PASS Targets

is designed to predict which protein targets a given drug-like organic compound may interact with using its 2D structure. The PASS Targets model was trained on public compound–target interaction data extracted from ChEMBL and additionally filtered/curated by us to reduce noise and contradictions in database records. As a result, the system leverages SAR knowledge for 589,107 distinct compounds and can predict interactions with 2,507 protein targets from different organisms. Prediction accuracy, validated by leave-one-out and 20-fold cross-validation, reaches an average AUC ROC of ~96–97% on the training set and ~90% on an external test set of known drugs across 206 targets, confirming the model's reliability for real-world applications.

An important evolution of the methodology was presented in the 2019 study, which addressed "sample selection bias" in the experimental testing of chemical compounds. The authors demonstrated that public database records are heavily biased because compounds are usually selected for testing based on prior knowledge rather than at random.

Using the PASS algorithm, researchers were able to accurately predict whether a given compound had ever been tested against a specific target. Crucially, if the algorithm "recognizes" a compound as one that would likely be tested (meaning it falls within the model's applicability domain), the prediction accuracy for its actual activity is significantly higher (an average ROC AUC of ~0.87 compared to ~0.75 for compounds outside this domain).

How it works?

In PASS Targets, a compound’s structure is encoded as a set of 2D descriptors called MNA (Multilevel Neighbourhoods of Atom), and a Naive Bayes–type approach (as implemented in PASS) is used to compute probabilistic interaction estimates for each target in the training set. A key distinction of PASS Targets versus “classic” PASS is that it predicts the fact of interaction with a protein target rather than a specific mechanism/type of action (e.g., inhibition), which enables broader use of heterogeneous public data and reduces bias from assay-condition differences. You submit the compound’s structural formula in the web interface, and the service returns a predicted interaction spectrum: a list of protein targets with probabilistic scores for each target. In PASS terminology, results are typically interpreted via probabilities of being “active” (Pa) and “inactive” (Pi) for each target; in practice, users often rank targets by Pa–Pi and/or focus on top targets with the highest confidence.

Practical use

PASS Targets is designed to support several concrete scenarios in early-stage drug research. First, it can identify the most probable molecular targets for a newly synthesized or repurposed compound, allowing researchers to narrow down which in vitro assays to run rather than testing blindly across thousands of proteins. Second, the tool can be used in reverse. It can start with a known protein target and search a library of compounds to find those predicted to interact with it. This effectively acts as a virtual screening filter. Third, predicted off-target interaction profiles can flag potential side effects or toxicity risks before costly preclinical experiments: for example, the we demonstrated this by predicting non-kinase off-targets of 28,000 submicromolar kinase inhibitors, revealing candidates relevant to pathogens such as Plasmodium falciparum and Mycobacterium tuberculosis. Finally, PASS Targets supports drug repurposing workflows, where the full predicted interaction spectrum for a known drug can uncover new therapeutic applications that were not originally intended.

The algorithm's predictive performance is highly robust: the average ROC AUC is roughly 96-97% in cross-validation procedures, and around 90% when evaluated on independent external test sets of known drugs.