Machine learning can be used to distinguish protein families and generate new proteins belonging to those families.

Proteins are classified into families based on evolutionary relationships and common structure-function characteristics. Availability of large data sets of gene-derived protein sequences drives this classification. Sequence space is exponentially large, making it difficult to characterize family differences. In this work, we show that Machine Learning (ML) methods can be trained to distinguish between protein families. A number of supervised ML algorithms are explored to this end. The most accurate is a Long Short Term Memory (LSTM) classification method that accounts for the sequence context of the amino acids. Sequences for a number of protein families where there are sufficient data to be used in ML are studied. By splitting the data into training and testing sets, we find that this LSTM classifier can be trained to successfully classify the test sequences for all pairs of the families. Also investigated is whether the addition of structural information increases the accuracy of the binary comparisons. It does, but because there is much less available structural than sequence information, the quality of the training degrades. Another variety of LSTM, LSTM_wordGen, a context-dependent word generation algorithm, is used to generate new protein sequences based on seed sequences for the families considered here. Using the original sequences as training data and the generated sequences as test data, the LSTM classification method classifies the generated sequences almost as accurately as the true family members do. Thus, in principle, we have generated new members of these protein families.

Reweighting ensemble probabilities with experimental histogram data constraints using a maximum entropy principle.

Entropy maximization methods that update a probability distribution P 0(x) to a new distribution P(x) with the use of externally known, averaged constraints find use in diverse areas. Jaynes developed a Maximum Entropy Procedure (MEP) that is an objective approach to incorporate external data to update P 0(x) to P(x). In this work, we consider the MEP in the context of external data known from a probability distribution versus that from a mean and a few higher moments. An immediate problem is that the conventional iterative Lagrange multiplier method, which relies on inverting a certain covariance matrix, is not applicable here because the covariance matrix is not invertible. We introduce an indicator function method that does not suffer from this problem. It leads to an analytic solution to this version of a MEP. As an example, a previously generated ensemble of peptide conformations used to characterize an intrinsically disordered protein is analyzed. The external constraint is on the radius of gyration probability distribution, p(RG), of this peptide. Ensemble observables such as geometric, shape characteristics, the residue end-to-end distance distribution, the all atom-pair distribution function related to the scattering intensity, the polyproline II content, and NMR 3JHNHα three bond couplings are evaluated with the initial and updated ensembles. Some observables are found to be insensitive and others sensitive to the external information. An example of a 24-residue peptide, histatin 5, where an experimentally derived p(RG) is available, is also analyzed.

Generating Intrinsically Disordered Protein Conformational Ensembles from a Database of Ramachandran Space Pair Residue Probabilities Using a Markov Chain.

Intrinsically disordered proteins (IDPs), involved in regulatory pathways and cell signaling, sample a range of conformations. Constructing structural ensembles of IDPs is a difficult task for both experiment and simulation. In this work, we produce potential IDP ensembles using an existing database of pair residue φ and ψ angle probabilities chosen from turn, coil, and bend parts of sequences from the Protein Data Bank. For all residue pair types, a k-means-based discretization is used to create a set of rotamers and their probabilities in this pair Ramachandran space. For a given sequence, a Markov-based probabilistic algorithm is used to create Ramachandran space database-Markov ensembles that are converted to Cartesian coordinates of the backbone atoms. From these Cartesian coordinates and φ and ψ dihedral angles of a sequence, various observables: the radius of gyration and shape parameters, the distance probability distribution that is related to the small-angle X-ray scattering intensity, atom-atom contact percentages, local structural information, NMR three-bond J couplings, CA chemical shifts, and residual dipolar couplings are evaluated. A benchmark set of ensembles for 16 residue, regular sequences is constructed and used to validate the method and to explore the implications of the database for some of the above-mentioned observables. Then, we examine a set of nonapeptides of the form EGAAXAASS where X denotes residues of different characters. These peptides were studied by NMR, and subsequent molecular dynamics (MD) simulations were carried out using various force fields to find which one best agrees with the NMR data. Our analysis of these peptides shows that the combination of the database and the Markov algorithm yields ensembles that agree very well with the NMR and MD results for the above-listed observables. Thus, this database-Markov method is a promising method to generate IDP conformational ensembles.

