Rapid interpretation of patent SAR using Forge

Biological data is now a regular feature of new patent applications and this is readily available for download from Bindingdb which has data on over 2,500 patents encompassing more than 300,000 binding measurements. Generating meaningful insights to this data is perceived as less straightforward. In this post I will use Forge™ V10.6 to demonstrate that it is possible to get an overview of the SAR from a single patent entry with minimal human intervention and time.

Application to PIM-1

Selection and processing of 288 compounds from US9321756, ‘Azole compounds as PIM inhibitors’ (detailed in Appendix I) gave the Activity Atlas™ model shown in Figure 1. The total time to generate and interpret this model was around 30 minutes. It would be relatively straightforward to automate the process.

Figure 1: Activity Atlas model generated in this case study. From data download to model took 30 minutes. The ‘Activity Cliff Summary of Electrostatics’ and ‘Activity Cliff Summary of Shape’ views are shown. These detail regions of acute SAR – Red / Blue = positive / negative electrostatics preferred for greater activity; Green / Pink = activity favors /disfavors atoms in this region.

SAR interpretation

Firstly, the oxadiazole is clearly required as demonstrated in Figure 2 by region of negative (blue) next to both nitrogen atoms and representing the interaction of this group with the side chain of Lys67. Perhaps this is not surprising given the title of the patent application. The model also shows that the amino group next to the oxadiazole is constrained (area of pink surface).

Figure 2: Activity Atlas model close to the oxadizole group. Red = positive electrostatics preferred; Blue = negative electrostatics preferred; Green = Atoms in this region favored; Pink = Atoms in this region disfavored.

On initial inspection there appears to be space in the protein to accommodate a substituent on the nitrogen. However, by viewing the aligned ligands in the context of the protein and showing contacts in Forge, Figure 3 shows it is clear that all N-substituted ligands clash with Asp186 and that the adjacent space is not accessible from this position in the ligand.

Figure 3: Clash of a ligand with a morpholino substituent to Asp186 (orange lines).

The model (Figure 4) shows that there is a clear preference for molecules that extend into the gap between the two arms of the ligand (green surface at the bottom of the model above). Whilst we would want to check the underlying data, the suggestion is that substitution on either R-group is tolerated. Indeed, the most active compound crosses this gap completely which raises the possibility of using a cyclized ligand.

Figure 4: A high active from the patent displayed in CPK. The N-trifluoroethyl group touches the cyclopropyl substituent on the opposite side of the molecule.

Surrounding the green (favorable volume) region between the two arms is large area of red surface. This suggests that positive electrostatics – edges of aromatics or H-bond donors etc. – is preferred in this region.
This summary is reinforced by looking at the individual compounds that make up the data, thankfully this is easy to do with the Activity Miner module of Forge. Using Activity Miner’s top pairs table (Figure 5) there are many pairs of molecules where introduction of a positive charge in the region below (as shown in the pictures) the ligand generates a more active molecule. Generally the difference is around 1 unit better activity for the charged species.

Figure 5: The top pairs table in the Activity Miner module of Forge showing a specific pair of molecules and the electrostatic difference map between them. Red regions indicate where that ligand in more positive than the comparator; Blue where that ligand is more negative. In this case the ligand on the left is over a log unit more active and contains a positive charge in the region at the bottom of the picture.

Looking at the protein structure does not reveal a specific interaction or reason for this gain in potency. However, by using the protein field surface in Flare, we can see that the protein is generating a negative potential in this region which would account for the gain in activity when introducing a positive charge.

Figure 6: The protein interaction potential contoured at 2kcal/mol, Red = positive; Blue = negative. The potential indicates the nature of atoms that to use in a region, positive atoms fit well in negative regions etc.

Lastly, in the region of the pyrimidine group the model has a large area of blue. This indicates that there is a clear preference for molecules with nitrogen atoms in the ring at these points (e.g., pyrazine). This area points towards solvent and hence this is quite surprising. From the crystal structure alone it would be expected that introduction of heteroatoms would have little effect on activity. Examination of the data using Activity Miner confirms that, for example, pyrazine is more active than pyridine. In this case the protein fields do not reveal anything significant in the underlying potential of the protein and we are left to speculate at the reason for the SAR.

Figure 7: PDB 4TY1 showing the region around the pyrimidine group of the ligand. There are few interactions between the protein and the edge of the ligand in this region.

Speculating that protein movement was at the root of the observed SAR, I downloaded into Flare all the PIM-1 structures from the PDB, sequence aligned them and superposed based on the sequence alignment. Looking at this region across the 150+ structures show no clear case for protein flexibility although a number of structures do have a water molecule in this region that would bridge the ligand to the side chain of Arg122.

Figure 8: Over 150 PIM-1 crystal structures superposed in Flare. The backbone is shown in tube, residues close to the depicted ligand of structure 4TY1 are shown in thin sticks. Only two structures have any variation in loop conformation in this region.

The reason for the observed SAR remains elusive and could be a function of protein-protein interaction, water mediated interaction or something else.


Rapid interpretation of Bindingdb patent data can be achieved using Forge. In this case the SAR of 288 ligands was condensed to a single Activity Atlas model in less than 30 minutes. Interpretation of the model over the next 30 minutes generated clear SAR insights that could be employed on competing projects. Inspecting the protein electrostatics using Flare provided further insights into the observed SAR.

Try Forge on your project

Request a free evaluation of Forge to try this on your data or condense a patent into a simple summary of the published SAR.

See all licensing options for Forge.

Appendix I

Background computational details

The raw data was downloaded in tab separated format from Bindingdb and pre-processed in Excel. The raw data contains data for two biological targets – ‘PIM’ and ‘PIM-1’. Compounds with ‘PIM-1’ data were selected and checked for duplicate values. One compound was excluded because of a large variation in the reported IC50 value and four molecules were excluded due to missing activity values. All other duplicate IC50 values were averaged and converted to a pIC50 value resulting in a dataset of 288 molecules in a csv file.

