UTSW

HHMI

We work at the interface of biology, computer science, mathematics and physics. Our group specializes in computational biology of proteins and combines sequence and structure analysis with evolutionary considerations to facilitate discoveries of biological significance. Two major directions are pursued:

Development of new computational approaches for analysis of proteins Application of available software tools to biological problems

This duality, i.e. combination of methods development with biological applications is beneficial for both directions, as we frequently find that existing approaches do not give satisfactory answers to specific biological questions; and availability of experts biologists in our group to validate the results in the process, stimulates methods development.

• Introduction •

What is the most important unsolved problem in computational biology of proteins? Apparently, protein energetics. 1) Protein folding problem (i.e. prediction of spatial structure from sequence); 2) precise modeling of interactions between proteins and of proteins with other molecules; 3) quantitative understanding of enzyme catalysis — are all various incarnations of the same challenge scientists were not able to overcome. Despite significant achievements, the exact solution from the physics perspective is still far from reach.

Bioinformatics approaches offer practically useful shortcuts to these problems. Deduction of protein properties (i.e. 3D structure or function) by homology to proteins with known properties has been most successful. For this method to reach its full potential, the following steps should be perfected. We need to: 1) find homologous proteins, i.e. do a database search; 2) compare them to the protein of interest, e.g. make an alignment; 3) decide on the boundaries of property transfer by similarity, i.e. at what level of similarity a property is shared between homologs, and thus can be deduced for a homolog without available experimental information from experimentally characterized homolog. Most projects in the lab deal with these questions.

Homology and evolution are the central themes of our research. By homology we mean similarity caused by common ancestry, not any kind of similarity. Proteins are homologous if they originated from a common ancestor. Similarity caused by other reasons, e.g. structural constraints on 3D packing, is termed analogy. When similarity is weak, it is not trivial to distinguish between the two scenarios leading to similarities between proteins: homology or analogy, and we are working on this problem.

• Main Research Directions •

The long-term objective of our research is to classify available sequence-structure data on proteins into a biologically relevant, hierarchical system analogous to the one currently used in zoology and botany, and to provide computational tools to maintain and update this classification. Since sequence and structural similarities usually imply functional similarity, such classification is of indispensable value for biologists to aid in experimental design. Combination of various approaches is best for addressing complex problems. Thus the questions we pursue are quite diverse and can be summarized as follows.

Biology problems: evolutionary classification of proteins homology vs. analogy, divergence vs. convergence structural fold change in evolution of proteins evolution of function & active sites variability of evolutionary rates between sites & proteins mathematical modeling of sequence & structure evolution definition of protein domains intrinsic structural disorder in proteins non-randomness of protein structure topologies dependence of folding on structure topology interpretation of clinically important mutations sequence-structure-function relationship in protein families

Methodology problems: sequence & profile similarity search clustering of proteins secondary structure delineation 3D structure pattern search multiple sequence alignment 3D structure alignment statistics of protein comparison ancestral sequence reconstruction phylogenetic trees prediction of functional sites protein fold & structure prediction assessment of structure prediction

Our publications give more specific ideas about research directions. We also are involved in many collaborations to study individual protein families, predict properties of proteins, help interpret the effects of clinical mutations and assist in data analysis–driven experimental design.

• Research Highlights •

A. Method development

During the last few years, we have been working on improvement of computational methods to assess protein similarity. We mainly concentrated on two levels of similarity between proteins: sequence and structure. We also directly addressed structure prediction as an important application of homology inference.

A.1. Sequence similarity detection methods

Two main steps constitute sequence analysis. First, homologues are found and separated from unrelated proteins; second, these homologues are aligned. We suggested strategies to improve both steps.

A.1.1. COMPASS – comparison of multiple sequence alignments

Compass is a profile-profile aligner based on a solid probabilistic basis.

