Usage and Strategy

Objectives

After studying the descriptive material in the HASL Strategy section and performing the later tutorials you should comfortable with the following operations:

1. Creating a HASL model for a series of molecules.
2. Predicitng activity for molecules not in the learning set.
3. Evaluating the partial activity profile with mapping tools.
4. Using the Multi-Conformer Protocol to discern conformations of bound ligands.

Methodology

Basic Steps for a HASL Calculation

1. Input and/or modify the input molecule "learning" set with known activities.

2. Calculate molecular lattices for each molecule in the data set.

3. Create the HASL model for the data set.

4. Analyze graphically the HASL model.

5. Predict activity for new molecules.

• Step 1: Defining the Input Molecule "Learning" Set

HASL is a Quantitiative Structure-Activity Relation (QSAR) technique that first requires a set of molecules from which the model is "learned". This is not dissimilar from CoMFA, another 3D QSAR method. These molecules must have known activity (hopefully measured on a common basis) and should be "aligned" so that they are superimposable on a common pharmacophore. Sybyl has a variety of tools for this purpose ranging from simple rigid fits, to flexible fits, to grid-based alignment tools. HASL expects the molecules to be available as rows in a Sybyl Molecular Spreadsheet (MSS). Usually the simplest way to get new molecules into the spreadsheet is by first building them and saving them in a Sybyl database. If you already have a spreadsheet with molecules and biological data, i.e., from a CoMFA experiment, that can be used by HASL without any problems.

The activity data seems to be best modeled if it is in the pK_i format. You can enter these data in a column cell-by-cell or you may wish to read an external file to fill the activity data column.

• Step 2: Calculating Molecular Lattices for Each Molecule

This is the critical step where you define the type of molecular property upon which the HASL model will eventually be built. This version of HASL offers up to six definitions for what we call "H-Val", the key property definition. The first two of these definitions are the internally defined parameters grouped as either the "Basic" or "Advanced" sets. These are simple integers representing the acid/base, hydrogen-bonding and/or hydrophobic character of the atoms. The "Basic" set is limited to one of +1/0/-1 while the "Advanced" set allows up to two integer values per atom type. This would be useful for atoms that may have two characteristics, e.g., hydrophobic and hydrogen-bond acceptor.

The third potential source of H-Val is the atomic charges for the molecules. In this case the charges are multiplied by the Gain Factor and rounded off to integers. It is important that the range of values for H-Val be maintained fairly small (-4 to 4?). So the choice of Gain Factor is important. It should be based on the estimated range of the partial atomic charges in the learning set molecules. A good starting point would seem to be about 2-5.

The fourth potential source of H-Val is available to HINT users -- the hydrophobic atom constants for the learning set molecules. The fifth and sixth sources are electrotopological state parameters and hydrogen electrotopological state parameters, respectively. These are available for Molconn-Z users. Again it is crucial that the Gain Factor be set to coincide with the atomistic values from these mathods.

You do have some control over how hydrogens are handled with these H-Val sources. It may be useful in some cases to use united atom models to reduce the size and complexity of the problem. In particular HINT parameters and the Advanced H-Val set may benefit from this approach.

Once the type of H-Val parameter is set, the next issue is to define paramaters specific to that type. The user is offered total control over the values in the Basic and Advanced H-Val parameter set. (The default set can be easily retrieved however!) The options for HINT and Molconn-Z should be familiar to users of those programs. The general idea with HASL is to offer a multitude of available options os that a variety of models can be explored.

Finally, the HASL region must be defined. By far the best method for the first run is to choose to "Pre-Calculate as Union..." which calculates the grid box into which all of the molecules in the learning set will fit and adds a margin around that. The other options are using a Sybyl region file previously saved and explicitly defining the grid box by coordinates. This last option is a bit tricky. On the other hand, using a region file is the best method if you are repeating a run on the same data set after adjusting a parameter or etc.