The original dataset included the ligands of PDB codes 4TY1 and 4WT6. The protein-ligand complexes were downloaded into Flare, sequence aligned and superposed. Looking at the binding site, either ligand would work well as a reference for initial alignment of the dataset. The ligand from 4WT6 was chosen for further experiments and both ligand and corresponding protein transferred to Forge (Copy-Paste). The csv file was loaded into Forge (Training Set) and the molecules processed using Accurate but Slow conformation hunting, Substructure alignment and an Activity Atlas model built.

The Forge processing window showing the options used in this case study.

Using the Cresset Engine Broker, the calculation took 15 minutes to complete. Examining the results shows excellent alignment through the common substructure but some variation beyond that.

288 aligned ligands from US9321756 that were used to prepare the Activity Atlas model.


About Activity Atlas

Activity Atlas models are created by comparing all pairs of molecules in terms of positive and negative electrostatics plus the hydrophobics and shape properties and then combining these together, weighted by the change in activity for the pair. The result is a simple, qualitative picture of the critical points in the SAR landscape.

The resulting Activity Atlas model was automatically displayed. I always start with the ‘Activity Cliff Summary of Electrostatics’ and ‘Activity Cliff Summary of Shape’ views to understand the data. As this was a quick experiment and the alignments were noisier than a fully curated experiment, the Activity Atlas model is also noisier than ideal. However, by increasing the Confidence Level to 3.0 concentrates on the clear signals in the data.

The display options used for the Activity Atlas models shown in this study.

Model validation

Activity Atlas is a qualitative technique and hence difficult to validate except through manual inspection. However, Forge is capable of building quantitative models that can be used to validate the alignment of the molecules (we believe that consistent alignment is the single biggest factor in generating reliable 3D QSAR models). Using the Automatic regression model building methods of Forge with a 20% activity stratified test set generated an SVM model with q2 0.64 (LOO) and an r2 on the independent Test set of 0.62. Given the noisy nature of the input data I believe this represents a good model and that the alignments are valid.

Comparing Forge’s command line utility to Blaze – which one should you use?

Here at Cresset we’re very interested in ligand-based virtual screening – it’s been a focus of the company ever since we started more than seventeen years ago. In that time there have been many advances and refinements of the techniques for both ligand-based virtual screening and structure-based methods. We have stuck by our fundamental principle that ligand similarity based on both electrostatics and shape is an excellent way to sort the wheat from the chaff. The results obtained by our services division, who have run more than 200 virtual screening campaigns with a better than 80% success rate, is testament to that.

Difference between falign and Blaze

One of the things our customers ask from time to time is which application should they be using to do virtual screening. The simple answer is that there are two, Forge (and its command-line utility ‘falign’) and Blaze, and the differences are readily apparent.

In falign, you can generate conformations for a large set of molecules, align them to one or more references, and rank them by the similarity score. You also have the option to bias the alignments and scores by adding field constraints, pharmacophore constraints, and protein excluded volumes.

By way of contrast, in Blaze, you can generate conformations for a large set of molecules, align them to one or more references, rank them by the similarity score, and… ok, point taken. So, given that falign and Blaze apparently do the same thing…

Why falign and Blaze?

The answer is scale. As anyone who’s ever played with large data sets knows, doing calculations on a few hundred compounds is fundamentally different to doing them on tens of millions of compounds. Once you are working at large scale, seemingly trivial operations such as filtering data sets become much more difficult if you want to be efficient. Blaze was designed from the ground up to work with large data sets of 107 molecules and more, with an emphasis on maximizing throughput on a computational cluster. Forge/falign on the other hand are much more aimed at small-scale work, enabling simple screening or analysis of relatively small sets of compounds where the big iron of Blaze is overkill.

Data preparation

As an example, let’s look at the preparation of the data set in the two software suites. In falign, this is relatively simple: you provide the compounds to falign in 2D or 3D form, it assigns protonation states as necessary, and computes conformations on-the-fly if required before aligning to the query:

Falign has a secondary mode for use when aligning structurally-related compounds, which ensures that the common substructure within the dataset is perfectly matched:

Blaze, on the other hand, is much more sophisticated in its conformer handling. The average user of Blaze has multiple data sets that they want to screen (in-house compounds, vendor screening compounds, virtual libraries, custom collections), and these often have significant overlap. In addition, these data sets are usually reused multiple times for multiple virtual screens. As a result, Blaze has a sophisticated deduplication and precomputation pipeline that maximizes computational efficiency. The Blaze workflow looks more like this:

Any given chemical structure is only present once within Blaze: it may have multiple different names, and be present in multiple collections, but we’ll only precompute its conformations once and we’ll only align it once in any given screen. The conformer computation pipeline is heavily optimized for performance: we’ve done extensive studies on our conformer generation algorithm XedeX to find the optimal trade-off between conformation space coverage, rejection of higher-energy conformations, calculation time and number of conformations required. In addition, we’ve developed a special-purpose file format that is highly compressed (less than 13 bytes per atom on average, including coordinates, atom types, element, charge and formal charge) while being unbelievably fast to parse.

Blaze has a multiple-step pipeline to filter the data set, so that the full 3D electrostatic shape and alignment algorithm is only applied to molecules that are likely to have a high score. For extremely large data sets there’s an initial filter by FieldPrint, an alignment-free fingerprint method that gives a crude measure of electrostatic similarity. The molecules that pass the filter then go into an ultrafast version of our 3D alignment and similarity algorithm, and the full similarity algorithm is applied only to the best 10% or so of these. As a result, Blaze can chew through millions of molecules very quickly on even a modest cluster. The processing capability of Blaze is further enhanced by the fact that there’s a GPU version which is even faster.

