Frequently Asked Questions

  1. Compared to the old version, this version of SwissTargetPrediction returns different predicted targets for the same query molecule. Is this normal?
  2. What kind of predictions can I make with SwissTargetPrediction?
  3. How does SwissTargetPrediction work?
  4. Should I run SwissTargetPrediciton to predict if my small molecule is active?
  5. Can I use SwissTargetPrediction with large molecules (e.g., peptides, antibodies,...)?
  6. What about other targets not included in SwissTargetPrediction?
  7. What do the green probability bars represent in the result page?
  8. How are the targets ranked?
  9. The output returned proteins with a probability of 1.0. What does this mean?
  10. What does Target Class mean?
  11. How is the number of known actives in the result page chosen for each similarity measure?
  12. What do the similarity values mean and can they be compared between 2D and 3D similarity?
  13. What does "(by homology)" stand for in the list of predicted targets?
  14. Why are most predictions made "by homology" in non-human organisms?
  15. I know my molecule is experimentally defined as active on a protein in another organisms but I do not see its ortholog in my predictions?
  16. I cannot find the published target of my query molecule. How is that possible?
  17. Compared to the previous version of SwissTargetPrediction, the number of similar known active are almost always shorter. Why this?
  18. Should I use SwissTargetPrediction to monitor literature or patent?
  19. What are "known actives similar in 2D"?
  20. What are "known actives similar in 3D"?
  21. Why is it useful to combine different similarity measures?
  22. I don't get any output but a message telling "No similar actives found. No target predicted". What has happened exactly?
  23. MarvinJS sketcher is not behaving as expected, what can I do?
  24. Can I copy/paste a molecule?
  25. Is SwissTargetPrediction freely available for academic users?

Answers

Compared to the old version, this version of SwissTargetPrediction returns different predicted targets for the same query molecule. Is this normal?

Yes it is! The current version represents a major update of the target prediction engine. First, the model was retrained on ChEMBL23 (the old version was on ChEMBL16), which implies new molecules, new targets and new activities, but also more severe curation of data. Besides, similarity thresholds are more strict so that the prediction relies on a shorter list of more similar known actives. This comes from a thorough analysis where we investigated above what level of similarity the meaningful information is located (see Daina A. et al. for details). As a result, predictions are different from the old version. In our applicative experience, the predictions are very much relevant and useful in drug design/discovery contextes. But if in some specific cases, for whatever reason, one wants to rely on the old version, this latter will still be available for at least one year at old.swisstargetprediction.ch

What kind of predictions can I make with SwissTargetPrediction?

SwissTargetPrediction is an online tool to predict the macromolecular targets (proteins from human, mouse and rat) of bioactive small molecules. This is useful to understand the molecular mechanisms underlying a given phenotype or bioactivity, to rationalize possible side-effects, to predict off-targets and to assess the possibility of repurposing therapeutically-relevant compounds.

How does SwissTargetPrediction work?

SwissTargetPrediction is based on the so-called 'similarity principle', which generally states that two similar molecules are prone to have similar properties. In the context of SwissTargetPrediction, we have statistically quantify that two similar bioactive molecules are likely to share their protein targets. Therefore, for a query molecule, we identify the most similar molecules among a set of about 370'000 actives on known targets. The predicted targets are those having the actives displaying the highest similarity with the query molecule.

Should I run SwissTargetPrediciton to predict if my small molecule is active?

Certainly not! SwissTargetPrediction is valide to estimate the most probable targets of bioactive molecules.

Can I use SwissTargetPrediction with large molecules (e.g., peptides, antibodies,...)?

This is outside of the validity domain of SwissTargetPrediction, because the model was trained on small molecules only. Therefore, even if you are able to technically launch the computation (which is not certain), the prediction has to be taken with due caution and not over interpreted. For short peptides you may still find small molecules that are similar (possibly interesting peptido-mimetics). Antibodies should be all together avoided. Also, very small molecules (less than 8 heavy atoms) should be avoided because similarity measures for these molecule tend not to perform well.

What about other targets not included in SwissTargetPrediction?

Currently SwissTargetPrediction only includes protein or protein complexes for which at least one molecule has been experimentally observed to bind to, or homologs thereof. Therefore predictions are restricted to these targets. However, it should be noted that they already cover most of the main 'druggable targets'.

What do the green probability bars represent in the result page?

These values correspond to the probability for a query molecule, assumed as bioactive, to have the listed proteins as biotargets. It is important to note that these are NOT the probability of the query molecule to be active. It is of course very advisable to look carefully at the ligands of each predicted target before planning any follow-up actions.

How are the targets ranked?

Targets are ranked according to a score that combines both 2D and 3D similarity values with the most similar known active to the query molecule. Therefore the top ranking target is not necessarily the ones bound by the most similar ligand with any of the similarity measure (see Gfeller et al. for details). Importantly, the ranking of the targets rather than the absolute values of scores or probabilities, is the most meaningful parameter. A maximum of 100 probable protein targets can be displayed and sorted on a dynamic table, the default is 15.

The output returned proteins with a probability of 1.0. What does this mean?