• Step 3: Creating HASL Models

Much like ComFA, HASL models can be created with cross-validation to obtain a measure of internal predictivity. The output of this calculation is, for the most part, only the cross-validated r². In HASL this step is termed Model Verification. To predict new molecules, etc. we also need to create the final, non-crossvalidated model. This is what we will also use to graphically analyze the results (Step 4). If the Verify step gives a poor result, it may be worthwhile to return to step 2 and try a different H-Val parameter set.
• Verifying With Crossvalidation

  To "Verify" a HASL model we use crossvalidation. This means that each molecule in the learning set is predicted from a model that did not include that molecule. The most conceptually simple, but most time consuming form of crossvalidation is "Leave-One-Out". Here, each molecule is predicted by a model constructed from all of the other molecules. Thus for crossvalidation using Leave-One-Out, n models will be constructed to predict n molecules. It is faster to use crossvalidation groups, e.g. if there are 40 molecules in the learning set, one can use 10 groups (of 4 molecules each) to Verify the model, but that may not give you the highest cv r². However, the point of crossvalidation is to determine whether a model is usable. It may be pragmatic to do a quicker form of crossvalidation while one is optimizing the model parameters, because crossvalidation with HASL can take a very long time.

  For this reason you should also consider how many Iterations to allow and what the actual Error Limit should be. For exploratory runs, 50 iterations may be sufficient to determine the value of a model. You can save the crossvalidation results (predicted activity for learning set molecules) in the spreadsheet for later examination. The crossvalidated r² will be reported to you in a dialog at the completion of the run. You can monitor the progress in the Sybyl text window. In particular the records "Remaining Rows:" list which rows have not yet been predicted.

  Note that the actual molecular HASL lattice data are stored in temporary files. This means that you cannot return to crossvalidate a HASL model if you have added a different HASL column to any spreadsheet in the interim. More detailed discussion of crossvalidation is available in the Sybyl manuals.

• The Final Model

  The final HASL model is one constructed using all of the molecules (rows) in the learning set. As this is a single calculation, as opposed to the multiple calculations in crossvalidation, it is worthwhile to choose fairly tight convergence criteria; the default vlaues of Iterations = 300 and Error Limit = 0.001 seem to work most of the time. You may save the results (which are predicted activities for the learning set molecules) in the spreadsheet if you wish, but these are always nearly identical to the input activities.

  What the HASL iteration does, actually, is distribute the activity to each HASL data point, which are actually four-dimensional: x, y, z and H-Val. The iteration optimizes these in such a way that when a molecule is placed in the HASL model lattice, the sum of data points activated by the atoms of the molecule is the activity of that molecule. For that reason the HASL models can predict activity very rapidly and the HASL activity maps are a unique synthesis of structure and activity.

• Step 4: Graphically Analyzing the Results

  The HASL contours display the relative spatial distribution of partial activity for the model. From this one can determine which regions of the molecule are particularly relevent for increasing activity, and which areas impact unfavorably on activity. Operationally, you choose a map type by H-Val selection and the HASL program selects the HASL points that meet that criterion and prepares a Sybyl format contour map of partial activity as a function of x, y and z that can be displayed graphically within Sybyl. The available criteria are based on whether the H-Val is greater than, equal to or less than zero. In addition there are a full set of combinations, i.e., less than or equal to, etc. This is a new feature of this version of HASL, so we at present have little advice to offer, other than the three simplest choices, greater than, equal to, and less than zero, approximately correspond with looking at basic, hydrophobic and acidic regions, respectively.

  Choosing appropriate contour levels for a specific map is really a matter of taste, but in general we have found that about 1/3 of the range (for both negative and positive contours) is a good starting point. A histogram is output to the Sybyl text window before the contour selection dialog appears to help in this selection. It is important to use the same value set on each map if you are trying to make visual conclusions from contours; i.e., if you contour the H-Val .gt. 0 map at +.01 and -.01, it would be confusing to choose other values for the other maps if you are going to directly compare them.

  In general regions enclosed by large volume contours are representative of a larger activity contribution, either positive or negative. From this one can visually determine a structure-activity relationship for the learning set molecules. The next step, obviously, is to apply that relationship to predict and understand molecules outside the learning set.

• Step 5: Predicting New Molecules

  The prediction step is very simple. Using the "final" model, one or more new molecules, which are spatially aligned with the learning set molecules in terms of their pharmacophores, etc., are sampled by HASL and their activities are predicted. The difficulties are actually in the alignment step and making sure that the molecule(s) to be tested are not too different from the learning set. This means that molecules having functional or other groups extending into regions not sampled by the learning set or molecules having chemical functionalities not sampled by the learning set will not be particularly well predicted.

  These are exactly the same problems plagueing CoMFA and other 3D QSAR methods.

