Rosetta 3.1 Release Manual

How things work in Rosetta

This document was orginally written 11 Nov 2007 by Chu Wang.

Metadata

This document is written for the purpose of helping developers who are used to Rosetta programming environment to grasp some key concepts when they first migrate over to use Rosetta The knowledge, based on which this document is written, is mainly from my personal experience of learning and using Rosetta and it is very likely that some parts of this document may not be accurate due to my limited understanding. Moreover, all credits should go to Phil Bradley and Andrew Leaver-Fay, who have laid a solid foundation for Rosetta, as well as all other Rosetta developers who have done tremendous amount of work for Rosetta development. Please feel free to add/modify/correct/update any part of its content as necessary.

Rosetta philosophy

A simple abstract on how Rosetta works is that the molecular system to be modeled has a certain number of degrees of freedom (DOFs), such as phi/psi/chi in protein folding, rigid-body translation/rotation in docking and amino acid sequence in design. Together with other fixed parameters by default assumption, these DOFs are translated into actual xyz positions of atoms in Cartesian space, via processes such as "refold" and "pack_rotamers". Afterwards, a physically-based energy function is used to evaluate the energy state of the current system based on these atom positions. The "DOF-XYZ-Energy" regime is coupled with algorithms such as Monte Carlo Minimization to sample the conformational space defined by those DOFs in order to locate the conformation with the lowest-energy state, which presumably should be close to the native state of the system.

Pose

Most of scientific principles and philosophies in Rosetta and Rosetta are identical, but in Rosetta, they are realized and coded in a relatively different way, taking full advantage of C++ object-oriented programming techniques. Among these objects, "Pose" is probably the essential one and it stays at the most upper level. "Pose" in Rosetta is similar to that in Rosetta in a way that it represents a certain state of molecular system to be modeled or simulated, such as protein, nucleic acid, small chemical ligand, and complexes with any combination of them. It stores information of chemical compositions and DOFs of the system. It can receive instructions on what changes are to be made on DOFs and know how to update atom XYZ positions given such changes. It keeps information of energy cached data and knows how to evaluate the energy of itself efficiently, based on the updated atom positions and energy evaluation methods to be used. In a standard modeling protocol, a Pose is initialized either from an input structure or from scratch, and then goes through a sequence/combination of operations which will change its DOFs to sample the conformational space and evaluate its energy so that the lowest-energy state can be located. There are two essential components in a Pose object, "Conformation" and "Energies", the former of which is in charge of translating changes on "DOF"s to XYZ atom positions and the latter of which is responsible for updating individual "Energy" components given the updated atom coordinates.

Conformation

As mentioned above, "Conformation" layer maintains a molecular representation of the system to be modeled and is responsible for translating any changes on DOFs of the system to actual changes of atom Cartesian coordinates. This object has three categories of data:

Residue

Each Pose/Conformation is broken down into small units called "Residues", which could be amino acids, nucleic acids or any group atoms with certain rules of what they are and how they are connected, such as a small chemical ligand moiety. The chemical content of a Residue is stored in an object called "ResidueType" and aside from that each Residue has other data storing actual coordinate information of each atom it contains as well as coordinate-related data such as mainchain/sidechain torsion angles, sequence position etc. For example, in a protein there might be multiple Leucine residues, each of which will be an individual "Residue" object. Each Leu Residue has its own coordinate data, but all Leu will have the same Leu ResidueType which contains information on what are the atoms, their names, chemical elements and connectivity. This setup also allows a sidechain Rotamer to be represented just as a Residue.

Chemical Layer

ResidueType

If you are familiar with amino acid data defined in read_aaproperties.cc in Rosetta, it will not be too hard to understand what a ResidueType object does. Basically, it has the same type of information, which is all related to the chemical representation of a residue, for example, things like how many atoms it has, what is the index of first hydrogen atom and which atoms could be potential hydrogen bonding donors and acceptors, etc. Each individual ResidueType is created by inputting data from a "param" file. (see rosetta_database/chemical/residue_type_sets/fa_standard/residue_types/ALA.param for an example), which contains information on how to define properties of this ResidueType. The advantage of using ResidueType param file is obvious:

Besides chemical data, a ResidueType also contains information how to generate an idealized copy of a Residue, similar to Rosetta "Icoord" data in read_aaproperties.cc. However, to be consistent with atom-tree representation in kinematic layer, A ResidueType Icoords are encoded in "internal coordinates" instead of "Cartesian coordinates". That is, for each atom in this ResidueType, a set of parameters are defined to instruct how to generate this atom from a set of (three) reference/stub atoms. These parameters, bond length, bond angle and torsion angle, are essentially DOFs in atom-tree. In addition to the internal coordinate data, a ResidueType object also contains information about how they can be connected to other ResidueTypes. For example, each amino acid ResidueType has two polymer connections, "backbone N" connecting to its lower end and "backbone C" connecting to its upper end. A cysteine ResidueType could have another non-polymer connection which allows its sulfur to be connected with a sulfur atom from another cysteine. Multi-residue ligands are also connected with those non-polymer connections.

