Current Research

My primary research goal is the understanding of the relationship between the amino acid sequence of a protein and its consequent three-dimensional structure and function. To this end I have been developing computational methods for the prediction of protein structure and function from sequence.

Currently I am developing a system for the prediction of protein-ligand specificity using a new machine learning technique called Support Vector Inductive Logic Programming (SVILP).

Fold recognition: Phyre

Myself together with PhD students Riccardo Bennett-Lovesy, Alex Herbert and Dr. Kieran Fleming, have developed a new protein fold recognition system (Phyre) to replace and surpass the popular 3D-PSSM system that I and Dr. Bob MacCallum developed 3 years ago.

Ab initio structure prediction using fragments.

Machine learning

Given the overwhelming amount of empirical information from protein sequence and structure databases, and the lack of substantial progress in understanding the protein folding problem from physical first principles, it is imperative to use modern techniques from the field of artificial intelligence and data mining to develop practical solutions to this problem.

In particular, I am interested in using Support Vector Machines, Random Forests, Graph Theory, and statistical schemes such as Monte Carlo simulation, Gibbs Sampling, Genetic algorithms and empirically derived potentials based on the Boltzmann hypothesis.

Pointless Research Finding of the Month April 2008

Well as anti-climaxes go, this ranks highly.

My PhD involved the development of various methods of analysing ensembles of protein structures determined by nuclear magnetic resonance spectroscopy. I was supervised by Dr. Mike Sutcliffe.

Designed and programmed NMRCLUST (Kelley et al., 1996), a tool to cluster ensembles of NMR-derived protein structures into conformationally-related sub-families. This required the development of an automated method of clustering without rigid cut-offs or user intervention. Spent two months in the Bioinformatics group at Oxford Molecular Ltd. incorporating NMRCLUST into Architect, the IDITIS protein database generation tool. Designed and programmed NMRCORE (Kelley et al., 1997), a tool to define automatically the core atoms and domains in an ensemble of protein structures. Developed OLDERADO: On Line Database of Ensemble Representatives And DOmains. This is a searchable database of the results of NMRCORE and NMRCLUST on the current set of PDB-deposited NMR-derived ensembles.

Protein Structure Prediction/Remote homology detection

Designed and developed the fold recognition algorithm (3D-PSSM) to detect remote homology relationships between an uncharacterised protein sequence and proteins of known structure. This has been my major focus to date. The system is described in detail in (Kelley et al., 2000). Briefly, the system scores the match between a user's sequence and each structure in a representative fold library. The scoring methodology involves the use of predicted and known secondary structure, sequence profiles (PSSMs) generated by the widely-used program PSI-Blast, a solvation potential, and an extended structural profile which we have called a 3D-PSSM. These are generated by using structural superpositions of proteins known to occupy the same structural superfamily as determined by the protein experts at SCOP (Structural Classification of Proteins). The structural superposition of proteins sharing very little sequence similarity permits the construction of sequence profiles that better represent the diversity of a given superfamily and thus cover a larger area of sequence space. All these factors are scored using a modified dynamic programming algorithm.

Designed and worked on the 3D-Crunch supercomputing project to assign folds to the then known bacterial genomes in collaboration with Dr. Manuel Peitsch.

Took part in CASP 3 and CASP 4 This is a meeting held every 2 years as an international blind trial of protein structure prediction techniques. At CASP4 3D-PSSM was found to be the best-performing fully automatic method for structure prediction. In addition, our manual, human-crafted predictions were ranked 3rd out of the 100+ groups attending.

Using textual annotation information

  SAWTED

There is a wealth of largely untapped information available on proteins in the form of human annotation and journal abstracts. Part of the reason for the superior performance of experts using structure prediction programs such as 3D-PSSM over the purely automatic use of such algorithms is the human expert's ability to read textual annotation and otherwise human-readable information. Myself and Dr. Bob Maccallum developed the SAWTED algorithm (MacCallum et al., 1999). The purpose of SAWTED is to use the extent of shared human-assigned keywords between potentially remote homologues to lend confidence to tenuous homology assignments. The SwissProt homologues of a user's query sequence may contain keywords such as "cytochrome" or "p450". Similarly, the SwissProt homologues of a known structure may contain keywords such as "cytochrome" and "mitochondria". These terms are represented as abstract "term vectors" which can be compared using the vector cosine model of text retrieval. Despite a very weak sequence match between these two proteins using an algorithm such as PSI-Blast, the shared keyword or SAWTED score can automatically provide an independent source of evidence for genuine homology.

Continuing on with this idea of using textual information automatically, Dr. Ben Stapley and myself used tools from machine learning (specifically Support Vector Machines (SVMs)) in conjunction with collated Medline abstracts to represent proteins as very high-dimensional (40,000+) vectors of English language terms. SVMs are binary classifiers that permit such large spaces to be quickly and automatically partitioned given training data. Once trained such systems can be used to classify new proteins based on text alone, or combinations of text and sequence features. (Stapley et al., 2001; Stapley et al., 2002).