A.1.2. Multiple Sequence Alignment

In the process of a tedious search for a tool to provide fast and accurate alignments, we developed this program that up until now demonstrated the best alignment accuracy. PCMA combines COMPASS scoring with the consistency strategy of T-Coffee

B.1.3. Sequence design assists homology searches

Sequence diversity in alignments is essential for profile-based methods. Profile improvement can be achieved by enriching alignments with more sequences sharing a common fold. These sequences can either come from sequencing projects, or be designed in silico. Structure-based design explores small but important covariation between positions, which is being ignored by profile methods. We extended the ideas of Koehl&Levitt to cover sequence-based searches. The design procedure may involve sequences that are expected to fold into a known structure^36,37, or ancestral sequences reconstructed from a multiple sequence alignment³⁸. Implementing both strategies, we demonstrated that they assist homology inference.

To capitalize on these preliminary experiments, we will combine evolutionary and structural design strategies to produce in silico sequences for many protein families, and will carry out sequence searches in the new expanded space.

 

B.2. Structure similarity detection methods

None of the numerous methods proposed for structure similarity search^39-42 achieves what we consider the most fundamental task, i.e. finding structural motifs that satisfy the general definition of a protein fold: globular units with the same secondary structure, topology and architecture irrespective of subtle differences in packing between elements. Such a program should be key in protein structure-mining. The closest available utility, TOPS⁴³ falls short, since it checks topology only. While developing a structural motif search algorithm, we were surprised to find that a method providing broad, yet accurate secondary structure delineation^44,45 was lacking as well.

 

B.2.1. PAL8E – delineation of secondary structural elements

We delineate secondary structural elements using C_a coordinates in order to generate input for the fold search⁴⁶. The elements we define (helices and strands) are linear, so they can be approximated by vectors; and cover about 85% of residues in the structure, so they can provide maximum representation in an element-based search. Our program is predictive by nature, i.e. it does not evaluate the actual existence of hydrogen bonds as DSSP⁴⁴, but rather checks if the general geometry is sufficiently close to provide H-bond presence in some conditions. As such, PAL8E is robust to coordinate errors up to 1.5Å RMSD and is more suitable as the first step to finding remote structural similarities.

 

B.2.2. ProSMoS – Protein Structure Motif Search

Programs that assess structural similarity compare two structures to each other and define common regions^39-42. Structural classification experts look for a particular structural motif instead^10,47. Most programs base similarity scores on superposition and closeness of either coordinates or contacts. Experts pay more attention to the main chain orientation and general match of secondary structural elements. We developed a program that emulates an expert⁴⁸. Starting from a structure, we use PAL8E⁴⁶ to delineate secondary structural elements. A matrix of element interactions (parallel or antiparallel) and handedness of connections is constructed. All structures are reduced to matrices that contain just enough information to define a fold, so the definition is very general and large deviations in coordinates are tolerated. A user supplies a matrix for a motif, and ProSMoS lists all structures that exactly match this motif. Application of ProSMoS to several tasks demonstrated its usefulness^49-51.

We have yet to 1) develop a good scoring function; 2) implement an algorithm that finds partial matches to the motif, and 3) develop an algorithm to produce a residue-to-residue alignment based on detected motif similarity.

 

B.3. Combination of methods for homology detection: SCOPmap

Likely the most difficult task is to identify whether proteins with remotely similar structures share a common ancestor⁹. Done largely by expert analysis, such work is very time-consuming^10,47. To tackle it computationally, we designed a strategy that checks sequence and structure similarity statistics using various existing programs, combines their scores, and attributes a protein structure to a previously defined evolutionary classification⁵². Our automatic method assigns about 95% of proteins to SCOP leaving the rest to expert analysis. Instances of incorrect assignments near domain boundaries need more work.

 

B.4. Structure modeling methods

After establishing a homology link, the structure of one protein can be modeled using the other one as a template⁵³. We tried to improve scoring functions for several specific problems in homology modeling.