Generating intrinsically disordered protein conformational ensembles from a Markov chain.

Intrinsically disordered proteins (IDPs) sample a diverse conformational space. They are important to signaling and regulatory pathways in cells. An entropy penalty must be payed when an IDP becomes ordered upon interaction with another protein or a ligand. Thus, the degree of conformational disorder of an IDP is of interest. We create a dichotomic Markov model that can explore entropic features of an IDP. The Markov condition introduces local (neighbor residues in a protein sequence) rotamer dependences that arise from van der Waals and other chemical constraints. A protein sequence of length N is characterized by its (information) entropy and mutual information, MIMC, the latter providing a measure of the dependence among the random variables describing the rotamer probabilities of the residues that comprise the sequence. For a Markov chain, the MIMC is proportional to the pair mutual information MI which depends on the singlet and pair probabilities of neighbor residue rotamer sampling. All 2N sequence states are generated, along with their probabilities, and contrasted with the probabilities under the assumption of independent residues. An efficient method to generate realizations of the chain is also provided. The chain entropy, MIMC, and state probabilities provide the ingredients to distinguish different scenarios using the terminologies: MoRF (molecular recognition feature), not-MoRF, and not-IDP. A MoRF corresponds to large entropy and large MIMC (strong dependence among the residues' rotamer sampling), a not-MoRF corresponds to large entropy but small MIMC, and not-IDP corresponds to low entropy irrespective of the MIMC. We show that MorFs are most appropriate as descriptors of IDPs. They provide a reasonable number of high-population states that reflect the dependences between neighbor residues, thus classifying them as IDPs, yet without very large entropy that might lead to a too high entropy penalty.

Conformational Ensembles Exhibit Extensive Molecular Recognition Features.

Intrinsically disordered proteins (IDPs) are important for signaling and regulatory pathways. In contrast to folded proteins, they sample a diverse conformational space. IDPs have residue ranges within a sequence that have been referred to as molecular recognition features (MoRFs). A MoRF can be viewed as contiguous residues exhibiting a conformational disorder that become ordered upon binding to another protein or ligand. In this work, we introduce a structural characterization of MoRFs based on entropy and mutual information (MI). In this view, a MoRF is a set of contiguous residues that exhibit a large entropy (from rotameric residue sampling) and large MI, the latter indicating a dependence among the residues' rotameric sampling comprising the MoRF. The methodology is first applied to a number of ubiquitin ensembles that were obtained based on nuclear magnetic resonance experiments. One is a denatured Ub ensemble that has a large entropy for various unitSizes (number of contiguous residues) but essentially zero MI, indicting no dependence among the residue rotamer sampling. Another ensemble does exhibit extensive regions along the sequence where there are MoRFs centered on nonsecondary structure regions. The MoRFs are present for unitSizes 2-10. That a substantial number of MoRFs are present in Ub strongly suggests a conformational selection mechanism for this protein. Two additional ensembles for the cyclin-dependent kinase inhibitor Sic1 and for the amyloid protein α-synuclein, which have been shown to be IDPs, are also analyzed. Both exhibit MoRF-like character.

Hinge action versus grip in translocation by RNA polymerase.

Based on molecular dynamics simulations and functional studies, a conformational mechanism is posited for forward translocation by RNA polymerase (RNAP). In a simulation of a ternary elongation complex, the clamp and downstream cleft were observed to close. Hinges within the bridge helix and trigger loop supported generation of translocation force against the RNA-DNA hybrid resulting in opening of the furthest upstream i-8 RNA-DNA bp, establishing conditions for RNAP sliding. The ß flap tip helix and the most N-terminal ß' Zn finger engage the RNA, indicating a path of RNA threading out of the exit channel. Because the ß flap tip connects to the RNAP active site through the ß subunit double-Ψ-ß-barrel and the associated sandwich barrel hybrid motif (also called the flap domain), the RNAP active site is coupled to the RNA exit channel and to the translocation of RNA-DNA. Using an exonuclease III assay to monitor translocation of RNAP elongation complexes, we show that K+ and Mg2+ and also an RNA 3'-OH or a 3'-H2 affect RNAP sliding. Because RNAP grip to template suggests a sticky translocation mechanism, and because grip is enhanced by increasing K+ and Mg2+concentration, biochemical assays are consistent with a conformational change that drives forward translocation as observed in simulations. Mutational analysis of the bridge helix indicates that 778-GARKGL-783 (Escherichia coli numbering) is a homeostatic hinge that undergoes multiple bends to compensate for complex conformational dynamics during phosphodiester bond formation and translocation.

