Rosetta 3.3
Documentation for ab initio protein structure modeling application (AbinitioRelax)
David E Kim


This document was last updated on November, 2010 by David E Kim <>. The PI is David Baker <>. The AbinitioRelax application was developed by numerous Rosetta Commons members, primarily:

Code and Demo

The ab initio executable is in rosetta_source/src/apps/public/ See the test/integration/tests/abinitio directory for an example ab initio run and input files.



This application was developed to predict the 3-dimensional structure of a protein from its amino acid sequence.


The AbinitioRelax application consists of two main steps. The first step is a coarse-grained fragment-based search through conformational space using a knowledge-based "centroid" score function that favors protein-like features (Abinitio). The second optional step is all-atom refinement using the Rosetta full-atom forcefield (Relax). The "Relax" step is considerably more compute-intensive and time-consuming than the first step. A single AbinitioRelax run can generate a user defined number of models via a command line option (see Options section below). For increased conformational sampling, this application is easily parallelized by executing numerous jobs each using a unique random number seed (see Options section below). This is typically done by submitting multiple jobs to a computer cluster or distributed grid.

Previously, the standard structure prediction protocol was to (1) generate a large sample of "low-resolution" models using the first step (typically up to 10,000), (2) cluster the low-energy models using a score cutoff of around 10-20 percent, and then (3) select cluster centers for all-atom refinement using the Relax application. The advantage of this protocol is that it is relatively time efficient since "Abinitio" folding is faster and "Relax" is more time-consuming. However, the potential drawback is that if no near-native models are sampled after the "Abinitio" folding step, it is impossible to correct them during the "Relax" stage. With more and more computational power available, the "abrelax" protocol (see Options section below) was created to streamline this process by doing "Abinitio" folding followed directly by "Relax". Obviously, this protocol is much more time-demanding and improvements are only realized with enough conformational sampling (partially due to the fact that the full-atom energy function is very sensitive to imperfect atomic interactions and more noise will exist with insufficient sampling); convergence towards the native structure may require a significant amount of sampling. Additionally, to increase your chance of sampling the correct topology, a diverse set of homologous sequences, preferably with sequence changes that may have a greater impact on sampling like deletions and differences in conserved positions, may also be run since a homologue may converge towards the native structure with significantly less sampling (see Bradley et al reference).

Input Files



You can run the AbinitioRelax application with the following flags (to list all relevant commands, run with -help option):

-in:file:native 1l2y.pdb                        Native structure (optional)
(or -in:file:fasta 1l2y_.fasta)                 Protein sequence in fasta format (required if native structure is not provided)
-in:file:frag3 aa1l2yA03_05.200_v1_3            3-residue fragments (fragments file)
-in:file:frag9 aa1l2yA09_05.200_v1_3            9-residue fragments (fragments file)
-database ../minirosetta_database               Path to rosetta database
-abinitio:relax                                 Do a relax after abinitio ("abrelax" protocol), default=false.

-nstruct 1                                      Number of output structures
-out:file:silent 1l2y_silent.out                Use silent file output, use filename after this flag, default=default.out
(or -out:pdb)                                   Use PDB file output, default=false
-out:path /my/path                              Path where PDB output files will be written to, default '.'

There are several optional settings which have been benchmarked and tested thoroughly for optimal performance (we recommend using these options):

-use_filters true                               Use radius of gyration (RG), contact-order, and sheet filters. This option conserves computing
                                                by not continuing with refinement if a filter fails. A caveat is that for some sequences, a large
                                                percentage of models may fail a filter. The filters are meant to identify models with non-protein
                                                like features.
-psipred_ss2 1l2y_.psipred_ss2                  psipred_ss2 secondary structure definition file (required for -use_filters)
-abinitio::increase_cycles 10                   Increase the number of cycles at each stage in ab initio by this factor.
-abinitio::rg_reweight 0.5                      Reweight contribution of radius of gyration to total score by this scale factor.
-abinitio::rsd_wt_helix 0.5                     Reweight env,pair,cb for helix residues by this factor.
-abinitio::rsd_wt_loop 0.5                      Reweight env,pair,cb for loop residues by this factor.
-relax::fast                                    Do a fastrelax which is significantly faster than the traditional relax protocol without a significant
                                                performance hit.
-kill_hairpins 1l2y_.psipred_ss2                Setup hairpin killing in score (kill hairpin file or psipred file). This option is useful for all-beta
                                                or alpha-beta proteins with predicted strands adjacent in sequence since hairpins are often sampled too

For running multiple jobs on a cluster the following options are useful:

-constant_seed                                  Use a constant seed (1111111 unless specified with -jran)
-jran 1234567                                   Specify seed. Should be unique among jobs (requires -constant_seed)

-seed_offset 10                                 This value will be added to the random number seed. Useful when using time as seed and submitting many
                                                jobs to a cluster.  If jobs are started in the same second they will still have different initial seeds
                                                when using a unique offset. If using Condor (, the Condor process id,
                                                $(Process), can be used for this. For example "-seed_offset $(Process)" can be used in the condor submit file.

