Fragment‐HMM: A new approach to protein structure prediction

Abstract

We designed a simple position-specific hidden Markov model to predict protein structure. Our new framework naturally repeats itself to converge to a final target, conglomerating fragment assembly, clustering, target selection, refinement, and consensus, all in one process. Our initial implementation of this theory converges to within 6 Å of the native structures for 100% of decoys on all six standard benchmark proteins used in ROSETTA (discussed by Simons and colleagues in a recent paper), which achieved only 14%-94% for the same data. The qualities of the best decoys and the final decoys our theory converges to are also notably better.Keywords: protein structure prediction; hidden Markov model; fragment assemblyWe wished to find a unified and the simplest model for protein structure prediction, one of the major open problems in science. We were not interested in trying PSI-BLAST for easy targets, threading by RAPTOR (Xu and Li 2003) for harder targets, fragment assembly by ROSETTA (Simons et al. 1997) for ab initio targets, or consensus for everything. We were also not interested in using different methods for different steps, such as Monte Carlo fragment assembly, clustering, selecting, and refinement. Nature does not do this. It does not fit with Occam's razor principle (Li and Vitanyi 1997;Baker 2000).Nature prefers simplicity. We wished to find one theory, one model, as simple as possible that goes from an input sequence to the final structure. This theory should embody homology modeling, threading, fragment assembly (all stages of it), loop modeling, refinement, side-chain packing, and consensus. The theory must be simple, robust, and effective.This work presents our initial efforts in building a theory toward this goal and our preliminary implementation of this theory, FALCON, together with clear-cut experimental results. Some ideas of our work come from three lines of research: fragment assembly, hidden Markov model sampling, and Ramachandran basins.The most successful approach for ab initio structure prediction is to use short structural fragments to model local interactions among the amino acids of a segment and utilize the nonlocal interactions to arrange these short structural fragments to form native-like structures (Simons et al. 1997). Despite the importance of nonlocal interactions in directing the search to discover the nativelike protein structures, the relationship between local structures and the interactions among amino acids within a local structure remain active issues of research. An accurate prediction of the local structural bias for a sequence segment is critically important to protein structure prediction. 4 These authors contributed equally to this work.

From local structure to a global framework: recognition of protein folds

“…FALCON employs a fragment-HMM approach [232] where both the preferred dihedral angles of each residue and sequencestructure relationships from fragment libraries are integrated to predict structural models for a sequence.…”

Can We Determine a Protein Structure Quickly?

“…ROSE-TTA [6] uses structural fragments of size 9 to assembly a protein structure. FALCON [11][12] uses these fragments to train a position specific HMM to model the structure. TESSER [7][8] uses more flexible fragments.…”

“…Let me demonstrate this point via FALCON [11][12] . FALCON uses a position specific HMM as in Fig.1.…”

“…Rather accurate structures can be obtained [6,11,13] . However, before we can consistently obtain high resolution structures and for larger proteins, our structures will continue to be labeled as "predicted targets", and they will have to be verified and improved by experimental methods before they are taken seriously.…”

A Probabilistic and Continuous Model of Protein Conformational Space for Template-Free Modeling

“…This discrete nature may exclude the native fold from the conformational space to be searched since even a slight change in backbone angles can result in a totally different fold. Fragment-HMM (Li et al, 2008), a variant of Robetta, can sample conformations from a continuous space, but still has the coverage problem since the Hidden Markov model (HMM) used in this method is built from 9-mer fragments. The lattice model used in the TOUCHSTONE programs (Kihara et al, 2001;Zhang et al, 2003) does not have the coverage problem, but it samples protein conformations from a 3D lattice with finite resolution.…”