Biopharmaceutical manufacturing is one of largest sectors in the world and purification steps are expensive. Packed-bed resins are widely used, but are limited by diffusion mass transfer. Convective mass transfer media offer improved productivities using high flowrates. Electrospun nanofibres are a non-woven with an open structure and high surface area. Cellulose acetate was electrospun into reproducible adsorbents and activation methodologies were evaluated. Aldehyde activation caused degradation. Epoxy, carboxyl and cyanuric chloride activations were successful and recommended for cellulose nanofibres. Compressing cellulose acetate nanofibres improved mechanical strength. Functionalisation to diethylaminoethyl and carboxyl adsorbents showed the highest DBC10% values at the lowest compression load of 1 MPa at 20 and 27 mg /mL, respectively. However, at this load the DBC decreased for increasing flowrate, whereas, the DBCs of higher loads were consistent. Glucose isomerase is an important industrial enzyme used as a sweetener in the drinks industry. Immobilised glucose isomerase by epoxy activation showed similar initial activities in static and dynamic assays, which represented batch and flowthrough process, respectively. The activity of free enzyme was notably higher in the static assay than the immobilised, but the poorer mixing in the dynamic assay reduced its activity. The retained activity of nanofibre-immobilised glucose isomerase in the dynamic assay supports its application to flowthrough reactors. Protein A chromatography is an expensive stage in antibody manufacture and was immobilised by cyanuric chloride and spacer arms. For comparison, an aldehyde activation was developed with glycidol modification, which was preferred to that of cyanuric chloride, recording equilibrium and DBC10% capacities of 4.37 and 2.99 mg/mL, respectively. The versatility of electrospun cellulose adsorbents are characterised in protein binding and biocatalysis applications, showing static and dynamic operations. The continued development of nanofibres into largescale systems and a whole process context will reveal their full benefits to bioprocessing.

The spontaneous formation of aggregates presents an obstacle in the developability, manufacturability and long term storage of numerous therapeutic proteins. These aggregates are a manifestation of the drug molecule's physical instability, potentially leading to reduced biological activity, differences in solution properties or increased immunogenic potential. Currently our understanding of the mechanisms behind aggregation and the role of a protein's extrinsic environment upon the molecule's aggregation propensity remains limited. In this study we investigated the aggregation behavior of a previously unstudied commercial biopharmaceutical, an α-TNF Fab' antibody fragment. Initially we mapped the differences in aggregation kinetics and product morphologies accompanying changes to the protein's aqueous environment, finding that the aggregation pathway adopted by the molecule depended upon the pH of the system. We then employed a combination of intrinsic and colloidal stability measurements to probe the causes of the behavioral differences seen. Together, these biophysical properties quantitatively captured the relative differences in aggregation rates observed. Furthermore, we were able to infer the existence of various partially folded intermediates whose populations govern the types and rates of aggregates formed. Finally, we sought to identify whether observable changes in the native state conformation precluded aggregate formation through low resolution structural analysis using small angle X-ray scattering. We showed that a drop in pH generated a conformationally expanded state which appeared to coincide with the titration of a key histidine residue within the protein's structure. We predict that an equivalent mechanism will lead to aggregation at low pH of other human derived antigen binding fragments.

The inter-disciplinary work in this thesis entails the development of a microfluidic device with bespoke imaging methodology to study flocculation growth kinetics dynamically in real-time. Flocculation is an advantageous downstream operation that increases the product-separation efficiency by selectively removing impurities. Yet, there is no unifying model defining the effect of different physico-chemical parameters on the rates of flocculation. Conventional setups for said analyses require large experimental space that are tedious to perform, and are limited by their dependence on end-point analysis, requiring sample-handling and further dispersion into typical particle-sizing instruments. In spite of the counter-intuitiveness of implementing microfluidics to study flocculation due to the anticipated channel-clogging issue, it is hypothesised that the growth kinetics can be measured by achieving a continuous, steady-state flocculation under a lower-shear environment. Flocculation within a spiral microfluidic device (~151.8 µl volume) is evaluated against a bench-scale setup (~50 ml volume) through the comparison of floc size and zeta potential. The fluid hydrodynamics in the microchannel is assessed by an experimental mixing-time analysis (tmix = 7.5 s) and a residence time distribution study (tm = ca. 70 s). In situ measurement of floc size and morphology is facilitated through high-speed imaging, with an image-processing script for robust analysis. Different flocculants are tested and growth rates calculated (~ 8 and ~12 µm s-1 for PEI and pDADMAC). Flocs grew linearly up to 250 µm for cationic polymers, while no growth was observed with a non-ionic PEG. Using an improved parameter-fitting step, the growth rates are compared to a simplified model for monodisperse perikinetic flocculation. The work presented should thus, enable an experimental estimation of flocculation growth kinetics and pave way for the development of accurate flocculation models for polydisperse particles. The developed system also facilitates a rapid screening of new flocculants useful for quicker process development.

