IMP Tutorial  for IMP version 2.1.1
The IMP C++/Python library

The core of IMP is the C++/Python library, which provides all of the necessary components, as a set of classes and modules, to allow methods developers to build an integrative modeling protocol from scratch. Most users of IMP will use one of the higher-level interfaces described in later sections of this tutorial; however, we will briefly demonstrate this library here to illustrate the core IMP concepts that these interfaces rely on.

The IMP library is designed as a set of modules. The 'kernel' module is a small collection of classes that define the storage of information about the system and the main interfaces used to interact with that information. The information is stored in a set of Particle objects; these are flexible and abstract data containers, able to hold whatever information is necessary to represent the system. For example, a given Particle may be assigned x, y, and z attributes to store point coordinates, another may be assigned x, y, z, and a radius to represent a sphere, and another may contain two pointers to other Particles to represent a bond or another relationship. The kernel defines only the abstract interfaces to manipulate the data in the Particles, but does not provide implementations; these are provided in other modules. For example, it merely defines a Restraint as any object that, given a set of Particles, returns a score, and an Optimizer as an object that changes the attributes of all Particles to yield an optimized score over all restraints. It is the core module that provides, for example, a concrete Restraint acting like a harmonic 'spring' between two point-like Particles, an Optimizer that utilizes the conjugate gradients minimization method, and much other functionality.

Some other IMP modules provide the basic building blocks needed to construct a protocol, such as the core module that provides functionality including harmonic restraints, point-like and spherical particles, and basic optimizers, and the atom module that provides atom-like particles, a molecular dynamics optimizer, etc. Other modules provide support for specific types of experimental data or specialized optimizers, such as the em module that supports electron microscopy data, and the domino module that provides an inference-based divide-and-conquer optimizer. IMP is designed so that it is easy to add a new module; for example, a developer working on incorporating data from a new experimental technique may add a new IMP module that translates the data from this technique into spatial restraints.

IMP is primarily implemented in C++ for speed; however, each of the classes is wrapped so that it can also be used from Python. A protocol can thus be developed from scratch by simply writing a Python script. As an example, we will first look at the script simple.py:

1 ## \example core/simple.py
2 # Illustration of simple usage of the IMP library from Python.
3 #
4 
5 import IMP
6 import IMP.algebra
7 import IMP.core
8 
10 
11 # Create two "untyped" Particles
12 p1 = IMP.kernel.Particle(m)
13 p2 = IMP.kernel.Particle(m)
14 
15 # "Decorate" the Particles with x,y,z attributes (point-like particles)
16 d1 = IMP.core.XYZ.setup_particle(p1)
17 d2 = IMP.core.XYZ.setup_particle(p2)
18 
19 # Use some XYZ-specific functionality (set coordinates)
20 d1.set_coordinates(IMP.algebra.Vector3D(10.0, 10.0, 10.0))
21 d2.set_coordinates(IMP.algebra.Vector3D(-10.0, -10.0, -10.0))
22 print d1, d2
23 
24 # Harmonically restrain p1 to be zero distance from the origin
25 f = IMP.core.Harmonic(0.0, 1.0)
28 m.add_restraint(r1)
29 
30 # Harmonically restrain p1 and p2 to be distance 5.0 apart
31 f = IMP.core.Harmonic(5.0, 1.0)
33 r2 = IMP.core.PairRestraint(s, (p1, p2))
34 m.add_restraint(r2)
35 
36 # Optimize the x,y,z coordinates of both particles with conjugate gradients
37 d1.set_coordinates_are_optimized(True)
38 d2.set_coordinates_are_optimized(True)
40 o.optimize(50)
41 print d1, d2

In the first part of the script, the IMP kernel and the algebra and core modules are loaded, as regular Python modules. We then proceed to set up the representation of the system, using the Model and Particle classes defined in the kernel. The Model class represents the entire system, and keeps track of all the Particles, Restraints, and links between them. As mentioned earlier, the Particle class is a flexible container, but here we give the two Particles (p1 and p2) point-like attributes using the XYZ class defined in the core module. This XYZ class is known as a 'decorator'; it does not create a new Particle, but merely presents a new interface to an existing Particle, in this case a point-like one. (Multiple decorators can be applied to a single Particle; for example, an atom-like Particle could be treated like a point, a sphere, an electrically charged particle, or an atom.) We can then treat each Particle like a point using methods in the XYZ class, here setting the x, y, and z coordinates to a provided vector.

In the second part, we set up the scoring of the system. We add two restraints to the Model, one of which harmonically restrains p1 to the origin and the other of which restrains p1 and p2 to be distance 5.0 apart. (IMP does not enforce any units of distance; however, some physical optimizers, such as molecular dynamics, expect distances to be in angstroms.) Note that the core module provides suitable building block restraints for this purpose. In the first case, we use the SingletonRestraint class that creates a restraint on a single particle (p1). It delegates the task of actually scoring the particle, however, to another class called SingletonScore that is simply given the Particle and asked for its score. In this example, we use a type of SingletonScore called a DistanceToSingletonScore that calculates the Cartesian distance between the point-like Particle and a fixed point (in this case the origin), and again delegates the task of scoring the distance to another class, a UnaryFunction. In this case, the UnaryFunction is a simple harmonic function with a mean of zero. Thus, the Particle p1 is harmonically restrained to be at the origin. The second restraint is set up similarly; however, in this case the restraints and scores act on a pair of particles. This building block functionality makes it easy to add a new type of restraint; for example, to implement a van der Waals potential it is only necessary to provide a suitable PairScore that scores a single pair of particles; the functionality for efficiently enumerating all pairs of such particles is already provided in IMP.

Finally, in the third part of the script, we tell IMP that it can move the two point-like particles, and to build a system configuration that is consistent with all the restraints. In this example, a simple conjugate gradients optimization is used.

The script is a regular Python script. Thus, provided that both IMP and Python are installed, it can be run on any machine, by typing on a command line, in the same directory as the script:

python simple.py

The script will run the optimization, printing IMP log messages as it goes, and finally print the coordinates of the optimized particles.

IMP is designed such that the C++ and Python interfaces are similar to use. Thus, IMP applications or protocols can be constructed either in C++ or in Python, and new IMP functionality (for example, new types of Restraint) can be implemented in either language. For a comparison, please inspect the simple.cpp file below. This file implements the same protocol as the first part of simple.py but uses the IMP C++ classes rather than their Python equivalents. The two programs are very similar; the only differences are in the language syntax (eg, the Python 'import IMP.algebra' translates to '#include <IMP/algebra.h>' in C++) and in memory handling (Python handles memory automatically; in C++, memory handling must be done explicitly by using the IMP::base::Pointer class or the IMP_NEW macro, which adds reference counting to automatically clean up after IMP objects when they are no longer in use).

/** \example core/simple.cpp
Simple example of using the IMP C++ library.
This should be equivalent to the first part of the Python example simple.py.
*/
#include <fstream>
#include <IMP/kernel.h>
#include <IMP/algebra.h>
#include <IMP/core.h>
int main()
{
// Create two "untyped" kernel::Particles
// "Decorate" the kernel::Particles with x,y,z attributes (point-like particles)
IMP::core::XYZ d1 = IMP::core::XYZ::setup_particle(p1);
IMP::core::XYZ d2 = IMP::core::XYZ::setup_particle(p2);
// Use some XYZ-specific functionality (set coordinates)
d1.set_coordinates(IMP::algebra::Vector3D(10.0, 10.0, 10.0));
d2.set_coordinates(IMP::algebra::Vector3D(-10.0, -10.0, -10.0));
std::cout << d1 << " " << d2 << std::endl;
return 0;
}