The standard command line for optimal performance is shown below (nstruct should be set depending on how many models you want to generate):

./bin/AbinitioRelax.linuxgccrelease \
        -database ../rosetta_database \
        -in:file:fasta 1l2y_.fasta \
        -in:file:native 1l2y.pdb \
        -in:file:frag3 aa1l2yA03_05.200_v1_3 \
        -in:file:frag9 aa1l2yA09_05.200_v1_3 \
        -abinitio:relax \
        -relax:fast \
        -abinitio::increase_cycles 10 \
        -abinitio::rg_reweight 0.5 \
        -abinitio::rsd_wt_helix 0.5 \
        -abinitio::rsd_wt_loop 0.5 \
        -use_filters true \
        -psipred_ss2 1l2y_.psipred_ss2 \
        -kill_hairpins 1l2y_.psipred_ss2 \
        -out:file:silent 1l2y_silent.out \
        -nstruct 10

Extracting PDB models from a silent output file using the score application

The resulting output using the command above is a silent output file (1l2y_silent.out) which contains the PDB models and Rosetta score information in a compact format. To extract the PDB models into individual PDB files from the silent file you can use the score.linuxgccrelease score application using the following command:

./bin/score.linuxgccrelease \
        -database ../rosetta_database \
        -in:file:silent 1l2y_silent.out \
        -in:file:fullatom \
        -output \

Clustering using the cluster application

Models from a single silent output file can be clustered using the cluster.linuxgccrelease cluster application using the following command:

./bin/cluster.linuxgccrelease \
        -database ../rosetta_database \
        -in:file:silent 1l2y_silent.out \
        -in:file:fullatom \
        -cluster:radius -1

PDB files of the cluster members are extracted from the silent output file by the cluster application.

Additional cluster options include (see cluster.linuxgccrelease for more information):

   -cluster:radius  <float>                    Cluster radius in A (for RMS clustering) or in inverse GDT_TS for GDT clustering. Use "-1" to trigger
                                               automatic radius detection
   -cluster:gdtmm                              Cluster by gdtmm instead of rms
   -cluster:input_score_filter  <float>        Ignore structures above certain energy
   -cluster:exclude_res <int> [<int> <int> ..] Exclude residue numbers from structural comparisons
   -cluster:radius  <float>                    Cluster radius
   -cluster:limit_cluster_size      <int>      Maximal cluster size
   -cluster:limit_clusters          <int>      Maximal number of clusters
   -cluster:limit_total_structures  <int>      Maximal number of structures in total
   -cluster:sort_groups_by_energy              Sort clusters by energy.


The AbinitioRelax application performs best for small monomeric proteins that are less than 100 residues in length. It is possible to get accurate predictions for some proteins up to around 150 residues, however, larger proteins require significantly more computing as the conformational space is vastly increased. It is not uncommon to sample in the range of 20,000 to 200,000 models in order to converge towards the native structure. The following references provide information relevant to the sampling problem:


As stated above, it is beneficial to try to identify homologous sequences to run along with the target sequence (see Bradley et al reference). Homologs can be identified using search tools like PSI-BLAST to search the non-redundant sequence database (NCBI nr database) or Pfam. Using a sequence alignment viewer like Jalview is very useful to help select an optimal set of homologs to run and also to aid in model selection. Typically we look for a diverse set of homologs (up to 10) with differences in conserved positions and deletions which may represent a truncated loop or disordered region. Small changes in sequence can have a large impact on the topologies that are sampled, for example, a polar residue at a conserved hydrophobic position can have a big effect. It is also important to identify and trim disordered termini using publicly available programs like Disopred or metaPrDOS. Signal sequences should also be identified and trimmed using publicly available programs like SignalP. This protocol is not developed for membrane proteins. If transmembrane helices are predicted using programs like TMHMM, please refer to our Membrane ab initio application.

Expected Outputs

Generates pdb files and an energy file, or a silent output file.

Post Processing

We recommend generating up to 20,000 to 30,000 models of the target sequence and of up to 10 homologs and then using the Cluster application to identify the most frequently sampled conformations. In a general case, at least one of the top 5-10 clusters by size may have models with the lowest rmsd to the native structure.

In an ideal case, your sequence will have many homologs identified by search tools like PSI-BLAST. Sequence alignments can be extremely helpful in model selection. For example, conserved hydrophobic positions most likely represent the core of the protein so models that have sidechains exposed in such positions may be discarded. The same logic applies to conserved polar positions which are most likely on the surface. Additionally, conserved cysteine pairs may represent disulphides. Tools like Jalview to view alignments and PyMOL to view models are extremely helpful for model selection in this respect.

Score versus RMSD plots may be helpful for identifying convergence towards the native structure for the target sequence and homologs. For example, the lowest scoring model can be used for the in:file:native input option when rescoring models with the score.linuxgccrelease score application. A score versus RMSD plot from the resulting score file may show convergence (an energy funnel) towards the lowest scoring model. If an energy funnel exists, the lowest scoring model has a greater chance of being near-native.

Lowest scoring models that are in a cluster and that have a topology represented in the PDB also have a greater chance of being correct. Structure-structure comparison tools like Dali or Mammoth can be used to search against the PDB database.

 All Classes Namespaces Files Functions Variables Typedefs Enumerations Enumerator Friends Defines