 

B.4.1. Side-chain modeling

Placement of given side-chains on a fixed backbone (=packing prediction) is a necessary step towards a realistic protein model⁵⁴. Careful refinement of weights for “energy” terms and usage of contact surface/overlapped volume allowed us to obtain a better side-chain rotamer predictor⁵⁵, which has been used by many other researchers, e.g.^56-58.

 

B.4.2. Sequence design

Finding optimal side-chains for a fixed backbone (=protein sequence prediction) is a natural extension of the side-chain modeling problem⁵⁹. Parameters refined on a set of high resolution structures correlated well with experimental data on mutant T4 lysozymes stability³⁷. We used the designed sequences to assist in remote homology detection.

We plan to 1) incorporate restricted backbone moves in the design algorithm; 2) as a collaboration with Dr. Zhang, to determine the X-ray structure of a protein with a new fold that we constructed de novo from short fragments of the backbone and designed a sequence for with our approach.

 

B.4.3. Local structure prediction

Fragment-based structure prediction methods⁶⁰ explore “horizontal” (local sequence) information in protein sequences, in contrast to sequence alignment methods that use “vertical” (positional) information assuming independence between positions⁶¹. To improve homology detection and sequence alignment construction using local conformation predictions, we developed a local fragment predictor²⁸.

We plan to add statistical significance estimation of local profile hits and to incorporate it into alignment algorithms.

 

B.5. Methods to evaluate the quality of structural models

Our group was invited to judge predictions in Critical Assessment of Structure Prediction competition (CASP5)⁶². Evaluating thousands of structural models, we found⁶³ that a consensus method combining many scores generated by various structure similarity scoring schemes correlates best with manual assessment of models that has been a standard in former CASPs. Approaches we developed have been successfully used by different assessors in last year’s CASP6.

 

C. Application to biological problems

As our main scientific interest is to find novel remote homology links between proteins, we applied various computational methods enhanced with the manual expert analysis to explore sequence and structure databases.

 

C.1. Large-scale evolutionary classification of proteins

The ultimate goal of our research is to classify all protein domains according to their evolutionary history. Moving towards this goal, we selected several protein groups for in-depth analysis. These groups can be structural, i.e. all proteins of a particular fold; functional, i.e. all proteins of a particular function; or genome-based, i.e. all proteins in a particular genome.

 

C.1.1. Fold-based

Small protein domains (20-60 residues) are notoriously difficult to analyze due to unreliable similarity statistics caused by insufficient numbers of residues⁶⁴. In an attempt to bring order to one of the small protein groups, we classified all zinc finger structures.

 

C.1.1.1. Zinc fingers

We define zinc fingers (ZF) as small domains where zinc contributes to the protein structural core. We classify all ZF structures into 8 fold groups based on structural similarity around the zinc-binding site⁶⁵. Each fold-group encompasses 1 to 11 evolutionary families. Prior to our work, there has been no targeted attempt to comprehensively classify all ZFs. However, rudiments of such classification existed within general structure databases^10,47.

We are currently extending our experience with ZFs over another very large group of small proteins: disulphide-rich domains.

 

C.1.1.2. Thioredoxin fold

For larger domains with the core formed by secondary structural elements, classification could start with ProSMoS search⁴⁸, as we have done for thioredoxin fold, an a/b domain of 6 secondary structural elements⁵⁰. Our definition of fold allows circular permutations, i.e. geometric transformations that link existing N- and C-termini of the chain and introduce new termini at another point. Since permutations regularly happen in protein evolution⁶⁶, their inclusion in the fold definition ensures completeness of coverage. We unify about 1000 protein structures from 5 different SCOP folds¹⁰, and divide them into 11 evolutionary families. For many of these proteins, structural similarity to thioredoxin remained unreported.

We are working on classification of RNAse H and Rossmann folds, and will plan to cover other wide-spread folds, such as Greek-key immunoglobulin-like, ferredoxin-like and a-folded leaf.