AtomType

A ResidueType has a group of atoms, more precisely, a group of atom names and their atom_type_index. By these atom_type_indexes, the actual chemical properties of individual atoms -- its AtomType, such as chemical element, charges, LJ radius and LK solvation volume can be looked up from AtomTypeSet. In rosetta_database/chemical/atom_type_sets/..., there are input files containing data of atom properties and a new AtomType can be easily created by supplying necessary data in that file. An AtomTypeSet object is linked with a ResidueType so that the ResidueType knows what atoms are supported and can look up their properties as necessary.

Kinematic Layer

We have covered most of the "chemical" layer (from Pose to Conformation to Residue to ResidueType to AtomType) and letĒs move on to the "kinematic" layer to see how the degrees of freedom in the system are actually translated into actual atomsĒ positions for scoring. The essential pieces of data of kinematic layer are FoldTree and AtomTree, which are solely the brainchildren of Phil Bradley.

FoldTree

FoldTree is a residue-based description of macromolecular system and has been extensively implemented in Rosettta++ recent development. For more information, one can refer to two reference papers: Bradley et al, Proteins, 2006; Wang et al, JMB, 2007 or see FoldTree Overview and Concepts. Basically, each residue is connected another residue in the tree either via a peptide connection or a rigid-body connection ("Jump"). DOFs of the system are backbone and sidechain torsion angles and rigid-body transformations. Changes to these DOFs can be propagated properly in the FoldTree to generate desired structural changes.

AtomTree

AtomTree provides an even more general framework to model biomacormolecular system. As its name stands for, it represents all atoms in the system in a tree-like topology as compared to the residue-based representation in FoldTree. Atoms are connected to each other via bond distance, bond angles and torision/improper angles, and rigid-body transformation, all of which are potential DOFs of the system. (FoldTree can be thought as a special AtomTree in which bond length/angles are fixed and only torsional/rigid-body DOFs are allowed). Not only does the AtomTree allow users to treat more DOFs, its atom-based representation also has great advantage for handling non-polymer-type molecules with arbitrary atom connections, such as small chemical ligands. For more information, one can refer to a very good overview description on some important concepts to understand AtomTree written by Ian Davis (AtomTree oveerview and concepts). Here are a few basic key points:

Rosetta internally uses only AtomTree to handle kinematics. FoldTree is still supported for defining kinematic representation of the modeled system as it is convenient to construct for protein-based modeling tasks. When a FoldTree is passed to Pose/Conformation, it will be converted into the corresponding AtomTree representation. For convenience purpose, the interface for setting most common types of DOFs are supported, for example, when pose.set_psi(18, 70.0) is called, it will look up an internal mapping function to find the corresponding AtomTree internal coordinate of a specific atom and set it accordingly. Whenever there is a change to any AtomTree internal coordinates, a mark is turned on indicating that atom XYZ coordinates are outdated. So next time any operation which will rely on XYZ coordinates, such as scoring or accessing the position of a certain atom, is requested, a "refold" -- update_xyz_coords() will be triggered to ensure the XYZ coordinates data is updated properly. Similarly, AtomTree also provides an interface for users to change XYZ coordinates for any atoms directly and in this situation, a similar tag will be set indicating internal coordinates need to be updated and the function of "update_internal_coords()" will be triggered if any internal coordinates data are to be used.

If you still remember, a Residue Object also has data for all its atom positions and its backbone/sidechain torsions. The source of these data will be updated from AtomTree kinematic layer whenever necessary. The AtomTree internally maintains a map of Atom Pointers so that given a residue number and atom number, a specific Atom can be looked up. Then its position can be accessed and any backbone/sidechain torsion can be mapped to a specific internal coordinate of a certain atom.

There are numerous member functions to manipulate AtomTree, such as, deleting/attaching/replacing a sub-tree by bond or by rigid-body, or re-rooting an AtomTree. An example would be during the construction an AtomTree for a protein, a Residue sub-tree is created first and then attached to the main AtomTree. Another one would be to change a Residue from centroid representation to full atom representation. Since they are treated as different residues in Rosetta now, a sub-tree for the fullatom residue will be created and replace the sub-tree of the centroid residue in the main AtomTree.

The last thing listed here about AtomTree is that change to any DOFs will be recorded in the cached data in Conformation, i.e., this torsion has moved or this rigid-body has moved so that it will be passed along to scoring routines for efficient and accurate energy evaluation.

Scoring

I have not plowed through the scoring/energies part yet and this probably needs more input from Andrew and Phil.