The manufacture of high quality virus at an industrial scale remains a challenge. In the present study, the ability of nanofiber adsorbent functionalised with quaternary (Q) amine ligands to purify adenovirus 5 (Ad5) from crude feeds was investigated. The hypothesis for this study is that nanofibres modified with Q amine ligands will enhance purification of Ad5, enabling recovery of the virus at high, infectious yields from crude cell lysate feeds. The nanofiber adsorbent technology are non-woven regenerated cellulose nanofibers that present an open pore structure 0.2-2.0µm with large inter-fiber space and shallow bed height. These are features previously reported to be advantageous for separation of viral vectors. The upstream process for propagation of Ad5 stocks involves use of a HEK293 cell line that can allow proliferation of the virus as it contains the E1 gene deleted from Ad5. Optimal culture conditions for reproducible production of infective Ad5 from HEK293 are poorly characterised. Therefore, we performed investigations of the influence of cell passage (P: P2, P5, P10), metabolic activity (mitochondrial activity: MTT assay), and rate of proliferation on generation of infective Ad5. HEK293 passage influenced the yield of Ad5 produced. Cells of a low passage (P2) generated lower amounts of virus than those of higher passage (P5, P10). This was due to lower growth rates of P2 cells than P5 and P10 cells. The intermediate passage cells (P10) presented with the highest growth rates that were significantly greater (P ≤ 0.05) than that of P2 cells. The mitochondrial activity of P2 cells were lower than that of P10 cells but P5 cells presented with greater mitochondrial activity than P10 cells. HEK293 cells were used at passages 10-15 to propagate Ad5 for the nanofibre optimisation and screening phases of the study. The Q amine modified nanofibres are novel and therefore preliminary investigations of their function for separation of proteins and viruses are necessary. Conditions for the use of nanofibre were optimised initially by assessing the ability of those matrices to purify bovine serum albumin (BSA) and thyroglobulin. High recovery separations of BSA and thyroglobulin were achieved with nanofibres compared to Q-Sepharose and POROS chromatography materials. A number of gradient elution methods were applied to purify Ad5 from a clarified bulk feed lysate. Ad5 was purified from a clarified bulk feed and protein VIII, a protein marker for infective mature capsids was identified (via mass spectrometry) in a high salt elution fraction. Our study is the first to describe evidence for identification of a cleaved mature capsid protein in a preparative chromatographic purification step. Evaluations of the effects of process variables that may induce stress on Ad5 during separation procedures showed that high salt concentrations in elution buffers and increased flow rates did not affect Ad5 recovery and infective yield. Lowering the Ad5 feed volume from 50mL to 20 mL lead to improvements in resolution of eluted virus peaks and increases in infective Ad5 yields from 69% to 78%. Further optimisation of nanofiber-mediated purification of Ad5 and screening of nanofibers functionalised with Q amine ligands at low (440 µg/mol), medium (750 µg/mol) and high (1029 µg/mol) densities was performed. Clarified crude and filtered (500 kDa hollow fiber tangential flow filtration system) Ad5 feeds were purified using nanofibers at low, medium and high Q amine ligand densities. Results showed that by maintaining short process times, infective Ad5 recoveries of over 90% were achieved and those are the highest infective recoveries of Ad5 achieved to date. Prolonged adsorption durations on Q amine nanofibers showed significant losses in Ad5 product quality on medium and high ligand densities over extended binding durations of up to 24 min. Each ligand density produced several Ad5 populations over a single run with unique infective ratios (1-16.04 virus particle/infectious virus particle). Increasing Q amine ligand density improved resolution and separation of intact Ad5 capsids from host cell protein impurities and product-related impurities including free hexon (a major capsid coat protein) and replication-defective Ad5 capsids that contained DNA. Using 0.125mL adsorbent, flow rates in excess of 70 mL/min could be applied to nanofiber adsorbents for separation of Ad5, indicating efficiency in the purification workflow. Nanofiber nanofibers exhibited high dynamic binding capacity for Ad5 vectors in excess of 2.39 x 1010 virus particles. This study demonstrates the utility of Q amine functionalised nanofibre nanofibers for high recovery purification of infective Ad5. The bioprocess workflow devised for separation of Ad5 from crude cell lysate feed generated from an optimised upstream process can be scaled up for industrial manufacture of therapeutic Ad5 containing genetic payloads. Quaternary amine functionalised nanofibre nanofibers present features that clearly indicate their potential as next generation bioprocessing tools.