Molecular Dynamics of Oxazole Yellow Dye in its Ground and First Excited Electronic States in Solution and when Intercalated in dsDNA.

Oxazole yellow (YOPRO), a cyanine dye consisting of benzoxazole and quinoline rings connected by a linker, is almost nonfluorescent in water, but its fluorescence is greatly enhanced after intercalation in double-stranded DNA, forming the basis of DNA concentration assays. To explore this difference, new potential energy surfaces for the two linker dihedral angles in the ground S0 and first excited S1 electronic states are developed. Umbrella sampling molecular dynamics is used to obtain the free energy of rotation around the two dihedral angles of the linker. The two-dimensional free energy surface of the S1 state, spanning the Franck-Condon transition point from the S0 electronic state minimum (dihedral 1 around 180°, dihedral 2 around 0°) to the S1 state minimum (∼90, ∼0), is obtained in water and when intercalated. In water, YOPRO's S1 free energy surface is completely downhill from the Franck-Condon point, whereas when intercalated, there is a barrier on the path. Thus, when intercalated in DNA, S1 YOPRO is more constrained than in water, supporting the hypothesis that intercalation does inhibit ring rotational motion around the linker and therefore strongly reduces the nonradiative relaxation, resulting in higher fluorescence intensity.

Conformational transition in signal transduction: metastable states and transition pathways in the activation of a signaling protein.

Signal transduction is of vital importance to the growth and adaptation of living organisms. The key to understand mechanisms of biological signal transduction is elucidation of the conformational dynamics of its signaling proteins, as the activation of a signaling protein is fundamentally a process of conformational transition from an inactive to an active state. A predominant form of signal transduction for bacterial sensing of environmental changes in the wild or inside their hosts is a variety of two-component systems, in which the conformational transition of a response regulator (RR) from an inactive to an active state initiates responses to the environmental changes. Here, RR activation has been investigated using RR468 as a model system by extensive unbiased all-atom molecular dynamics (MD) simulations in explicit solvent, starting from snapshots along a targeted MD trajectory that covers the conformational transition. Markov state modeling, transition path theory, and geometric analyses of the wealth of the MD data have provided a comprehensive description of the RR activation. It involves a network of metastable states, with one metastable state essentially the same as the inactive state and another very similar to the active state that are connected via a small set of intermediates. Five major pathways account for >75% of the fluxes of the conformational transition from the inactive to the active-like state. The thermodynamic stability of the states and the activation barriers between states are found, to identify rate-limiting steps. The conformal transition is initiated predominantly by movements of the ß3α3 loop, followed by movements of the ß4α4-loop and neighboring α4 helix region, and capped by additional movements of the ß3α3 loop. A number of transient hydrophobic and hydrogen bond interactions are revealed, and they may be important for the conformational transition.

Dihedral angle entropy measures for intrinsically disordered proteins.

Protein stability is based on a delicate balance between energetic and entropic factors. Intrinsically disordered proteins (IDPs) interacting with a folded partner protein in the act of binding can order the IDP to form the correct functional interface by decrease in the overall free energy. In this work, we evaluate the part of the entropic cost of ordering an IDP arising from their dihedral states. The IDP studied is a leucine zipper dimer that we simulate with molecular dynamics and find that it does show disorder in six phi and psi dihedral angles of the N terminal sequence of one monomer. Essential to ascertain is the degree of disorder in the IDP, and we do so by considering the entire, discretized probability distribution function of N dihedrals with M conformers per dihedral. A compositional clustering method is introduced, whereby the NS = N(M) states are formed from the Cartesian product of each dihedral's conformational space. Clustering is carried out with a version of a k-means algorithm that accounts for the circular nature of dihedral angles. For the 12 dihedrals each found to have three conformers, among the resulting 531441 states, their populations show that the first 100 (500) most populated states account for ∼65% (∼90%) of the entire population, indicating that there are strong dependencies among the dihedrals' conformations. These state populations are used to evaluate a Kullback-Leibler divergence entropy measure and obtain the dihedral configurational entropy S. At 300 K, TS ∼ 3 kcal/mol, showing that IDP entropy, while roughly half that would be expected from independently distributed dihedrals, can be a decisive contributor to the free energy of this IDP binding and ordering.

