Sequence-based learning algorithms for understanding and improving protein characteristics

Abstract

The development of high-throughput measurement techniques resulted in rapidlyincreasing amounts of biological data, which made computational methodsessential for biological research. Hence, the field of bioinformatics emergedthat since plays an important role in storing, making accessible, integrating,and analysing different types of biological data. Recently, the use ofcomputational methods to (re)design biological molecules is emerging. Thisthesis describes an approach to determine which protein features influenceproduction and secretion under industrial conditions, and how these results areused to redesign proteins in order to enhance their production levels. The starting point of this thesis work was a set of measured extracellularconcentrations of proteins that were produced under the condition ofover-expression in Aspergillus niger, a filamentous fungus that is often used forindustrial protein production because of its excellent secretion ability. Usingthis data, classifiers where developed that can predict successful proteinover-production based on its sequence. In practice, these classifiers can beused to select proteins that have potential for industrial production. Subsequent classifier analysis was used to infer what sequence-propertiesrelate to low and high production levels respectively. While taking intoaccount a large set of sequence-derived characteristics, the amino acidcomposition, i.e. the relative occurrence of the twenty different amino acidsin a protein, was found to be best discriminating. Classifier analysis showedthe importance of tyrosines for high production levels, whereas lysines andmethionines were found to occur relatively often in proteins with lowproduction levels. A similar classifier-analysis method was also used to findcharacteristic properties of the sequence surrounding neutral anddisease-associated missense mutations in humans. One of the developed classifiers was used in a protein redesign approach toselect amino acid substitutions that are expected to have a positive effect ona protein's production level, while additional objectives and restrictions wereused to reduce the probability that the amino acid substitution will affect theprotein structure. Application of this method resulted in up to ten-foldextracellular concentrations for two A. niger enzymes. These results show thatthis approach is a promising new tool to improve production levels ofindustrial enzymes. Furthermore, additional research to the redesigned proteinsmight help to better understand the mechanisms involved in protein productionand secretion in A. niger.