In majority of virtual screening studies, what we see is the poor or absence of correlation between the predicted scores and experimental activity of hits. Also self docking of compound rarely gives the accurate binding affinity in terms of score or Delta G. The above issue is very important to address, because in course of virtual screening , we may loose the better hits. Besides developing several scoring functions for protein-ligand interaction, why not the experts come together to develop more accurate scoring function.
The difficulty to make a correlation between scores and activity of hits is not only a question of docking. You have many other reasons which can explain this difficulty, particularly in vivo: Solubility of hits, Accessibility to the target like membrane permeability, better interaction with other proteins, modification of hits by enzymes or degradation. Even if you have a measure of affinity on your purified target, it is not always true that a good affinity will give a good activity ! A given ligand can bind your target without giving inhibitory effect. Your target is rigid, this is not the reality. You see, it is not a simple problem. Many people are trying to fit the scoring functions to find better hits, and it works sometime and the reason why it works better is sometimes obscure, even for the researcher who made the job...
A recent trend is using ligand efficiency indices e.g. it can be calculated as binding affinity/number of non-Hydrogen atoms. I've enclosed a link to an excellent review paper on this issue.
http://tigger.uic.edu/labs/caz/downloads/Abad-Zapatero_EODC_final_printed.pdf
Dear Vikash,
In my opinion, the aim of virtual screening is to retrieve the most probable active structures that will help at filtering chemical libraries to be screened in vitro. The final goal being finding hit compounds. In that case, you surely don't need an accurate quantification of the ligand-protein goodness of fit. This approach only needs a scoring function able to discriminate few actives from a bench of inactives. Anyway you surely couldn't afford a complex quantification as the number of structures to be docked will probably reach multiple hundreds of thousands of inputs.
Said that, there are many closely related activities in which an accurate determination of molecular recognition is central. In our group, it was demonstrated that computing delta G with implicit solvent models can greatly help the prediction of correct binding modes and energies by molecular docking. Among interesting articles on the subject, I can mention swissdock, a free and handy web-based docking tool relying on genetic algorithm and implicit solvation simulations.
Hope this may help. Antoine.
http://www.swissdock.ch
http://www.molecular-modelling.ch/drug-design.php
Article SwissDock, a protein-small molecule docking web service base...
This is a good question. I would say a qualified no, there is no strong need for much more accurate scoring functions that better represent binding affinity. The first reason lies in the thermodynamics. Binding energy is a function of dynamical behaviour as much as binding interactions and solvation/de-solvation so a scoring function that uses a single pose in a single protein structure to estimate binding affinity has strong limitations as to what it can represent.
The second reason lies in the practical usage. Not a complete no-no but something that needs to be considered. Such an improved scoring function would no doubt be calibrated on a wide range of very high quality protein structures. But it may be applied practically using protein structures that are much poorer quality. The scoring function would have to be robust enough to still perform well. Experience tells us that often the more accurately we try to parametrize a tool the weaker that tool performs in non-optimum conditions.
Of course there may still be worthwhile improvements to be made in existing scoring functions. But I recommend using multiple methods of assessing whether a docking pose (consensus methods, removal of bad geometries, etc etc) is viable rather than relying on a single docking score value. These methods can separate possible actives from structures unlikely to be active when properly applied, even with the scoring functions we have today.
John
In virtual screening the most challenging thing is to fight with a huge amount of false positives. The majority of current scoring functions are indeed capable of predicting binding affinities of ACTIVE compounds.
This is not surprising at all. If one simply compares the distribution of predicted binding affinity of millions of compounds from a commercial library with that of experimental binding affinities of active compounds, one may find that the majority of compounds would be predicted as ACTIVE. Since experimental binding affinities usually span just a few magnitude of order, any scoring function more or less can predict the binding affinity of ACTIVE compounds.
It's not so difficult to predict affinities of ACTIVE compounds, but remains very challenging to discern ACTIVE from INACTIVE.
If binding of compounds to a protein target follows the mechanism of induced fit, it would be better to use ensemble docking to reduce the chance of missing a good hit.
Hope this may help.
Article In silico structure-based screening of versatile P-glycoprot...
- Badges
- Science topic
- Similar topics
- Thesis Research