Dear Kuang-Lung Hsueh,

Without target's 3d structure you can use ligand-based approaches to conduct virtual screening.

http://pubs.acs.org/doi/abs/10.1021/ci900419k?journalCode=jcisd8

Of course, known ligands of the target are required. You can check the existence of ligands in ChEMBL, for example.

There are plenty opportunites to conduct ligand-based virtual screening. You can buy specialized software, or use R-packages or Python-libraries to build a model and to train it, or use one of the web-applications developed for this purpose (PASS online, SWISS-Target, SEA, etc.).

Good luck!

If your protein is closely relate to another protein for which a structure is available, you can use a threading program to create a homology model and use it for virtual screening. Such an approach would, unfortunately, be even less reliable than virtual screening with the actual structure.  Ligand-based screening, as suggested by Pavel Pogodin, is an alternative. My preference would be to do an actual screen.

Dear Pavel V Pogodin, I have downloaded the reference. However, I had difficulties to ask the right question. Here is what I came up with, please let me know what kind of reference articles that I still need to learn. Thank you!

  1. For ligand-based screening, say we have only n ligands known, with possibly n unique sites on the protein, then the algorithm will not learn much from the ligands, am I right?

  2. Under the same condition, but only one single site, then the machine can learn to tell based on the ligands. However, I cannot stop thinking the binding site on the protein. Without structural information from the protein, what does the machine develop? They can deduce from the similarity and bioactivity from what we feed; then there is immense chemical space, how does virtual screen help us to get information of the potential new ligand?

  3. Also, do we need dynamics information from either ligand or the protein? Alternatively, the virtual screening of the software does not consider this?

Dear Kuang-Lung Hsueh,

1. You are absolutely right, in this case you can use molecular similarity to simply find the most similar compounds to the one of the known ones. However, usually it is not very interesting.

2. Well, the output depends on the input. Suppose, we provide the information on the chemical structures of known compounds and label for them ("active", i.e. this compound is the actual ligand of our target, "inactive" - compound does not interact with our target). Note, that information on the chemical structure is not chemical structure itself. We feed to the machine set of features (descriptors) which describe the compound's chemical structure (or its physico-chemical properties) with reasonable accuracy.

There are thousands of chemical descriptors and you can select the ones, that fit best to your goal.

To these data we apply some algorithm to elucidate the features, which are specific to the particularly labeled compounds (class).

After that, we feed to machine the representation of our novel compound(s). Machine analyzes and gives us a some score.

This score corresponds to probability for our compound(s) to belong to one of the classes. Here is our output in this very general and simplified case.

Note, that input may vary, it's only one example.

So, we don't use any data on the target's structure. But, implicitly, its structure projects on the structures of "actives" and "inactives". I mean, that "actives" have configuration that allows them to get in to binding site and their functional groups arraтged in the way that allows compound to be binded for some time. "Inactives" have not appropriate configuration or their functional groups are not arranged in an appropriate way, or both.

Chemical space is enormous and heterogeneous, it is an issue. Applicability domain of the model should be determined to adress this issue.

3. I'm not aware about ligand-based approaches that deal with the dynamics.

Dear Adam B Shapiro,

  I am not familiar with the topic, and thus, could you point out the threading program for the homology model building? Could you also point out how to increase the chance to get a better prediction? You seemed to mention that actual protein structure is better than ligand-based virtual screening than homology model of the protein for virtual screening. Also, did anyone of you or your team use CS-Rosetta that employs NMR chemical shift information? If the target protein of interest does not have a crystal structure, but NMR data or structure is available, what would you do?

As a matter of fact, a number of papers have been published in literature describing virtual screening of small-molecule ligands based on merely on the primary sequence of the target protein. My personal opinion is that this is not the right science. The success claimed in those studies are simply due to other reasons, e.g. homologous proteins of course have similar sequences. In QSAR studies, it is well-known that if A correlates with B, A is not necessarily the reason of B or vice versa.  Unfortunately this simple fact is ignored by so many researchers.

I am not the person to ask about the technical details of these computer programs, or virtual screening based on NMR structures. Perhaps others in the RG community can help.

My experience observing the use of structure-based virtual screening to try to enrich for active compounds in a library was that it had no advantage over random screening in terms of hit rate when screening novel targets. Therefore, I believe that any type of virtual screening that uses lower quality structural information (homology model, NMR structure) will not perform any better. In contrast, ligand-based screening should be enriching, although it is unlikely to find novel binding modes.

Dear Pavel V Pogodin,

  In summary, we need a list of ligands with active and inactive categories. Regarding the input, however, we use their physicochemical properties instead of their real chemical structure. One or series of algorithms will be employed to label these ligands into different classes. Moreover, then by feeding novel compounds and use the same type of algorithms, the machine will tell us how likely the novel compounds might be active or inactive. Therefore, the process indicates that the score is a representative, sometimes the simplified version of the docking process, this is also very dependent on how ligands were categorized, how ligands were sorted into classes and how the machine (algorithms) analyzed the novel compound. Do you agree or disagree?

Dear Kuang-Lung Hsueh,

I'm agree with you in general, but

1) "we use their physicochemical properties" Not neccessary, we can use another types of descriptors like fragment descriptors. You can read this book to know more

http://onlinelibrary.wiley.com/book/10.1002/9783527613106

Or read papers about molecular descriptors. I'm sure you will find a lot of info.

2) "sometimes the simplified version of the docking process" Approaches are very different. When you dock compound to target - you get the energy assessments, when you use some classifier to find ligand - you get the assessment of probability for compound to be ligand. But results (list of hits) may be the same. That I can tell, everything else is very case-specific and needs proofs.

3) "this is also very dependent on how ligands were categorized, how ligands were sorted into classes and how the machine (algorithms) analyzed the novel compound" True, and also it is depends on the descriptors selection.

You can read this book to know more about virtual screening

https://books.google.ru/books?hl=ru&lr=&id=0-lgYMb14JoC&oi=fnd&pg=PA1&dq=varnek+Tropsha&ots=F3k9pLC2Hg&sig=7YW1LN6CCDpBxsjWyGd4KaQsP5g&redir_esc=y#v=onepage&q=varnek%20Tropsha&f=false

And, as Adam Shapiro mentioned earlier, experimental testing is the best option and it is needs to be done to confirm the results of any type of virtual screening.

Good luck!