Five checkpoints maintaining the fidelity of transcription by RNA polymerases in structural and energetic details.

Transcriptional fidelity, which prevents the misincorporation of incorrect nucleoside monophosphates in RNA, is essential for life. Results from molecular dynamics (MD) simulations of eukaryotic RNA polymerase (RNAP) II and bacterial RNAP with experimental data suggest that fidelity may involve as many as five checkpoints. Using MD simulations, the effects of different active site NTPs in both open and closed trigger loop (TL) structures of RNAPs are compared. Unfavorable initial binding of mismatched substrates in the active site with an open TL is proposed to be the first fidelity checkpoint. The leaving of an incorrect substrate is much easier than a correct one energetically from the umbrella sampling simulations. Then, the closing motion of the TL, required for catalysis, is hindered by the presence of mismatched NTPs. Mismatched NTPs also lead to conformational changes in the active site, which perturb the coordination of magnesium ions and likely affect the ability to proceed with catalysis. This step appears to be the most important checkpoint for deoxy-NTP discrimination. Finally, structural perturbations in the template DNA and the nascent RNA in the presence of mismatches likely hinder nucleotide addition and provide the structural foundation for backtracking followed by removing erroneously incorporated nucleotides during proofreading.

Simulations of potentials of mean force for separating a leucine zipper dimer and the basic region of a basic region leucine zipper dimer.

Basic region leucine zipper (bZIP) transcription factors involved in DNA recognition are dimeric proteins. The monomers consist of two subdomains, a leucine zipper sequence responsible for dimerization and a highly basic DNA recognition sequence. Leucine zippers are strongly dimerized, and in a bZIP, the basic region can, in the absence of DNA, undergo extensive relative monomer-to-monomer fluctuations. In this work, LZ and bZIP potentials of mean force (PMFs), which provide free energies along reaction coordinates, are simulated with a distance replica exchange method. The method uses restraint potentials to provide sampling along a reaction coordinate and enhances configuration space exploration by exchanging information between neighboring restraint potential configurations. Restraint potentials that are constructed from sums over a number of atom distances are employed. Their use requires a modification of the Weighted Histogram Analysis Method (WHAM) procedure to combine and unbias the data from the different restraint-potential-biased window densities to provide a PMF. These methods are first used to obtain a PMF for separating a leucine zipper (GCN4-p1) of the yeast transcriptional activator GCN4. The PMF indicates a very strong binding free energy that only weakens when the monomers are separated by about 12 Å, which is about 6 Å beyond their bound, dimer equilibrium distance. PMFs are also obtained for separating the basic subdomain monomer parts of the GCN4 bZIP transcriptional factor, in the absence of DNA. In a monomer separation range spanning the open, crystal-based structure to closer configurations, the basic subdomain PMF is quite flat, implying essentially thermal sampling in this distance range. A PMF generated starting from a "collapsed" state, taken from a previous simulation ( J. Phys. Chem. B 2012 , 116 , 6071 ), where collapsed refers to the feature that the basic subdomain monomers are also effectively dimerized, shows that this state is bound in free energy, though much less so than the leucine zipper dimer.

Conformational transition of response regulator RR468 in a two-component system signal transduction process.