Small versus large data sets

So, falign is designed for the simple use case on small sets of molecules, while Blaze is aimed at maximum computational and I/O efficiency on very large data sets. There is another important difference between the two. As anyone who’s been in charge of maintaining a virtual screening system knows, keeping it up to date is often a painful and thankless task. It’s bad enough keeping up with the weekly additions to the internal compound collection but keeping track of updates to external vendor’s collections is difficult: not only are new compounds being added but old ones are being retired. Blaze makes handling this situation easy. You simply provide Blaze with the new set of compounds that you want to be in the collection, and Blaze will automatically handle the update.

Any new compounds will be added, no-longer-available compounds will be marked and removed from the screening process, and any unchanged compounds will be left alone. This is far more computationally efficient than fully rebuilding the conformations for everything. Blaze can even be directly connected to your internal compound database, so that the Blaze collection holding your in-house compounds is always right up to date.

Given how great Blaze is at handling virtual screening, why would you ever want to use falign?

Blaze is optimized for throughput and computational efficiency, but the downside of this is latency. If you have a set of compounds you want to align and score in Blaze, you have to upload them, wait for Blaze to process them and build the conformations, wait for Blaze to build its indices, initiate a search, and wait for it to be submitted to your cluster queueing system. There’s five- or ten-minute’s latency in all of this, which is fine for a million molecules but is overkill if you have only one hundred. Falign, by contrast, will start work straight away on your local machine with no waiting at all.

The answer to the falign vs Blaze question, then, is largely a question of scale. Got a dataset of a million molecules that you want to run repeated virtual screens against? Blaze is just the ticket. Got a small set of compounds that you want to align and score as a one-off? Forge and falign are just what you need. For our in-house work we tend to find that the tipping point occurs around a few thousand molecules. Falign can easily chew through this many in an hour or so (especially if plugged into your computing cluster using the Cresset Engine Broker). However, if there’s more compounds than this or we’re going to want to run multiple queries, then Blaze it is. Since Blaze is accessible through the Forge front end, and both are accessible through KNIME and Pipeline Pilot, it’s as easy as pie to pick the right tool for the job.

Try on your project

Request a free evaluation to try out Forge or Blaze for your small or large-scale virtual screening needs. Don’t have a cluster? Blaze Cloud and Blaze AWS provide simple ways to access cloud resources to do the number crunching for you.


Forge Design: New name, familiar environment

Forge Design is a new licensing level of Forge™ for medicinal and synthetic chemists. It replaces Torch™ and benefits from the familiar GUI, but with V10.6 enhancements.

What can Torch users expect from Forge Design?

The new graphics engine generates enhanced 3D objects, thus delivering strong performance, improved pictures and new smooth transitions between storyboard scenes (Figure 1).

Figure 1: The new graphics engine in Forge Design generates great pictures in an enhanced GUI providing strong performance, faster calculations, improved 2D display of molecules and smooth transitions between storyboard scenes.

For larger projects the GUI will be more responsive, with improved performance on common operations such as application of filters, calculation and interaction with custom plots, exporting data. Activity Miner users will experience faster calculations which are less memory intensive.

The 2D display of molecules has been improved to make it clearer and more appealing.

New functionality

Forge Design has a significant number of new features and improvements compared to Torch V10.5. For example:

  • If you are working on a large project, there is a new function to automatically assign the selected molecules to roles, based on their Murcko scaffold
  • The Filters window includes a green/red toggle to control whether each filter is enabled or disabled, and pre-defined structural filters (for example, for groups like Ring, Aromatic Ring, H-bond donor and H-bond acceptor)
  • There is a new function to show the chosen field surfaces as a difference between the two molecules in the 3D display.

Figure 2: New functions in Forge Design include an option to automatically assign molecule to roles based on their Murcko scaffold, improved filters and the new ‘Field Difference’ button to show the chosen field surfaces as a difference between the two molecules in the 3D display. In this case the mono-fluoro derivative on the right is more positive (red) where the F is changed to H but also has a more negative aromatic ring.

  • Improved Molecule Editor, enabling you to conformation hunt and align all the molecules created during the same editing session as you exit the editor
  • Improved Blaze™ results window, which now shows an enrichment plot and statistics for each Blaze refinement level
  • The ideal radial plot profile for your project and the custom settings used for the conformation hunt and alignment can be shared with your team using the new import/export functions
  • Improved display of bonds in the 3D window, greater readability of constraints labels, more intuitive display of interactions between the reference molecules and the protein
  • New function to clip only the display of the protein leaving the ligands untouched
  • New functions to choose column content as the molecule title, or as a label in the 3D window
  • Improved plots now showing a regression line for selected molecules
  • For Activity Miner users, the Disparity matrix can now be filtered by Similarity, Disparity and Δ Activity; there is also a new ‘Find Molecule’ function. Also, you can now tag all the molecules visible in the Activity View in the Forge Molecules table.

Figure 3: In the Forge 10.6 Activity Miner window, the disparity matrix can be filtered by Similarity, Disparity and Δ Activity; molecules that do not pass the filter(s) are shown in gray.

How does Forge Design compare to Torch and Forge?

Figure 4 below shows the modules which are common between Forge and Forge Design (red), and the optional modules in Forge only (blue).

Figure 4: Modules available in Forge only (blue), and modules available in Forge and Forge Design (red).

Forge Design uses wizards for common operations just as with the full Forge package.  However, the wizards for building activity models and pharmacophores will be greyed out, as these are optional in Forge Design. The processing window will look slightly different (Figure 5 – right), with the optional Build Model section greyed out.

Figure 5: The wizard and functions available only in full Forge are greyed out in Forge Design.

You will find new, pre-defined roles in the Molecules table for training, test and prediction sets. In Forge Design, these roles have no special meaning and you can use them as any other user-created role or ignore them completely.

Simplicity and integration

Having a streamlined platform for ligand-based software makes it easier for existing Forge and Torch users to upgrade their installation to the newer release, with a single installer for both Forge and Forge Design.

