explains how to read the results.

Yes. The application and its Dockerfile are at MaedehDarsaraee/PPB4, and the trained models are downloaded separately and mounted into the container. Running predictions needs roughly 0.9 GB of data and 8 GB of RAM; the similar-compounds search needs up to 3.7 GB more. The README lists every file and the commands to start it.

The web interface comfortably handles batches up to a few dozen compounds.