Transketolase is a ubiquitous enzyme of the thiamine diphosphate-dependent (TPP) family involved in the Calvin cycle and the pentose phosphate pathway. Substrate-walking of E. coli transketolase progressively shifted the target acceptor substrate from phosphorylated aldehydes to non-polar aromatic aldehydes. However, its applicability as an industrial biocatalyst is limited by the lack of combination mutants exhibiting satisfactory substrate breadth and stability. The S385Y/D469T/R520Q variant, which had previously been thought to exhibit differential binding to aromatic substrates, was analysed. Three model substrates were docked into its active site thus revealing two binding pockets supporting π-π stacking interactions. Screening of this variant with other cyclic compounds revealed evolved activities towards valuable industrial building blocks including 4-(methylsulfonyl)benzaldehyde (4-MSBA), a precursor to thiamphenicol. A quadruple mutant was consequently engineered by recombining a stabilising mutation and used as a template for further evolution towards bulky aromatics. Site-directed mutagenesis of a key residue generated the H192P/S385Y/L466M/D469T/R520Q variant which exhibited 5.6-fold improved kinetics towards 4-MSBA compared to the triple mutant. The transition of TK from a model enzyme to a robust industrial biocatalyst however does not only rely on its ability to synthesise novel therapeutic molecules, but also on its thermo- and solvent-stability. 52 variants of TK across the tree of life were consequently aligned to engineer a consensus variant and reconstruct a common ancestor to TK speculated to have branched from proteobacteria, firmicutes and fungi. The resulting common ancestor exhibited trace levels of non-native activity towards non-phosphorylated sugars and provided an initial soluble enzyme to explore the stability/activity relationship of future de novo TKs.

Pichia pastoris (P. pastoris) is an attractive industrial host cell due to its ability to grow up to 60% wet cell weight (WCW) by volume, a far higher level of biomass than the typical values reached by Escherichia coli (E. coli) and Saccharomyces cerevisiae (S. cerevisiae). This thesis seeks to explore how the genetic tractability and high cell densities characteristic of P. pastoris can be exploited to intensify whole-cell biocatalysis. Chiral amino alcohols such as 2-amino-1,3,4-butanetriol (ABT) are key building blocks of small molecule pharmaceuticals and have previously been produced by whole-cell biocatalysis using cells engineered to express a de novo enzyme pathway consisting of transketolase and transaminase. Within this work, native and foreign P. pastoris transaminases were characterised with respect to their biocatalytic potential. Genomic data mining was performed to explore the GS115 strain genome, allowing the selection of three putative Class III transaminase genes and the construction of overexpressor strains PpTAm107, PpTAm677 and PpTAm410. The well-studied ω-transaminase CV2025 from Chromobacterium violaceum (C. violaceum) was also successfully engineered to generate two strains; PpTAmCV708 for single expression of CV2025, and PpTAm-TK16 strain for CV2025 co-expression alongside a native transketolase previously characterised for L-erythrulose production. The rapid growth and high biomass characteristics of P. pastoris were successfully exploited for production of ABT by whole-cell biocatalysis. At high cell density, the best performance for the de novo pathway was obtained with the engineered PpTAm-TK16 strain, which tolerated high concentration of substrates to achieve STY 0.29 g L-1 h-1 of ACP, 49-fold higher than levels previously achieved with E. coli in the presence of the same substrates.

Polishing chromatography is a critical element of a bioprocess, because it is currently the only scalable separation technique that can remove process-related impurities, thereby achieving the high purity required of a biotherapeutic. Optimising the polishing chromatography of complex feeds has not been systematically addressed in the literature. This thesis identified a novel, academically affordable ternary protein mixture and systematically developed an optimal two-column polishing train for it. The ternary protein feed mixture was selected using many criteria, but had no special feature to aid identification, such as a chromophore, making it more difficult to characterise. The resulting analytical chromatogram could not be fully resolved, which is typical of industrially relevant products, such as glycoproteins. The selected HPLC column produced fast separations, resulting in a comparatively rapid quantification of preparative chromatograms. Many chromatographic resins and operating conditions were screened, resulting in the non-obvious sequence a hydrophobic interaction (HIC) followed by an anion-exchange (AX) adsorbent. Systematic experimental studies optimised the sequence with respect to yield, purity and amount recovered. Although the loading exceeded the binding capacity of the HIC column, runs at extremely high loadings (60 - 150 g/L) gave very efficient separation in an unusual combination of flow-through and bind-and-elute modes. It was found to achieve >200 mg of acceptably pure product from a single run. A variety of problems were encountered during the development of this polishing train, to which solutions were developed. While these problems are not uncommon, the literature does not contain systematic solutions to them. Examples include decisions about sequence design, protein solubility issues, and the detailed characterisation of samples from preparative runs (not achieved by analytical HPLC). In particular, a system-specific deconvolution methodology was developed that allowed complete characterisation of the mixture; the approach is likely to be widely applicable to industrially relevant biological feed mixtures.