This solution also makes it easier to upgrade Forge Design with additional functionality (for example, Activity Miner or the model building package), if desired.

This is a further step towards the integration of all Cresset ligand-based and structure-based functionality and simplifies the product installation and distribution for most customers.

Try Forge Design

If you are an existing Torch user, we will be in contact soon with more information on Forge Design. If you don’t have Forge or Forge Design, contact us to learn more.

Forge V10.6: Choose the molecules to make, and understand why you should make them


I am delighted to announce the availability of Forge™ V10.6, our powerful computational chemistry suite for understanding structure-activity relationship (SAR) and new molecule design. The focus of this release is on new and improved methods to generate robust Quantitative Structure-Activity Relationship (QSAR) models with strong predictive ability.

Choose the molecules to make next

Project chemists generally know which molecules they can make with a reasonably good chance of them being active. They often have too many clever ideas and are looking for ways of filtering and prioritizing lists of tangible compounds, arrays and small libraries.

Having a predictive QSAR model is a terrific way of doing this – you send your molecules into the model and get immediate feedback on whether making a compound is a good or bad idea.

However, getting a robust, predictive QSAR model is not always straightforward, and this is still a pain point for many of our users. You need a training data set of reasonable size, good activity data (e.g., pKi, pIC50) spanning a sufficiently large range, good descriptors and good modeling algorithms.

While we can’t help with the need of having a training data set of reasonable size and spread of activity, we can help with the rest.

The new Machine Learning (ML) methods in Forge, namely Support Vector Machines (SVM), Relevance Vector Machines (RVM) and Random Forests (RF) significantly expand the range of available QSAR model building options beyond the previous Field QSAR and k-Nearest Neighbors (kNN) regression options (Figure 1). Having access to a panel of well known, robust statistical tools gives you more opportunities to build a predictive model useful in project work.

Figure 1. The new Machine Learning methods significantly expand the range of QSAR model building options in Forge V10.6.

What about the descriptors?

Forge 3D electrostatic (based on Cresset’s XED force field) and volume descriptors are relevant for molecular recognition, and accordingly work very well for modeling activity and selectivity. These are used by Field QSAR and the new ML methods, while kNN can use either 3D electrostatic/shape or 2D fingerprint similarity.

New methods in action in a practical example

For this experiment, I have re-used an aligned data set of Orexin 2 Receptor ligands from the US patent literature,1 which I previously presented in a case study on Activity Atlas™, a method in Forge for qualitatively summarizing the SAR for a series into a visual 3D model.

I split the 377 Orexin 2 ligands into two subsets: a training set of 302 compounds which I used to build the QSAR models, and a test set of 75 molecules which were used solely to assess their predictive ability.

Figure 2 shows the results obtained with Field QSAR and the ML methods in generating predictive models for OX2R pKi. Field QSAR, kNN and RF models were built using default conditions; for SVM and RVM, Forge suggested a fine tuning of the model building conditions as the training set is large.

Figure 2. Performance of Field QSAR and ML methods on the Orexin 2 data set. Training set = 302 molecules used to build the models. Test set = 75 additional molecules used solely to assess the predictive activity of the models.

‘r2 Training Set’ is used to check the ability of each model to fit the data in the training set. It ranges from 1 (perfect fit) to 0 (no fit). From Figure 1, I can see that all models (except kNN in this case) give excellent results in fitting. However, this is hardly surprising as ML methods are well known for their ability to fit data of any type.

A more realistic check of the quality of the model comes from ‘r2 Training Set CV’. In cross-validation (CV), a part of the compounds in the training set is temporarily excluded from the model and the remaining compounds are used to build a model which is then used to predict the activity of the excluded compounds. Not surprisingly, ‘r2 Training Set CV’ is always lower than ‘r2 Training Set’, but the results for Field QSAR, RF, RVM and especially SVM are still good (kNN does not calculate this statistics).

Finally, ‘r2 Test Set’ gives an idea as realistic as possible of the performance of the model in real project work, as the model is asked to predict the activity of compounds it has never seen before. Most methods give reasonably good results, with SVM clearly outperforming the other methods with a more than respectable r2 test set = 0.59.

In a real project, I would not hesitate to choose SVM for filtering and prioritizing my list of ‘to-make’ compounds, with confidence that this is the best predictive power I can get for this specific data set.

What about kNN? It didn’t work very well on this data set; does it mean that it is not a good method? Not really. kNN is a robust, well known method particularly useful when working with multiple compound series, or with biological data which are derived from different sources. The fact that it didn’t work particularly well in this case does not exclude good performance in other projects.

This is the whole point of having several model building methods available: you can choose the one which gives best performance in your specific project.

If you think it must have been boring to calculate all these models separately, then I have good news: you don’t really have to. The default option in Forge is to automatically run all the ML models and pick the best one for you (Figure 3).

Figure 3. The Automatic model building option in Forge runs all the available ML methods and picks the best model for the output.

Understand why you should make the molecules you have chosen

A significant part of a project chemist’s work is to design the next generation of active molecules. To achieve this, you need to understand what are the features which make some compounds active, and which are those that undermine the activity in others. In other words, you need to interpret the model.

Unfortunately, ML algorithms won’t help you here: they are complicated equations which cannot be easily translated back to 3D in terms of ligand-protein interactions.

Luckily, Forge provides you with two additional tools: Field QSAR 3D views and the Activity Cliffs Summary in Activity Atlas.

Field QSAR, when successful, gives you the best of both worlds, i.e., predictions and interpretation.

Activity Atlas is qualitative only (no predictions) and is great for understanding the SAR for your data using activity cliffs analysis, especially when the SAR landscape is jagged.

Activity Atlas in V10.6 includes a new Activity Cliffs Summary algorithm which generates more detailed SAR maps reducing the reliance on individual compounds, especially useful for small and medium sized data sets.

In Figure 4, you can see the Field QSAR maps compared to the new Activity Cliffs Summary maps for the Orexin 2 data set.