Signal transduction can be accomplished via a two-component system (TCS) consisting of a histidine kinase (HK) and a response regulator (RR). In this work, we simulate the response regulator RR468 from Thermotoga maritima, in which phosphorylation and dephosphorylation of a conserved aspartate residue acts as a switch via a large conformational change concentrated in three proximal loops. A detailed view of the conformational transition is obscured by the lack of stability of the intermediate states, which are difficult to detect using common structural biology techniques. Molecular dynamics (MD) trajectories of the inactive and active conformations were run, and show that the inactive (or active) trajectories do not exhibit sampling of the active (or inactive) conformations on this time scale. Targeted MD (TMD) was used to generate trajectories that span the inactive and active conformations and provide a view of how a localized event like phosphorylation can lead to conformational changes elsewhere in the protein, especially in the three proximal loops. The TMD trajectories are clustered to identify stages along the transition path. Residue interaction networks are identified that point to key residues having to rearrange in the process of transition. These are identified using both hydrogen bond analysis and residue interaction strength measurements. Potentials of mean force are generated for key residue rearrangements to ascertain their free energy barriers. We introduce methods that attempt to extrapolate from one conformation to the other and find that the most fluctuating proximal loop can transit part way from one to the other, suggesting that this conformational information is embedded in the sequence.

Transition paths of Met-enkephalin from Markov state modeling of a molecular dynamics trajectory.

Conformational states and their interconversion pathways of the zwitterionic form of the pentapeptide Met-enkephalin (MetEnk) are identified. An explicit solvent molecular dynamics (MD) trajectory is used to construct a Markov state model (MSM) based on dihedral space clustering of the trajectory, and transition path theory (TPT) is applied to identify pathways between open and closed conformers. In the MD trajectory, only four of the eight backbone dihedrals exhibit bistable behavior. Defining a conformer as the string XXXX with X = "+" or "-" denoting, respectively, positive or negative values of a given dihedral angle and obtaining the populations of these conformers shows that only four conformers are highly populated, implying a strong correlation among these dihedrals. Clustering in dihedral space to construct the MSM finds the same four bistable dihedral angles. These state populations are very similar to those found directly from the MD trajectory. TPT is used to obtain pathways, parametrized by committor values, in dihedral state space that are followed in transitioning from closed to open states. Pathway costs are estimated by introducing a kinetics-based procedure that orders pathways from least (shortest) to greater cost paths. The least costly pathways in dihedral space are found to only involve the same XXXX set of dihedral angles, and the conformers accessed in the closed to open transition pathways are identified. For these major pathways, a correlation between reaction path progress (committors) and the end-to-end distance is identified. A dihedral space principal component analysis of the MD trajectory shows that the first three modes capture most of the overall fluctuation, and pick out the same four dihedrals having essentially all the weight in those modes. A MSM based on root-mean-square backbone clustering was also carried out, with good agreement found with dihedral clustering for the static information, but with results that differ significantly for the pathway analysis.

Glucose-Promoted Localization Dynamics of Excess Electrons in Aqueous Glucose Solution Revealed by Ab Initio Molecular Dynamics Simulation.

Ab initio molecular dynamics simulations reveal that an excess electron (EE) can be more efficiently localized as a cavity-shaped state in aqueous glucose solution (AGS) than in water. Compared with that (∼1.5 ps) in water, the localization time is shortened by ∼0.7-1.2 ps in three AGSs (0.56, 1.12, and 2.87 M). Although the radii of gyration of the solvated EEs are all close to 2.6 Šin the four solutions, the solvated EE cavities in the AGSs become more compact and can localize ∼80% of an EE, which is considerably larger than that (∼40-60% and occasionally ∼80%) in water. These observations are attributed to a modification of the hydrogen-bonded network by the introduction of glucose molecules into water. The water acts as a promoter and stabilizer, by forming voids around glucose molecules and, in this fashion, favoring the localization of an EE with high efficiency. This study provides important information about EEs in physiological AGSs and suggests a new strategy to efficiently localize an EE in a stable cavity for further exploration of biological function.

Variance of a potential of mean force obtained using the weighted histogram analysis method.

