Steven Plotkin

This thesis aims to advance the development pipeline of protein and peptide therapeutics from a biophysical perspective, and covers a spectrum of contributions, from methodologies to applications. For methodology contributions, an unbiased molecular dynamics (MD) simulation tool, Reservoir REMD (Res-REMD), has been developed and integrated into GROMACS. It has been benchmarked and shown to give the same results for different initial conformations, even when starting the simulation from a kinetically trapped initial state. Res-REMD and other enhanced MD methods were used to calculate the dimer binding free energy of SOD1 to study how disease-associated mutations affect homodimer binding. The results reveal that while the A4V mutation decreases the binding affinity, the D101N mutation does not. These findings challenge the hypothesis that dimer dissociation initiates SOD1 misfolding, a pathological event that is known to contribute to ALS disease progression. For application-related contributions, enhanced MD and free energy calculations were performed to guide the design of vaccine immunogens for neurodegenerative diseases and mutation-robust therapeutics for COVID19. For neurodegenerative diseases, flexible cyclic peptide immunogens were designed to mimic the conformations of accessible epitopes in toxic oligomers mode of proteins such as tau in Alzheimer's disease, or alpha-synuclein in Parkinson's disease. This approach, called "Glycindel scaffolding", can be extended to other protein misfolding diseases. For COVID19, we hypothesized that a conserved region on the spike S2 region could be scaffolded to become a mutation-robust vaccine. Free energy calculations predicted that the S2 region would be exposed under unglycosylated conditions and would be stable in the pre-fusion state. With this information, Rosetta was used to scaffold this S2 region, and design several protein constructs, which have been successfully expressed and shown to be functional in wet-lab experiments. A portion of ACE2, the human receptor to SARS-CoV-2, was engineered to develop mutation-robust protein binders for spike RBD. Several engineered ACE2 decoys have been successfully expressed, showcasing the power of integrating machine-learning tools into protein design.

View record

This thesis describes applications of computer simulation and bioinformatics techniques in conjunction with experiments to understand various biological systems.In Chapter 2, we used molecular dynamics simulations to uncover the structural details of an experimentally observed interaction between the ALS-associated protein superoxide dismutase 1 (SOD1), and TNF receptor-associated factor 6. Residues present in their heterodimer binding interface were predicted through unbiased and metadynamics simulations and tested in cultured cells.In Chapter 3, we used a quick computational scan to identify two de novo mutations of SOD1, A89R and K128N that were expected to be destabilizing and stabilizing respectively. Expression in cell cultures and zebrafish confirmed that A89R produces pathologies similar to the ALS-associated mutant A4V, and that K128N is WT-like. Interestingly, unlike WT-SOD1, K128N rescued the aberrant phenotype of zebrafish motor neurons when coinjected with A4V-SOD1. To explain this, we used computational alchemy to calculate heterodimerization free energies for A4V-SOD1 with WT-SOD1 and K128N-SOD1, but could not confirm a "heterodimer-rescue" mechanism.In Chapter 4, we studied the conformational landscape of the SARS-CoV-2 spike protein. The conserved residues 980-990, normally buried under the receptor binding domain, have been reported to be transiently accessible to antibodies. Through umbrella sampling simulations we found that direct exposure of the epitope through dynamic motions is not a likely mechanism for this accessibility. Further, glycans play an important role in preventing such dynamics. During its normal function, this epitope undergoes a large-scale conformational change from a bent to an extended state. To aid development of a vaccine antigen containing the conserved fragment, we studied the free energy cost of this change and found that a bent pre-fusion-like conformation is preferred.Moving on from protein studies, in Chapter 5 we were interested in the biophysics of genome organization, and its evolution in early ancestors of all animals. The basal metazoan Mnemiopsis leidyi is well-suited for such studies but has not been standardized as a model organism. We developed protocols for laboratory culture of this organism, and obtained a highly contiguous reference genome for an inbred lineage.