Figure 4. Top: Field QSAR electrostatic (left) and steric (right) coefficients.  Bottom: Activity Cliffs Summary of Electrostatics (left) and Activity Cliffs Summary of Shape (right). Color coding: red = more positive electrostatic favors activity; blue = more negative electrostatic favors activity; green = favorable steric bulk; magenta = unfavorable steric clash.

Both types of maps clearly and consistently indicate where more positive (red) or negative (blue) electrostatics favors activity, and where steric bulk is favorable (green) or forbidden (magenta), providing invaluable indications for ligand design.

I don’t have ‘top quality’ data, but I still need a model

Sometimes the data you have are not as clean as you would like for the purposes of QSAR modeling. You may have % of inhibition data rather than pIC50s or pKis; data generated with different assays; or simply data which are qualitative in nature.

The new ML methods in Forge will work just as well to build classification models for sorting new molecules into existing categories (e.g., active/inactive). Forge will also provide appropriate visual tools (such as the confusion matrix, Figure 5) and classification performance metrics (Precision, Recall, Informedness) to assess the performance of the model and decide if it is good enough to be used in project work.

Figure 5. Confusion matrix for and useful statistics for an Orexin 2 classification model.

Improved graphics and GUI

In Forge V10.6 you will experience strong performance, great pictures and new smooth transitions between storyboard scenes thanks to new graphic engine which generates enhanced 3D objects (Figure 6).

Figure 6. The new graphic engine in Forge V10.6 generates great pictures.

This release includes also many other GUI and usability improvements, including:

  • An improved QSAR Model widget including relevant information and plots for the regression and classification models, PCA component plots, notes, and a ‘pop-up’ button to visually compare different models (Figure 7)
  • An improved interface for handling categorical data in support of classification models
  • Improved Blaze™ results window, showing an enrichment plot and statistics for each Blaze refinement level
  • New function to automatically assign selected molecules to roles, based on their Murcko scaffold
  • New function to run clustering from the main Forge GUI, specifying the desired similarity metric and threshold
  • New option to use all the available local CPUs, relaxing the 16-CPUs limitation of previous Forge releases
  • More responsive GUI for large projects, with improved performance on common operations such as application of filters, calculation and interaction with custom plots, exporting data
  • Faster, more robust and less memory-consuming calculation of Activity Miner™ and Activity Atlas large similarity matrices
  • Improved 2D display of molecules
  • Improved Activity Miner GUI
  • Improved plots now showing a regression line for selected molecules
  • Improved structural filters now including pre-defined filters for Ring, Aromatic Ring, Non-ring atom, Chiral atom, H-bond donor and H-bond acceptor
  • Improved Filters window, now including a green/red toggle to control whether each filter is enabled or disabled.

Figure 7. Compare different QSAR models with the new ‘pop-up’ button in the QSAR Model widget.

Stay tuned for more

Sign up for our newsletter to receive product release announcements, request your free evaluation or contact us to learn more about how Forge can help advance your project.

  1. US patent number 8,653,263B2

Sneak peek at Forge V10.6: Model building focus and much more


While the development team is busy giving the finishing touches to Forge V10.6, let’s have a quick look at what is new in this release.

Improved predictions through new models

Forge users told us that the development of QSAR models with strong predictive ability was still a pain point for their projects. Not surprisingly, this is what made us focus on model building in this release.

Forge V10.6 comes with a full panel of well-known and robust Machine Learning (ML) methods (Support Vector Machines, Relevance Vector Machines, Random Forests, kNN classification) which complement those available in previous versions (Field QSAR and kNN regression).

These ML methods can be used to build both regression and classification models, and this is reflected in a QSAR Model widget completely re-designed to provide relevant visualizations and statistics for both model types (Figure 1). While each regression and classification model can be built individually, there is an option in Forge to automatically run all the ML models and pick the best one for you.

Figure 1. Left: Observed vs. Predicted Activity graph for a SVM regression model. Right: Confusion matrix and statistics for a SVM classification model.

Generating qualitative models on small datasets

Activity Atlas is a qualitative method for summarizing the SAR for a series into visual 3D maps that can be used to inform new molecule design. Forge V10.6 includes a new Activity Cliff Summary method which generates more detailed SAR maps by slightly downsizing the importance of the strongest activity cliffs.

You may want to use the new flavor of the method for understanding the SAR of small to medium size data sets, as this will provide a finer level of detail. For larger data sets (e.g., for quickly understanding patent SAR information), the original algorithm will help you focus on the prevalent SAR signals.

More responsive GUI for larger projects

Working with large projects (more than 1,000 molecules with multiple alignments and QSAR models) will be much more efficient in Forge V10.6. You will see improvements in the performance of common operations such as application of filters, calculation and interaction with custom plots, exporting data. The calculation of the large similarity matrices in Activity Miner and Activity Atlas will also be faster, more robust and use less memory.

Furthermore, there is now an option to set-up Forge to use all the available local CPUs, if appropriate, as we have relaxed the 16-CPUs limitation in the previous release of the software.

Figure 2. Forge running on multiple local CPUs.

Improved interface to Blaze for virtual screening

The improved Blaze results window now shows an enrichment plot and statistics for each Blaze refinement level.

Figure 3. Improved interface to Blaze in Forge.

Stay tuned for more

Subscribe to our newsletter to receive the product release announcement, or contact us to learn more about Forge.

Pharmacophore constraints – when should they be used?

Cresset’s alignment and scoring technology is based around the idea of electrostatic matching: the ‘best’ alignment of a pair of conformations is generally the one that provides the most similar molecular interaction potentials. In most cases this is true. However, we recognize that often some portions of the electrostatic potential, for example those involved in a strong hydrogen bond, can be more important than others. This isn’t information intrinsic to the molecules themselves, but to the context in which they are being used. As a result, we have for many years offered the possibility of adding ‘field constraints’, which force the alignment algorithm to prioritize matching the electrostatic potential in particular places around a ligand.