Energies

The Energies object contains cached energy components for a Pose/Conformation. It contains classic one-body and two-body energiy components, but the implementation of their storage are different. Residue pairwise (two-body) energies are stored in a EnergyGraph in which each node is a residue and an edge between the two nodes is the interaction energy between those two residues. This type of Graph has also been applied in many scoring-related implementations, such as residue-pair-wise distances for setting up neighbor lists. When structure has moved and a scoring is requested, the Energies object will check if any old cached data is re-usable and only update those which are necessary to be recalculated.

ScoreFunction

A ScoreFunction has a set of weights, each of which is for an individual energy component. Each energy component will be evaluated by a certain type of EnergyMethod, for example, LJ-attractive will be evaluated by a Context-independent-2-body method (ci_2B_method) and Dunbrack will be evaluated by a Context-independent-1-body method. The list of all supported EnergyMethods is also maintained by A ScoreFucntion object. When a ScoreFunction is applied to a Pose to evaluate its energy, all non-zero-weighted energy components will be calculated using their corresponding scoring methods. One can find more information about currently supported energy methods and their categories in src/core/scoring/methods/.

ScoreFunctionFactory

This object mainly facilitates initializing ScoreFunction weights from an external file and additionally applying a patch to add or remove or change subset of the weights as necessary. A list of defined scoring function weights can be found in rosetta_database/scoring/weights.

Protocol related

Once a Pose is initialized, either from an existing structure or from scratch, it can be passed to different protocol-level functions to perform certain types of modeling task. This includes but is not limited to MonteCarlo, gradient-based minimization, packer, rotamer_trials and a more general object wrapper called "Mover". These upper-level protocols are under intensive development and the hope is that based on the machineries provided by Pose and all its internal building modules, users will be granted freedom to construct their own protocols, tailed specifically for their research goals. Here are very short descriptions about these upper-level building blocks:

MonteCarlo

This object should look very similar as it is one of the core principles used by Rosetta. It internally stores two Poses, which records the last accepted state and the lowest energy state so far visited, a ScoreFunction for scoring purpose and a temperature factor to adjust Boltzmann probability for acceptance/rejection.

AtomTreeMinimizer

Since the kinematic setup for a Pose is solely via AtomTree now, the optimization function is carried out by this AtomTreeMinimizer object. This object has a member function called "run" which takes a Pose, a MoveMap, a ScoreFunction and a MinimizerOption wrapper object to perform the task of energy minimization. The object energy function to be minimized is obviously from ScoreFunction and the variables are those DOFs in the Pose/Conformation/AtomTree. Which subset of DOFs are to be minimized is specified by this MoveMap object and what type of minimization to be carried out is defined in the MinimizerOption object (it also has minimization tolerance value and other minimization-related parameters). For more information on minimization, see Minimization overview and concepts.

MoveMap

The purpose of MoveMap is similar to those "set_allow_move" functions in Rosetta, that is, to enable/disable certain types of DOFs in the system. For example, omega is fixed but phi/psi/chi are flexible. However, unlike that those "allow_**_move" data are internal to Pose in Rosetta, MoveMap is isolated from Pose in Rosetta and are passed separately to the AtomTreeMinimzer. A MoveMap supports enabling/disabling a specific type of DOFs, such as bond length and angles; or certain DOFs of a single residue, such as backbone torsion of residue i; or individual DOF in the AtomTree, such as torsion angle associated with atom j in residue i. This provides more freedom for users to control what DOFs to be optimized.

Packer and Rotamer_trials

Both objects perform rotamer-based sidechain/sequence optimization given a Pose. The information about what to be repacked or redesigned is wrapped into the PackerTask object, which can be supported by many command line options that are quite similar to those used in Rosetta.

Mover

The design of this object has recently brainstormed by many Rosetta developers, with the aim of creating a prototype class for representing any type of operations on a Pose which will change its structure. The setup of this class allows it to be easily extended to create new composite types of moves from single-type moves. Each Mover will take in a Pose and has a "apply" method which will do some certain changes to the Pose. Examples of single-type moves are SmallMover/ShearMove will do perform classical Rosetta small/shear backbone moves, RigidBodyMover that will change a rigid-body (jump) DOF. MonteCarlo, Minimization and Packer can also be considered as special type of Movers. So combining them can create Movers that carry out tasks like SmallMinTrial or RigidBodyPackerTrial etc. A single Mover or a series of Movers can also be applied repeatedly in a number of cycles. Moreover, a Protocol itself can be deemed as a composite Mover which is essentially constructed by combining any single-type Mover objects.
Generated on Tue Apr 20 07:50:05 2010 for Rosetta Projects by  doxygen 1.5.2

© Copyright Rosetta Commons Member Institutions. For more information, see http://www.rosettacommons.org.