Accelerating and enhancing the PSIPRED Workbench with deep learning

With the growing number of completely sequenced genomes, life scientists now face the challenge of characterizing the biological role of the encoded proteins as to advance our understanding of cell physiology. Most genes are designed to code for proteins which have useful functions in an organism. Proteins are essentially strings of simpler molecules, called amino acids and these strings can self-assemble into a complex 3-D structure as soon as the protein is formed by the protein-making machinery (ribosomes) in the cell. It is this unique structure which determines the precise chemical function of the protein (i.e. what is does in the cell and how it does it). By firing X-rays at crystallised proteins, scientists can determine their structure, but this process can take many months or even years. With hundreds of thousands of proteins for which the native structure is unknown, it is not surprising that scientists want to find a clever shortcut to working out the structure of proteins. We, like many other scientists have been trying to "crack the code" of protein structure i.e. working out the rules which govern how the protein finds its unique structure and then trying to program a computer with these rules to allow scientists to quickly "predict" what the structure of their protein of interest might be.

The PSIPRED Workbench is a collection of Web servers maintained at UCL which does just this i.e. it allows biologists to predict the structure of their protein structure given just its amino acid sequence. Over the years it has helped many thousands of scientists with their work by providing these services and we now wish not only to upgrade and maintain these existing servers but also to implement new methods which allow the structures of even the most difficult proteins to be deduced by computer simulations.

More recently, however, PSIPRED has been given a wider range of features to cover other important problems in biology. For example, using PSIPRED, a scientist can predict which proteins do not fold into stable shapes (called disordered proteins) or which chemical substances are likely to bind to a protein. Even where a protein does not appear to fold into a single stable structure, PSIPRED can still help scientists deduce what the function of his or her protein is likely to be. Generating such information on a large scale using computer algorithms can help expand our knowledge base of the biological role of proteins at a cellular level, and such understanding will be a key stepping stone in the development of techniques and pharmaceuticals to target diseased genes and their products as well as proteins from pathological organisms such as bacteria or viruses. In a similar way, knowledge on the function of certain bacterial genes can, for example, help develop new industrial processes by modifying the genes to make them produce novel chemical compounds, or even helping to detoxify industrial waste by producing friendly bacteria that can use the poisonous chemicals as food.

The Jones Group at UCL has been developing a widely-used suite of web-based tools based on cutting edge protein structure prediction methods since 1998. The methods allow users to predict a variety of protein structural features, including secondary structure and natively disordered regions, protein domain boundaries and 3D models of tertiary structure. More recently we have been developing new services to assist users in prediction gene function and protein-protein interactions - all of which we believe are vital developments to make PSIPRED a vital and unique tool for biologists.

PSIPRED employs a number of features to help users become familiar with the software e.g. via online tutorials. Through work done in the original BBR grant, we have successfully integrated our suite of tools, resulting in the only single site worldwide which, after learning one simple user interface, provides all of the following prediction services to biologists: comparative modelling, fold recognition, ab initio (new fold) prediction, transmembrane protein structure prediction, disorder prediction, domain boundary prediction, binding hotspot prediction, ligand binding site prediction, and several novel approaches to gene function prediction.

The key bioinformatics developments in this proposal will be to harness the power of deep learning methods to hugely accelerate the PSIPRED toolset. In particular we plan to circumvent the need for running time consuming databank searches using PSIBLAST or HHblits by using sequence-sequence learning models to maintain a continuously updated embedding of UniProt. This means we will be able to extract sequence neighbourhood information for any given sequence with a related sequence in UniProt almost instantaneously, compared to minutes or even hours using standard sequence databank searching methods. Furthermore we wish to implement user-friendly implementations of the recent breakthroughs in deep learning-based covariation-based modelling.

SUMMARY OF RESOURCE

This proposal is to maintain and further develop a set of Web-accessible tools and services that has been developed at UCL (namely the PSIPRED Workbench - originally called the PSIPRED Server) since 2001 (and at the University Warwick since 1998). This portal provides a wide variety of very well-known tools (e.g. PSIPRED/DISOPRED/GenTHREADER/MEMSAT/FFPRED) to the general life science research community, and is available for use (free of charge) to both academic and commercial researchers. In many independent tests (e.g. every CASP experiment since 1994), these tools have proven to be amongst the very best worldwide, and are widely used by other resources around the world as part of their own pipelines and workflows. The PSIPRED Workbench is probably one of the most widely accepted and used bioinformatics resources that is operated from a UK University, and is frequently referenced in many textbooks and training courses. The close association between a world-class bioinformatics research group and such a widely-used tool means that the methods are kept fully up to date with changing technological and demand-based trends.

IMPACT OVERVIEW

The PSIPRED portal was used over 170,000 times in the last year, with nearly 1000 jobs handled per day during busy periods, and has over 5,200 unique visitors per month. The overall usage is up 20% since our last application to the BBR Fund, which demonstrates a clear growth in demand. Users are spread further across the globe than before, with 18% of users coming from the US and 9% of users from the UK. This testifies to the importance of this resource, particularly to the UK bioscience community given the ratio of researcher headcounts in the two countries. Users typically also come from a wide variety of scientific research areas. Based on our user support enquiries and user surveys, we can identify users in areas across the whole BBSRC remit e.g. bio-energy, ageing research, biotechnology, synthetic biology, vaccine design, plant biology, animal health and even nanotechnology.

In summary, the immediate beneficiaries of this research are the broad community of experimental biologists needing additional functional or structural clues for proteins of interest. Both academic and industry scientists will benefit in a similar way as the results of this research will be available freely to all users. Commercial scientists with sensitive data, or companies wishing to released closed-source code, will be able to license the software through UCL Business so that they can exploit the resource without revealing their research interests to other users. Being able to determine even some clue as to the structure or function of uncharacterised proteins can have significant impact in the broad variety of areas mentioned above.

Beyond industrial applications of this research, filling in the major gaps in our knowledge of what the full complement of genes and the products of these genes do and how the proteins interact can have wider implications in understanding the working of healthy cells and how they age. Ultimately this work can make a contribution to our overall understanding of how life processes arise from interactions between a relatively small number of genes in our genomes and the genomes of other organisms.

We also note that many users of our servers use the resources for teaching purposes. It's clearly vital that for maximum impact, the next generations of graduates and postgraduates in the biosciences be trained in advanced computational biology techniques. We are therefore pleased that our tools, because of our focus on good quality visual output and speed of returning jobs, find use in teaching laboratories around the world.