It’s worth noting that the field constraint algorithm works by applying a score penalty to alignments which violate the constraint, not by forcing the constraint to be matched (Figure 1). In general, soft constraints delivered via score penalties are preferred to hard constraints where we simply disallow alignments that violate the constraint. There are two reasons behind this. Firstly, using a score penalty allows the strength of the constraint to be adjusted, so that it can be anything from a mild hint to a strict requirement. Second, in most cases we feel that it is better to get a best-effort alignment that violates the constraint than nothing at all, especially when using the algorithms in lead optimization and trying to align a set of known actives.

Figure 1: Field constraint requires the field at the position of a particular field point to have a certain minimum value.

While field constraints work extremely well in the majority of cases, there are a few exceptions. In some cases there isn’t a field point at the place you want to add a constraint (although this problem is alleviated by the new field editor in Forge and Spark). More commonly, the constraint that you want to add is either stronger or more specific than can be expressed through fields: sometimes generic positive potential isn’t enough and you want to explicitly require a hydrogen-bond donor, for example, and sometimes you want to ensure that the metal-chelating group in a set of ligands is aligned, and ‘good metal chelator’ is similarly difficult to pin down in terms of pure electrostatics. In addition, the donor-acceptor motif that is often required in the hinge binding region of kinases can be difficult to describe through fields alone, as the donor and acceptor electrostatics cancel out to some extent making this region contribute less to the overall alignment score than it should.

For these reasons we introduced the option of pharmacophoric constraints to our core algorithms. These operate in a similar fashion to field constraints, in that they are constraints not requirements. Traditional pharmacophore algorithms treat pharmacophores as binary switches: you either have a H-bond donor in a region of space, and hence a full match, or you don’t and therefore don’t have a match. Our pharmacophore constraints instead operate on a sliding scale. You specify how strong the constraint is, and hence how much alignment score should be lost if the constraint is not satisfied. We spent a while looking at what form the penalty function should take, and the best form overall was a simple linear penalty with a cap at 2Å distance (Figure 2).

Figure 2: Pharmacophore constraints require an atom of the specified type to be within 2Å of the constrained atom, or a penalty is applied to the score.

We’ve talked elsewhere about the validation of our pharmacophore constraints on virtual screening performance within Blaze, where for some targets they can make a significant difference. However, the new constraints are useful for more than just virtual screening. What effect does adding constraints have on the probability of a successful ligand alignment?

We investigated this by using the standard AZ/CCDC dataset from Giangreco et al. (J. Chem. Inf. Model. 2013, 53, 852−866), which contains a large number of targets, and for each target a collection of aligned co-crystallized ligands. The data set can be used to validate ligand alignment algorithms, by measuring what proportion of the ligands within a target class can be successfully aligned. However, running a validation over the entire data set would be challenging, as it would pose the question for each target as to what pharmacophore constraints should be applied. We decided to address this by only testing (in the first instance) 4 kinase targets, as the donor-acceptor-donor motif for kinases is well understood and fairly uncontroversial. For each ligand in these 4 target sets, we manually applied between 1 and 3 pharmacophore constraints to the hinge binding region based on how many H-bonds were actually made by that ligand. For the purposes of this experiment we only included ‘true’ H-bonds from hydrogens on O or N, not the weaker C-H form. Note that all of the alignments were performed in ligand space only, with no information from the proteins applied.

The distribution of results is shown in Figure 3. We used a cutoff of 2Å to distinguish between correct and incorrect alignments. As can be seen, a significant number (around 13%) of alignments that were incorrect when unconstrained were rescued by adding the constraints – that’s a significant improvement. However, at the same time, around 3% of the alignments that we had been getting correct were made worse by adding the constraints, and it’s also apparent that there are a number of incorrect alignments which are left unchanged by the imposition of constraints.  Overall the pharmacophores are a net win, especially for these kinase data sets, but it’s not a magic bullet.

Figure 3: Effect of adding pharmacophore constraints to 4 kinase alignment data sets. The alignments in the ‘Win’ section are improved by the addition of constraints, while those in the ‘Loss’ section are made incorrect.

Let’s look at a couple of examples. Figure 4 shows a case where adding hinge binder constraints converts an incorrect alignment to a correct one. The unconstrained result on the left matches the dimethoxyphenyl ring on the left and the phenol on the right very well, but at the expense of the quinazoline nitrogen which makes the hinge interaction. Adding a single acceptor constraint to that nitrogen switches the best solution to one where the indazole nitrogen satisfies the constraint, which happens to be the experimentally-observed alignment.

Figure 4: CDK2, aligning EZV to DTQ.

Similarly, in Figure 5, the best alignment in terms of raw shape and electrostatics of MTW and LZ8 is shown on the left. The 2-aminopyrimidine in MTW which makes the hinge interactions is partially matched by LZ8’s pyrazole and carbonyl, but the alignment in that part of the molecule isn’t that good. However, the alignment matches the phenyl ring in both molecules beautifully. Applying a donor/acceptor pharmacophore constraint pair forces the alignment algorithm to try and match the 2-aminopyrimidine much more closely, which leads to the correct alignment on the right. However, note that the phenyl rings are much more poorly aligned.

Figure 5: CDK2, aligning LZ8 to MTW.

Finally, Figure 6 shows a case where adding field constraints converts a correct alignment to an incorrect one. When aligning to DTQ as the reference, as in Figure 4, the field/shape alignment gets the correct alignment for C94. However, C94 does not make the hydrogen bond to the hinge that we have constrained in the reference. The constraint coerces the alignment to be flipped so that one of the sulfonamide oxygens matches the H-bond acceptor requirement, thus forcing the wrong answer to be obtained.

Figure 6: CDK2, aligning C94 to DTQ.