The human motor system controls a large number of independent degrees of freedom simultaneously, and is capable of learning a seemingly infinite amount of movement skills, vastly surpassing such abilities of any man-made robot. The computational and neuronal mechanisms of human dexterity and adaptation abilities remain elusive. It has been recently suggested that one of the computational brain mechanisms allowing such a rich movement repertoire might be “structure learning” (Braun et al., 2009b). After extensive practice with motor tasks sharing structural similarities (e.g. different dancing movements, or different sword techniques), new tasks of the same type can be learnt faster. According to the structure learning hypothesis, such rapid generalisation of related motor skills relies on learning the dynamic and kinematic relationships shared by this set of skills. As a consequence, motor adaptation becomes constrained, effectively leading to a dimensionality reduction of the learning problem; at the same time, adaptation to tasks lying outside the structure becomes biased towards the structure. We tested these predictions by investigating how previously learnt structures influence subsequent motor adaptation and found that after extensive training with both kinematic or dynamic perturbations, adaptation to unpractised, diagonal, perturbations happened along the previously learnt structure (vertical or horizontal), and resulting adaptation trajectories were curved. We further present several computational models that can account for this behaviour: correlated distribution of motor primitives, changed Bayesian prior or Bayesian network with a hidden variable. These models make different predictions with respect to structure extrapolation; in a series of experiments we did not observe any evidence for structure extrapolation and conclude that the observed effects are probably explained by the changed Bayesian priors. Finally, we present a series of experiments on path tracking, where subjects develop a skill of path tracking in the absence of any external perturbations. Relationship with structure learning is discussed.

The brain controls behaviour and has to manage the body’s resources (including energy) at the same time. How the brain coordinates and combines computations for controlling behaviour in response to metabolic state is little understood. I examined internal metabolic state and its role in motor coordination. I found that internal metabolic state modulates human motor coordination, with a lower energy expenditure associated with performing a velocity-controlled centre-out reaching task when in a low metabolic state. One approach to understanding human motor coordination is to consider motor cost functions. Many cost functions have been proposed yet the form and implementation of the cost function in the human motor system remains largely unknown. I have shown how an approximately quadratic metabolic energy cost function can be derived from the physiological properties of muscles and muscle fibres, producing a biophysically plausible cost of motor control. I then used this cost function to predict the manner in which coordination would change during an isometric force production task. I showed my predictions were correct, with motor effort shifting from muscles with higher metabolic energy costs towards muscles with lower metabolic energy costs. I examined the effect of internal metabolic state on muscle fibre recruitment regimes. I found no significant effect here, suggesting that fibre recruitment is computed in an independent manner to muscle coordination and supporting hierarchical control of human motor coordination. To directly uncover the composite cost function of reaching movements, I used model based inverse optimal control to show how differences in hand reaching trajectories between metabolic states can be described by a single parameter representing a trade-off between motor variability and energy expenditure.

Evidence gathered across various animal models suggests that in the cerebellum, sensory and motor coding integrate. I therefore aimed to characterise the activity of cerebellar vermis neurons in mice performing a task involving both motor and sensory learning. I used multi-electrode-arrays to record from populations of cerebellar neurons in head-fixed mice running in a virtual reality (VR) environment. I show that cerebellar vermis neurons modulate their activity according to the speed of locomotion, having either a positive or negative relationship with increasing speed or responding maximally to a given speed. Moreover, some neurons tune their firing rate to the stepping frequency; this relationship becomes more distinct with increasing speed. I then demonstrate that a subset of neurons can detect the initiation and the termination of locomotion. I also show that a minority of neurons display preferred responses to either clockwise or counter-clockwise yaw direction, suggesting that some neurons receive lateralised proprioceptive information. Finally, I compare the cerebellar vermis neural responses with the activity of neurons recorded from the lateral cerebellum of mice performing the same behavioural task. I show that neurons from both cerebellar areas are able to encode locomotion-related kinematic parameters, but the differences in the way they do so is indicative of their efferent and afferent connections with the rest of the nervous system. This work is the first to characterize a population of lateral cerebellum neurons in a VR environment, providing new data for sensorimotor integration and navigation models. In the future, population recordings, further characterization of the functional connectivity of the lateral cerebellum, and targeted behavioural tasks can improve understanding of how the different areas of the cerebellum work together to achieve sensorimotor control and learning.