The PolyPharma app provides a variety of modelling and visualisation capabilities that can be applied to molecular structures. There are hundreds of available models, each of which addresses a disease target or a biological function that should be either activated or avoided, depending on the objective. The app is designed to make it easy to search for structure-activity relationships for the targets of interest, but also easy to spot serendipitous effects, such as alternate disease targets that were not originally anticipated, or possible problems with ADME/tox.
The models are prebuilt and prepackaged within the app, which means that they are immediately accessible from the moment the app is installed. Also, the cheminformatics algorithms for applying models and viewing the results are implemented natively on the device, which means that all of the functionality is available without the need to use the internet, or be concerned with the security implications of doing so.
When opened for the first time on an iPhone or iPad, the PolyPharma app presents a front page with 3 sections, entitled Profiles, Predictions and Molecules:
A profile is a selection of targets and off-targets of particular interest, which is typically associated with a disease campaign. The molecules are a customisable list of molecular structures for potential drugs. In order to find out what the app is capable of, it is necessary to select a single profile, and a single molecule, simply by tapping on them. At that point, the app begins to undertake a series of calculations, and provides a cornucopia of predictions to visualise.
To initiate the prediction process, it is necessary to select a profile. Tapping on any of the disease target buttons listed at the top of the front page will cause that profile to become the current one:
Besides highlighting the button, the app begins by going through all of the custom molecules, and adding a colour summary along the bottom of their button glyphs:
Selecting the first molecule by tapping on it gives the PolyPharma app enough to focus on and generate a slew of predictions. The front page view quickly becomes quite tall, and can be scrolled by swiping up and down. At the top of the newly filled in predictions section is a "flower petal" outline, which displays a prediction for the models corresponding to each of the targets selected within the profile:
For primary and secondary targets (in this case Tuberculosis and solubility respectively), the predictions are shown with a sliding colour scale using the "heatmap" signature (red = low, yellow = intermediate, green = high). Each of these represents a Bayesian model that was created using ECFP6 fingerprints. The angular wedge for the primary target is twice as large as for the secondary target. There are also three off-targets which indicate bioactive targets that should be avoided, being in this case cytochrome P450, hERG and p-glycoprotein. These are coloured slightly differently, using the "doppler" signature (blue = low, white = intermediate, red = high), since instances where a high activity is predicted are indicative of potential problem molecules.
Underneath this is are a series of frames that show the molecular structure repeated 5 times, once for each target indicated within the profile. These represent a detail view of each of the predictions, wherein the atoms are individually coloured in a way that represents the atomic contribution to the activity of the overall model. The average atom is coloured according to the overall prediction, whereas those that are unusually more or less represented within active molecules from the training set have more or less intense colour schemes. The contrast (offset) can be adjusted using the slider control directly above the structures. This molecule-specific colour coding provides potential clues as to which particular parts of the molecule may be responsible for notable properties.
For each of the targets within the profile, the app stores a selection of representative molecules from the training set that was used to build the Bayesian model. These are shown as a series of thumbnail summaries using the honeycomb-style clustering layout, with the current molecule at the centre:
In the front page view, these honeycomb patterns are simple previews. They can be viewed in detail using the zoom button (at the top right), and will be described in the next section.
The currently selected profile typically identifies just a handful of the desired and undesired biological targets for the disease or subject of interest. The app is provided with a large collection of models, any one of which might just happen to be fortuitously (or otherwise) relevant to the selected molecule. Once the main predictions have been completed, the app continues on to make predictions with all the other available models. In the following section, the top 20 targets and off-targets respectively are listed:
The most high scoring models are presented, on the off chance that they are relevant. Disease targets that score well represent the possibility that the proposed molecule has polypharmacological properties, or has activity against an unexpected target. ADME/tox targets that score highly represent potential problem cases, as a red flag for possible unwanted side effects.
Each of the honeycomb thumbnails has an arrow button at the top right edge, the pressing of which brings up the detail view:
The hexagonal layout is created by placing the selected molecule in the middle, and iteratively placing the selected compounds from the underlying model around it. The compounds from the model are placed in order or decreasing similarity to the current molecule, and their position on the hexagonal grid is chosen in order to maximise similarity to neighbouring structures:
The view can be pinched & zoomed interactively. In the above case, the hexagons are coloured about their rims with red for inactives and green for actives, since the underlying Tuberculosis dataset was classified in a binary fashion (active or inactive). The current molecule is shown in the middle, distinctively white-on-black. Note that some of the structures are drawn on a grey background: these are the other custom molecules, which are not part of the model, but are included in the clustering layout to provide perspective.
All targets that should be optimised in favour of are coloured using the heatmap scheme, and those which have source data as continuous properties are displayed as such, for example with solubility:
Off-target properties are coloured using the Doppler scheme, where blue indicates the safe zone (according to the prediction), while red is potentially troublesome:
Besides the handful of profiles that are provided with the app by default, it is possible to edit the existing cases (e.g. to add new targets) or to create new profiles, such as for new diseases. It is straightforward to add or remote biological targets using the editing dialog:
Existing targets and off-targets can be reordered or deleted. Underneath the current selection is a long list of additional targets, each of which provides some information about the constituent data that went into the Bayesian model. Each target is indicated by two buttons: Target and Avoid. Helpfully the default state is indicated by highlighted in turquoise (i.e. for targets that are commonly considered to be of the ADME/tox variety, the Avoid option is default). Tapping on any of these buttons adds the target to the corresponding list.
The list of custom molecules is shown at the bottom of the view. By default, the app includes a selection of molecules that happen to be interesting relative to the default profiles, but they are not necessarily interesting to your drug discovery project, so may be removed at leisure. The touch-and-hold gesture on any of the molecules brings up an action selection:
Along the divider bar that demarcks the molecules section are three buttons, which bring up menu selections relating to bringing in molecules or applying to the whole collection.
The import button provides ways to defer to other sources for bringing in new molecules:
The Lookup by ID action allows database identifiers to be entered in order to fetch a compound from an online source. Download File invokes the iCloud integration features which allows files to be downloaded from compatible services, which also includes Dropbox.
Adding a new molecule manually currently presents just one option:
Molecules can be Pasted from the clipboard, which is a classic method of interprocess communication.
Several options for exporting the entire list of structures are available:
Sending the compounds by Email includes both the native XML datasheet and MDL SDfile formats. Launching in either of these formats provides a chance for other apps to receive the content, while Uploading invokes the iCloud API, which also applies to other compatible services, such as Dropbox.
If a profile has been selected, then each of the molecules will be displaying a thumbnail indication of the predicted activities as a coloured mnemonic. The actual numeric prediction values will be included in the files that are exported.
The PolyPharma provides as selection of powerful prediction and visualisation methods, based on precurated data and models that are included within the app. The modelling techniques are instantly accessible by tapping on several default values. Making use of this functionality with custom molecules and selected target profiles requires minimal effort.