Pharmacophore constraints can, as you can see, be very useful. If you know that particular interactions are required by the protein, you can provide our tools with that information to help them find the correct alignment. However, it must be kept in mind that by adding these constraints you are adding bias to the experiment: you are nudging it to provide the answers that you expect. Sometimes, the unexpected answers turn out to be correct! We recommend, therefore, that pharmacophore constraints are used sparingly. I personally will always run an unconstrained alignment experiment as a control, just to see if Forge will find an alignment that I hadn’t expected but which might illuminate a previously-puzzling piece of SAR.

So, when should you use a field constraint and when a pharmacophore constraint? The answer depends both on your requirements and on the chemistry inside the active site. If there’s a pharmacophoric interaction that you are certain must be preserved, then using a pharmacophore constraint is probably justified. However, if the interaction is not always present in the same form, a field constraint may be more appropriate. Figure 7 shows a BTK example. The ligand on the left makes two hydrogen bonds to the backbone from the indazole. We might be tempted to add pharmacophore constraints to both the donor and the acceptor, except that there are known highly-active ligands which make a C-H hydrogen bond instead (see the ligand on the right). As a result, it is probably more appropriate, if constraints are required, to add a pharmacophore constraint to the acceptor nitrogen, but use a field constraint to indicate a donor preference in the NH/CH location.

Figure 7: BTK ligands – which constraint type should I use?

Try yourself

Pharmacophore constraints are now available in all Cresset ligand-based computational chemistry applications:

  • Forge: Powerful ligand-focused suite for understanding SAR and design
  • Torch: Molecular design and SAR for medicinal and synthetic chemists
  • Spark: Discover new directions for your project using bioisosteres
  • Blaze: Effective virtual screening optimized to return diverse new structures

Request an evaluation and try them yourself.

At the heart of medicinal chemistry design: The synergistic relationship between Forge and Spark


We present an exercise in in silico medicinal chemistry: using Activity Atlas1 to assess and understand the features that drive activity for a collection of CDK2 inhibitors; using Spark2 to suggest new ideas to explore; then visually assessing those new ideas in Forge3 in the context of the qualitative 3D-SAR models generated by Activity Atlas.


The efforts of medicinal chemists in their quest to design new, active ligands can be summarized by answering three deceptively simple questions:

  1. What to make next?
  2. Why should I make it?
  3. Has this chemical space been explored previously?

These questions are frequently addressed separately. However, there is much value in considering the questions simultaneously with the help of Forge and Spark.

In this case study, Cresset ligand-based design tools are used to address the questions above with respect to the well-studied target for oncology Cyclin-dependent kinase 2 (CDK2).


Compound collection and alignment

A collection of 38 imidazopyridine containing CDK2 inhibitors was gathered from CHEMBL4 and curated to create a molecular dataset spanning a pIC50 range of 4-9.

Prior to assessing any relationship between activity and 3D structure, it is necessary to align the collection of molecules. In this case we chose to align the molecules to the ligand derived from the crystal structure 1OIT (CHEMBL73303: pIC50 =9.0). As many of the compounds in the dataset contained a solvent-exposed group that was not present in this reference an additional active ligand was identified (CHEMBL70808: pIC50 = 8.52) that maintained the core, but expanded beyond the terminal sulfonamide of the 1OIT ligand. The CHEMBL70808 ligand was aligned to the 1OIT ligand using manual and automated methods to ensure good correspondence of the imidazo[1,2-a]pyridine groups that are deeply embedded in the binding site. Further manual alignment was employed to ensure comparable orientation of the sulfonamide, despite being solvent-exposed. The pair of ligands (Figure 1) was then used as a combined, equally weighted, reference for the subsequent alignment of the 38 compounds gathered.

Figure 1_ Alignment of PDB1OIT ligand
Figure 1. Alignment of PDB:1OIT ligand (pink) and CHEMBL70808 (grey).

Compound collection and alignment

A collection of 38 imidazopyridine containing CDK2 inhibitors was gathered from CHEMBL4 and curated to create a molecular dataset spanning a pIC50 range of 4-9.

Prior to assessing any relationship between activity and 3D structure, it is necessary to align the collection of molecules. In this case we chose to align the molecules to the ligand derived from the crystal structure 1OIT (CHEMBL73303: pIC50 =9.0). As many of the compounds in the dataset contained a solvent-exposed group that was not present in this reference from an additional active ligand was identified (CHEMBL70808: pIC50 = 8.52) that maintained the core, but expanded beyond the terminal sulfonamide of the 1OIT ligand. The CHEMBL70808 ligand was aligned to the 1OIT ligand and subjected to additional manual alignment in order to better align the imidazo[1,2-

a]pyridine groups that are deeply embedded in the binding site. Further manual alignment was employed to ensure comparable orientation of the sulfonamide, despite being solvent-exposed. The pair of ligands (Figure 1) was then used as a combined, equally weighted, reference for the subsequent alignment of the 38 compounds gathered.

Within Forge, the collection of CDK2 inhibitors was aligned to the two equally weighted references, using 50% fields and 50% shape and a soft protein excluded volume to generate an alignment (Figure 2) and similarity score (range 0.59 through 0.85). As alignment is at the heart of every 3D-QSAR approach, the alignments were visually inspected and manually tweaked to ensure consistent orientation of all side chains.

Figure 2_ Alignment of 38 CDK2 inhibitors to PDB1OIT and CHEMBL70808
Figure 2. Alignment of 38 CDK2 inhibitors to PDB:1OIT and CHEMBL70808.

Modeling 3D SAR with Activity Atlas

To generate a qualitative assessment of the features driving activity, an Activity Atlas model calculation was performed.

Activity Atlas is available in Forge, and generates visually striking maps based on the aligned molecules. The models depict: the average electrostatics and shape for active compounds; the summary of activity cliffs in electrostatics, shape and hydrophobicity; and the regions explored. The models generated by Activity Atlas are intended to help answer the questions asked earlier and are shown in Figures 3-5.