Most probably the query molecule is part of the known actives screened. As such the similarity is total and the probability is absolute. This case stands for the retrieval of a known experimental bioactivity and not an actual prediction. However, it can be greatly informative.

What does Target Class mean?

In SwissTargetPrediction, the classes have been taken from ChEMBL. The 'Target Class' column in the result page shows the most specific target class. The pie chart indicating the repartition of Target Class among the predicted proteins can be displayed for the top 15, top 25, top 50 and all ranks. For predictions based on homology, the target classes have been mapped from the homologous proteins.

How is the number of known actives in the result page chosen for each similarity measure?

We use thresholds on each similarity values to decide which ligands are to be displayed for each predicted target. These thresholds are set at 0.85 for ES5D similarity (shape) and 0.65 for FP2 similarity (2D). Below such thresholds the known actives are not similar enough to be displayed.

What do the similarity values mean and can they be compared between 2D and 3D similarity?

For all actives used to predict a target, we provide the similarity with the query molecule. These values are the exact values as given by the algorithm used to quantify the similarity. However, they can NOT be compared between different similarity measures as the typical scale is completely different. Because this may vary between ligands and targets. So it is always the ranking that is meaningful, and not the absolute similarity values.

What does "(by homology)" stand for in the list of predicted targets?

A query molecule may have very similar known active that target a in another species than the one chosen in the input form. This is a strong indication that the molecule may also bind to orthologs of this target. We therefore include such predictions and mark them as "(by homology)".

Why are most predictions made "by homology" in non-human organisms?

The vast majority of experimental activities compiled to built our datasets involve human targets. Therefore, for any query molecule, you are much more likely to find a similar known actives on a human target with an ortholog in your chosen organism. This results in many predictions being done "by homology" for rat and mouse.

I know my molecule is experimentally defined as active on a protein in another organisms but I do not see its ortholog in my predictions?

We cannot exclude that some orthology relationships will be missed, e.g. because they are distant orthologs.

I cannot find the published target of my query molecule. How is that possible?

Even though the training set is several hundred of thousands of data points, these have been filtered with strict criteria. Also, the current status is as of ChEMBL 23, we have not update the list of known actives to screen afterwards. As such, the information of very recent publications cannot be included in the knowledge-based tool.

Compared to the previous version of SwissTargetPrediction, the number of similar known active are almost always shorter. Why this?

The dataset is now based on ChEMBL23, compared to ChEMBL17 for the old version of SwissTargetPrediction. Active molecules and targets are both more numerous, but strict thresholds of similarity (0.65 for 2D and 0.85 for 3D) were defined to focus on the most meaningful information and to reduce noise in the model training. As a result, the lists of known actives (the compounds that drive the prediction) are shorter because only containing highly similar (above thresholds) molecules. In our experience, we found that the shorter lists of more similar known actives (in 2D and in 3D) selected among a larger pool of screenable molecules, returned by SwissTargetPrediction 2019 compared to the old version, are more descriptive and informative for drug design/discovery applications.

Should I use SwissTargetPrediction to monitor literature or patent?

Certainly not. This tool is meant for predicting probable targets to support drug design/discovery. The prediction is based on ChEMBL 23 knowledge. We do not update. As such, the latter data on medicinal chemistry are not explicitly part of the backend.

What are "known actives similar in 2D"?

These are the molecules known as active on a given target, whose 2D chemical structure is highly similar to those of the query molecule. In SwissTargetPrediction, this 2D similarity is quantified by the Tanimoto coefficient between the fingerprints vectors (FP2) of both the query and the screened molecules.

What are "known actives similar in 3D"?

These are the molecules known as active on a given target, whose 3D shape together with the projection of charge and lipophilicity (5D) is highly similar to those of the query molecule. In SwissTargetPrediction, this 3D similarity is quantified by the Manhattan distance between the Electroshape vectors (ES5D) of 20 conformers for both the query and the screened molecules.

Why is it useful to combine different similarity measures?

Each similarity measure have pros and cons. For instance, in our work, we observed that by combining different similarity measures, the predictive capacity of screening is significantly improved, especially for druglike molecules.

I don't get any output but a message telling "No similar actives found. No target predicted". What has happened exactly?

To produce a list of probable targets, SwissTargetPredicition has to find at least one molecule similar either in 3D or in 2D to the query molecule. In the given case, the input molecule is under the thresholds of similarity to any of the known actives, it was screened against.

MarvinJS sketcher is not behaving as expected, what can I do?

Try to clean the cache memory and cookies of your browser and restart it. If not sufficient, please ensure that your internet connection allows binding to web services. Some private IP subnets and public WiFi accesses restrict binding to such protocols.

Can I copy/paste a molecule?

Yes. You can paste SMILES (ctrl-v) either into the sketcher (on the right) or in the text-box (on the left). Other standard molecular formats (e.g. mol, sdf, mol2, mrv, common name) can be pasted into the sketcher. If not working, it is possible to open a file in the sketcher (folder icon, second on the top).

Is SwissTargetPrediction freely available for academic users?

Yes, SwissTargetPrediction can be used without charges for non-profit research purposes.