View record

Protein aggregation-related diseases, in particular neurodegenerative diseases, are characterized by the aberrant perturbation of the underlying protein conformational ensemble. Effectively presenting epitopes using vaccines, to raise conformationally selective antibodies, is a central problem in treating neurodegenerative diseases. Parkinson’s and Alzheimer’s disease, which pathogenesis has been attributed to aberrant aggregates of alpha-synuclein and tau protein, respectively, are the two most common neurodegenerative diseases. Designing conformationally selective immunogens in silico often requires: 1.) an effective epitope scaffolding strategy that selectively targets pathologic aggregates known as oligomers while sparing the more abundant healthy monomers, with both the oligomeric and the monomeric forms having essentially identical amino acid sequences, and 2.) efficient methods and software for quantitative comparison of large conformational ensembles, which are currently not readily available.In this thesis, we apply various computational techniques including those based on information theory and principles of physics to address these two challenges. We computationally modeled and designed cyclic peptide and beta-helix protein immunogens to best mimic toxic oligomeric conformational ensembles of alpha-synuclein and tau protein computationally predicted epitopes,respectively. In both Parkinson’s and Alzheimer’s disease, our designed immunogens are predicted to be conformationally selective for toxic oligomers. Additionally, we developed a new generalized method for efficient representation and comparison of protein conformational ensembles. The method is up to 88 times faster while utilizing 48 times fewer computing cores than the readily available Encore software on a molecular dynamics-generated ensemble dataset.The methods developed and results presented in this thesis will not onlyaccelerate the process of in silico conformation-specific immunogen designfor protein aggregation-related diseases but have potential applications toantibody drug discovery and development in pharmaceutical biotechnology.

View record

We use classical density functional theory to investigate the interactions between solvents and proteins. We examine a diverse experimental literature to establish thermodynamic properties of protein-cosolute interaction, particularly the compensation between transfer entropy and transfer enthalpy. We develop a method of analysing the uncertainties in such measurements and use the method to resolve a long-standing debate over entropy-enthalpy compensation. We develop a classical density functional theory for interactions between proteins and cosolutes. The theory developed here ignores the solvent-solvent interaction but is nonetheless quite accurate. We use this approach to reproduce transfer free energies reported elsewhere, and show that the cDFT model captures the desolvation barrier and the temperature dependence of the transfer free energy. We use experimental values that we have analyzed to define the parameter space of a model density functional theory approach. We then extend the classical density functional theory to capture protein-water interactions, thus developing a new implicit solvent model. Along the way we give a proof that the free energy of a bath of particles in a finite external potential is independent of the external potential in the isothermal-isobaric ensemble. We finally discuss the challenges remaining in implementing our implicit solvent model.

View record

This thesis is divided into three chapters. Chapter 1 is a study of statistical models of graphs, in order to explore possible realizations of emergent manifolds. Graphs with given numbers of vertices and edges are considered, governed by a Hamiltonian that favors graphs with near-constant valency and local rotational symmetry. The model is simulated numerically in the canonical ensemble. It is found that the model exhibits a first-order phase transition, and that the low energy states are almost triangulations of two dimensional manifolds. The resulting manifold shows topological "handles" and surface intersections in a higher embedding space as well as non-trivial fractal dimension. The model exhibits a phase transition temperature of zero in the bulk limit. We explore the effects of adding long-range interactions to the model, which restore a finite transition temperature in the bulk limit.In Chapter 2, aspects of Chern-Simons theory are studied. The relations between Chern-Simons theory, a model known as BF theory named after the fields that appear in the actions, and 3D gravity, are explored and generalized to the case of non-orientable spacetime manifolds. U(1) Chern-Simons theory is quantized canonically on orientable manifolds, and U(1) BF theory is similarly quantized on non-orientable manifolds. By requiring the quantum states to form a representation of the deformed holonomy group and the deformed large gauge transformation group, we find that the mapping class group of the spacetime manifold can be consistently represented, provided the prefactor k of the Chern-Simon action satisfies quantization conditions which in general are non-trivial. We also find a k 1/k duality for the representations. Motivated by open questions about interpreting the finite size results from Chapter 1, models of finite size scaling for systems with a first-order phase transition are discussed in Chapter 3. Three physics models -- the Potts model, the Go model for protein folding, and the graph model in Chapter 1 -- are simulated. Several finite size scaling models, including three functional forms to fit the energy distributions, and a capillarity model, are compared with simulations of the corresponding physics models.