In Figure 3, the average active molecule map is depicted. It indicates the common electrostatics and hydrophobic features of active molecules in the dataset – essentially molecules that do not have these features are unlikely to have any significant activity. The central H-bond donor (amine)/H-bond acceptor (pyrimidine) motif mapping the interaction with the hinge region of CDK2 is present in all actives. Negative regions near the H-bond acceptors of sulfonamide and imidazopyridine are also present on all actives.

Figure 3_ Activity Atlas map for average electrostatics and hydrophobics for active molecules
Figure 3. Activity Atlas map for average electrostatics and hydrophobics for active molecules. Blue = negative electrostatics; Red = positive electrostatics; Gold = hydrophobics.

Figure 4 shows the summary of activity cliffs in electrostatics and steric space – the fine detail on the average of actives map. Areas that are blue suggest that making this area more negative (or less positive) could enhance activity; areas that are red suggest that making the area more positive (or less negative) can drive activity; and finally, steric bulk is favorable in the green regions, whereas it is not tolerated in the pink regions.

Figure 4 (right) displays the activity cliff analysis in an orientation better suited to examining the imidazopyridine. Tilted on its side, the Activity Atlas model suggests that a more positive/less negative π-cloud should enhance activity, as well as noting that a stronger negative near the H-bond accepting ring nitrogen should also favor activity.

Figure 4_ Two views of the Activity cliff summary for electrostatics and sterics
Figure 4. Two views of the Activity cliff summary for electrostatics and sterics. Left: Electrostatics and sterics. Right: Electrostatics only in a view rotated around the x axis. Areas in blue suggest that negative electrostatics are favored; areas in red favor positive electrostatics. Areas in green are favorable sterically, while areas in pink are unfavorable for sterically bulky groups.

The assessment of the electrostatic regions explored suggests that at least 10 of 38 molecules have contributions as shown in Figure 5. This means that regions that are not explored may present opportunities for modification and further optimization.

The region explored analysis also performs a calculation of novelty for all the data set compounds as well and novel designs. The calculated novelty can be used as an indication for how much information would be gained from the molecule should it be made.

Figure 5_ Aligned ligands with a display of the regions of electrostatic space explored
Figure 5: Aligned ligands with a display of the regions of electrostatic space explored.

Generating ideas in Spark

With a SAR analysis in hand, Spark was used to find bioisosteric replacements for the highlighted portion of

the 1OIT ligand as shown in Figure 6. The fragments were sourced from the Spark reagent database of boronic acids derived from eMolecules.

Figure 6_ The Spark eMolecules boronic acids reagent database
Figure 6. The Spark eMolecules boronic acids reagent database was designated to identify bioisosteric replacements for the moiety shown in pink in the PDB:1OIT ligand.

Results and analysis

Four results were selected from the Spark experiment were selected for further analysis and are shown in Figure 7. Each of the selected Spark hits have been chosen to either enhance activity, and/or to challenge the model since it was built on a relatively small number of compounds.

Each of the proposed bioisosteric replacements are derived from commercially available reagents that are shipped from the supplier within 2 days to 4 weeks.

In (a), the pyrazolopyridine replacement would test the importance of the H-bond acceptor strength on the original imidazopyridine. Moving the heteroatom across the ring results in a weaker H-bond acceptor while maintaining a reasonable density above the ring. Note that this replacement results in a richer central pyrimidine ring.

The dihdropyrroloimidazole (b) has similar electrostatics to that starting ligand but tests the requirement for an electron deficient aromatic by saturating this ring, removing all pi-electrons from this region.

In (c), the benzofuran might be suitable for testing the model as it removes the strong H-bond acceptor altogether, replacing it with a weak feature. Whilst this might not be productive in activity the molecule would add knowledge to the model and the reagent is available immediately.

Finally, (d) suggests replacing the imidazopyridine with a quinoline. The change from 5,6 to 6,6 ring introduces a different geometry for the H-bond acceptor but maintains many of the electrostatic features of the starting ligand in a reagent that is available for immediate dispatch. If this substitution was tolerated then the spark results contain other, more exotic suggestions that include a H-bond acceptor in a 6 membered ring in this position.

Figure 7. The starting ligand from 1OIT and selected Spark results (a-d). Top: in 2D. Center: in 3D showing positive and negative interaction potentials. Bottom: presented in the context of the Activity Atlas activity cliff summary for electrostatics.


This article presents our approach to answering the three questions at the heart of medicinal chemistry design:

1. What to make next?

Every chemist has a collection of tricks and substitutions that have been built up from their laboratory experience. Spark enhances the generation of ideas in an unbiased way, and in combination with a chemist’s intuition, can provide ideas about what compounds can be made that are in potentially open IP space, but still retain the characteristics of active molecules. The Spark experiment using eMolecules reagent databases also provides tier and ordering information for applicable reagents exploiting specific chemical reactions to improve laboratory efficiency.

2. Why should it be made?

Taking the Spark and chemist-designed ideas and examining them in the context of the Activity Atlas models is a way to ensure that new compounds meet the characteristics of active molecules, or conversely, are designed to test specific aspects of the models. The combination of average active molecule and activity cliff summary maps can provide the rationale needed to give confidence that taking a new design to the laboratory is worth the effort.

3. Has this chemical space been explored previously?

The region explored analysis combined with the novelty calculation within Activity Atlas can be used to assess whether the Spark and chemist-design ideas lie within the already explored chemical space, or whether they map regions of this space so far unexplored.

References and links




4 P. Bento, A. Gaulton, A. Hersey, L.J. Bellis, J. Chambers, M. Davies, F.A. Krüger, Y. Light, L. Mak, S. McGlinchey, M. Nowotka, G. Papadatos, R. Santos and J.P. Overington (2014) ‘The ChEMBL bioactivity database: an update.’ Nucleic Acids Res., 42 1083-1090.