 

C.1.2. Function-based: kinases

Bringing together all proteins that share significant functional similarities and looking at the whole group from evolutionary perspective is informative for understanding both function and evolution^67,68. In addition, such work results in many challenging structure predictions. Our study of kinases defined as enzymes transferring terminal P-group from ATP to another molecule is included among the five references^69,70. Although protein kinases are frequently analyzed and classified⁷¹, no comprehensive classification of all kinases, including many families of metabolic small molecule kinases, existed.

We are extending our experience with kinases on several evolutionary diverse protein functional classes, such as proteases, phosphatases and acyltransferases.

 

C.1.3. Genome-based: human Y chromosome

Computational genomics necessitates annotation of proteins encoded in complete genomes⁷². Meticulous expert-driven study of proteins from a single organism is capable of finding interesting functional annotations and structure predictions that are missed in large-scale automatic analyses⁷³. We applied our experience to human Y-chromosome⁷⁴ (see one of the five publications).

We are planning to undertake diligent domain analysis of all human proteins with the goal of evolutionary classification and structural prediction, crucial for our understanding of human biology.

 

C.2. Studies of individual protein families

This aspect of our work is arguably closest to the bench and richest in immediate biological applications. Taking an individual protein and a few of its relatives, we embark upon solving a problem that other researchers pursue at the bench, namely, we try to find out what that protein does and how it functions. Many of such cases result in unexpected structure predictions.

 

C.2.1. Structure predictions

During the last few years, our studies of two dozen families resulted in non-trivial and yet confident structure predictions^75-90 Many of our published predictions remain unverified by experimental structures. However, more structures become available every year. Some of these correspond to proteins for which we have published predictions. On occasion, structural biologists discuss our predictions in their publications describing experimental structures⁹¹. Several such articles are included in the Appendix to this report. So far, not a single fold prediction reported by us has been found incorrect. Naturally, careful analysis of alignments reveals that some details of our models are not always accurate. The most interesting deviation from the model was found by crystal structures of GyrA and ParC domains^92,93 that were predicted to form six-bladed b–propellers⁸². Crystal structures revealed that both domains are indeed comprised of six blades with 4 b–strands each. However, each blade is involved in a swap of 2 b–strands with the adjacent blade. This swap could not be deduced by us.

We plan to continue our work on non-trivial structure predictions of interesting domains.

 

C.2.2. Homology predictions

The most important aspect of our research is remote homology prediction, i.e. deduction that proteins A and B share a common ancestor despite low similarity between them. Although such statement cannot be verified by direct experimentation, it predicts the possibility of existence of a protein C, such that both links between A and C, and between C and B are strong and clear. Thus, homology between A and B can be verified by transitivity. Protein C may be currently unknown, but frequently its sequence or structure eventually becomes available. The best argument for remote homology prediction is made when multiple evidence (sequence, structure, function) is considered^94-103. Despite remoteness of homology, it could be helpful for the functional understanding of a protein. One such example, MH1 domain of SMAD, is discussed in an included publication¹⁰⁰.

Our runs of SCOPmap program⁵² on weekly updates of PDB frequently produce possible, but not fully justified assignments. We will be inspecting them to find new non-trivial homology links that we could support by manual analysis.

 

C.3. Fold changes in protein evolution

Analyzing distantly related proteins, at times we see structural changes large enough to warrant placing homologues in different folds. We were pioneers in recognizing this universal trend^104,105. We try to understand how protein structures, usually considered to be rigid, change in evolution^106,107. One particular aspect of the problem is distinction between homologues and analogues, and finding examples of analogues^108-110.

Our goal for the next several years is to comprehensively catalogue all instances of homologues with significant structural differences, and to understand mechanisms used most commonly by evolution to change protein structures.

 

to be more efficient - predone classifications where expert analyze. because computers are not very good at it Evolutionary classification of proteins. Why remote homology is important: 1. Majority of proteins are remote homologs Can we teach computers to do as good as experts? Bioinformatics algorithms provide biologists with powerful tools to gain insights into biological functions.

• More info about significant projects •

Disclaimer: we try our best to update these pages, but inevitably fall behind. Please tell us what you want to see.