View record

The Euclidean distance, D, between two points is generalized to the distance between strings or polymers. The problem is of great mathematical beauty and very rich in structure even for the simplest of cases. The necessary and sufficient conditions for finding minimal distance transformations are presented. Locally minimal solutions for one-link and two-link chains are discussed, and the large N limit of a polymer is studied. Applications of D to protein folding and structural alignment are explored, in particular for finding minimal folding pathways. Non-crossing constraints and the resulting untangling moves in folding pathways are discussed as well. It is observed that, compared to the total distance, these extra untangling moves constitute a small fraction of the total movement. The resulting extra distance from untangling movements (Dnx ) are used to distinguish different protein classes, e.g. knotted proteins from unknotted proteins. By studying the ensembles of untangling moves, dominant folding pathways are constructed for three proteins, in particular a knotted protein. Finally, applications of D, and related metrics to protein folding rate prediction are discussed. It is seen that distance metrics are good at predicting the folding rates of 3-state folders.

View record

Protein misfolding diseases represent a large burden to human health for which only symptomatictreatment is generally available. These diseases, such as Creutzfeldt-Jakob disease, amyotrophiclateral sclerosis, and the systemic amyloidoses, are characterized by conversion of globular, nativelyfoldedproteins into pathologic β-sheet rich protein aggregates deposited in affected tissues. Understandingthe thermodynamic and kinetic details of protein misfolding on a molecular level dependson accurately appraising the free energies of the folded, partially unfolded intermediate,and misfolded protein conformers. There are multiple energetic and entropic contributions to thetotal free energy, including nonpolar, electrostatic, solvation, and configurational terms. To accuratelyassess the electrostatic contribution, a method to calculate the spatially-varying dielectricconstant in a protein/water system was developed using a generalization of Kirkwood Frohlich theoryalong with brief all-atom molecular dynamics simulations. This method was combined withpreviously validated models for nonpolar solvation and configurational entropy in an algorithm tocalculate the free energy change on partial unfolding of contiguous protein subsequences. Resultswere compared with those from a minimal, topologically-based Gō model and direct calculationof free energies by steered all-atom molecular dynamics simulations. This algorithm was appliedto understand the early steps in the misfolding mechanism for β₂-microglobulin, prion protein,and superoxide dismutase 1 (SOD1). It was hypothesized that SOD1 misfolding may follow atemplate-directed mechanism like that discovered previously for prion protein, so misfolding ofSOD1 was induced in cell culture by transfection with mutant SOD1 constructs and observed tostably propagate intracellularly and intercellularly much like an infectious prion. A defined minimalassay with recombinant SOD protein demonstrated the sufficiency of mutant SOD1 aloneto trigger wtSOD1 misfolding, reminiscent of the “protein-only” hypothesis of prion spread. Finally,protein misfolding as a feature of disease may extend beyond neurodegeneration and amyloidformation to cancer, in which derangement of protein folding quality control may lead to antibodyrecognizablemisfolded protein present selectively on cancer cell surfaces. The evidence for thishypothesis and possible therapeutic targets are discussed as a future direction.

View record

Living cells are composed of a variety of biological macromolecules such as nucleic acid, metabolites, proteins and cytoskeletal filaments as well as other particles. The fraction of the cellular interior volume that is taken by these biomolecules is about 30%, leading to a highly crowded environment. Biomolecules present in an extremely dense environment inside a cell have a completely different set of kinetic and thermodynamic behavior than in a test tube. Therefore comprehending the effect of crowding conditions on biological molecules is crucial to broad research fields such as biochemical, medical and pharmaceutical sciences. Experimentally, we are able to mimic such crowded environments; which are of more physiological relevance, by adding high concentrations of synthetic macromolecules into uncrowded buffers. Theoretically, very little attention has been paid to the effects of the dense cellular cytoplasm on biological reactions. The purpose of this work is to investigate analytically the effects of crowding agents on protein folding and stability. We present a new parameter as the measure of the polymer size, which will substitute the traditional measurements of the radius of gyration of the polymer and the end to end distance of a polymeric chain. Using this quantity we derive an expression for the free energy of the polymer which can easily be generalized to provide the free energy of a protein. This mechanism enables us to study the effect of crowding on folding and stability of a protein. The stabilization effects of the crowding particles depend on the concentration and the size of the crowders and also the type of the crowding particles that are present in the system. In our calculations the type of the crowders is controlled by the energetic parameter between the protein and surrounding macromolecules.

View record