  From this data and the visualization offered by the HASL contour maps it should be possible to develop an understanding of the relationship between structure and activity and perhaps design new agents with enhanced properties.

Basic Steps for the Multi-Conformer Protocol

1. Create a collection of ligands, each with potentially with a multiplicity of conformers.

2. Calculate molecular lattices for each ligand and conformer in the data set.

3. Create multiple cross-validated HASL models for randomly selected subsets of the ligands and confomers data set.

4. Analyze the results of the randomly constructed and optimized models to determine the optimally matched conformer for each ligand.

5. Create a HASL model for this ligand/conformer set.

• Step 1: Creating the MCP Ligand/Conformer Set

As described above, the HASL methodology requires a set of molecules from which the model is "learned". However, in using the Multi-Conformer Protocol, the object is to learn which conformer for each ligand are the likely bound conformers. In other words, when there is a set of conformers for each of a collection of ligands, which conformer set (one for each ligand) produces the best 3D QSAR model, and is presumably correct.

The sets of conformers can come from ligand docking or other methods. The MCP is designed to ultimately deal with a database for which each conformer is named "ligand_n" where "ligand" is the name of the ligand and "n" is a sequence number for the conformer, e.g., 001, 002, etc. The current implementation of MCP has been designed to easily interface with the output from the Sybyl FlexX module. The Create MCP Database/MSS command will guide the user through building the appropriate database and spreadsheet and entering the activity data into the correct spreadsheet cells. See Step 1 above for more information on activity data.

• Step 2: Calculating Molecular Lattices for Each Ligand/Conformer

This is the same as Step 2 of Basic Steps for a HASL Calculation, above. Of course, now there are more rows in the spreadsheet.
• Step 3: Create Multiple Cross-Validated HASL Models for Subsets of Ligands/Conformers

Since it would be impossible to productively examine each possible combination of conformers for each ligand in HASL models, the approach used in the multi-conformer protocol is to randomly examine subsets of ligands, and optimize the conformers for these subsets, while maintaining a record of the accuracy of these models. Then the a model including all of the most optimum conformers can be constructed. In this step, the parameters for the stochastic search process are set. In general the idea is to cover as amny combinations of liagnds and conformers as possible, in as short a time as possible. The MCP has many features to optimally select conformers, by "remembering" the success of that conformer in previous models. Nonetheless, the MCP is a fairly slow operation and may require a few days to complete even on a relatively fast processor.

Another feature available at this step is that certain ligands can be "forced" into each model. This would perhaps be attractive if those ligands were particularly well characterized or known to be effective binder(s). Also, for "forced" ligands there is the option of "freezing" certain conformers if, for example, there was an x-ray crystal structure for that ligand bound in the receptor site.

The MCP ends by one of three conditions to be set in this step: a) total number of unique models examined, b) minimum occurrence of each ligand in a specified number of unique models, or c) the cross-validated r² of all-lignad HASL models crosses a specified threshhold. (The all-ligand models are calculated periodically each time the counter for condition b is incremented.) While the running statistics of these three conditions are importnt for determing the course of the MCP run, the results for each calculated model are tabulated for the analysis of step 4.

• Step 4: Analyze MCP Results to Determine Best-Matched Conformers

The Analyze function exams the record of completed models and calculates metrics that rank each conformer's performance in these models based on three observables: cross-validated r² of models including the conformer, average error of ligands in models including the conformer and the slope (predicted vs. actual) of models including the conformer. A fourth metric, termed the AMD Score, combines the other three metrics. In addition the Analyze function calculates a rank (for each metric) of the set of conformer for each ligand so that the "best" conformer can easily be identified.

These data can be added to the Molecular SpreadSheet as columns that can be used as sorting keys. If, for example, you sort on the AMD Score Rank, then all of the number one ranked conformers will be accumulated in rows at the top of the spreadsheet. Presumably these are the set of conformations "most" consistent with binding at the site.

• Step 5: Create a HASL Model for a Ligand/Conformer Set

At this point you can perform steps 3 and 4 of Basic Steps for a HASL Calculation, above, to calculate a 3D QSAR of the conformer set. The best starting point would be the number one ranked conformers, but examination of some of the other well-ranked conformers may be prudent.

Certainly with a validated overlap of conformers, and 3D QSAR models of activity, it should be possible to begin a design phase that would further optimize the structure of therapeutically interesting inhibitors or ligands.