A potential of mean force (PMF) that provides the free energy of a thermally driven system along some chosen reaction coordinate (RC) is a useful descriptor of systems characterized by complex, high dimensional potential energy surfaces. Umbrella sampling window simulations use potential energy restraints to provide more uniform sampling along a RC so that potential energy barriers that would otherwise make equilibrium sampling computationally difficult can be overcome. Combining the results from the different biased window trajectories can be accomplished using the Weighted Histogram Analysis Method (WHAM). Here, we provide an analysis of the variance of a PMF along the reaction coordinate. We assume that the potential restraints used for each window lead to Gaussian distributions for the window reaction coordinate densities and that the data sampling in each window is from an equilibrium ensemble sampled so that successive points are statistically independent. Also, we assume that neighbor window densities overlap, as required in WHAM, and that further-than-neighbor window density overlap is negligible. Then, an analytic expression for the variance of the PMF along the reaction coordinate at a desired level of spatial resolution can be generated. The variance separates into a sum over all windows with two kinds of contributions: One from the variance of the biased window density normalized by the total biased window density and the other from the variance of the local (for each window's coordinate range) PMF. Based on the desired spatial resolution of the PMF, the former variance can be minimized relative to that from the latter. The method is applied to a model system that has features of a complex energy landscape evocative of a protein with two conformational states separated by a free energy barrier along a collective reaction coordinate. The variance can be constructed from data that is already available from the WHAM PMF construction.

Excess dielectron in an ionic liquid as a dynamic bipolaron.

We report an ab initio molecular dynamics simulation study on the accommodation of a dielectron in a pyridinium ionic liquid in both the singlet and triplet state. In contrast to water and liquid ammonia, a dielectron does not prefer to reside in cavity-shaped structures in the ionic liquid. Instead, it prefers to be distributed over more cations, with long-lived diffuse and short-lived localized distributions, and with a triplet ground state and a low-lying, open-shell singlet excited state. The two electrons evolve nonsynchronously in both states via a diffuse-versus-localized interconversion mechanism that features a dynamic bipolaron with a modest mobility, slightly lower than a hydrated electron. This work presents the first detailed study on the structures and dynamics of a dielectron in ionic liquids.

Bending Vibration-Governed Solvation Dynamics of an Excess Electron in Liquid Acetonitrile Revealed by Ab Initio Molecular Dynamics Simulation.

We report an ab initio molecular dynamics simulation study of the solvation and dynamics of an excess electron in liquid acetonitrile (ACN). Four families of states are observed: a diffusely solvated state and three ACN core-localized states with monomer core, quasi-dimer (π*-Rydberg mode) core, and dual-core/dimer core (a coupled dual-core). These core localized states cannot be simply described as the corresponding anions because only a part of the excess electron resides in the core molecule(s). The quasi-dimer core state actually is a mixture that features cooperative excess electron capture by the π* and Rydberg orbitals of two ACNs. Well-defined dimer anion and solvated electron cavity were not observed in the 5-10 ps simulations, which may be attributed to slow dynamics of the formation of the dimer anion and difficulty of the formation of a cavity in such a fluxional medium. All of the above observed states have near-IR absorptions and thus can be regarded as the solvated electron states but with different structures, which can interpret the experimentally observed IR band. These states undergo continuous conversions via a combination of long-lasting breathing oscillation and core switching, characterized by highly cooperative oscillations of the electron cloud volume and vertical detachment energy. The quasi-dimer core and diffusely solvated states dominate the time evolution, with the monomer core and dual-core/dimer core states occurring occasionally during the breathing and core switching processes, respectively. All these oscillations and core switchings are governed by a combination of the electron-impacted bending vibration of the core ACN molecule(s) and thermal fluctuations.

The RNA polymerase bridge helix YFI motif in catalysis, fidelity and translocation.

The bridge α-helix in the ß' subunit of RNA polymerase (RNAP) borders the active site and may have roles in catalysis and translocation. In Escherichia coli RNAP, a bulky hydrophobic segment near the N-terminal end of the bridge helix is identified (ß' 772-YFI-774; the YFI motif). YFI is located at a distance from the active center and adjacent to a glycine hinge (ß' 778-GARKG-782) involved in dynamic bending of the bridge helix. Remarkably, amino acid substitutions in YFI significantly alter intrinsic termination, pausing, fidelity and translocation of RNAP. F773V RNAP largely ignores the λ tR2 terminator at 200µM NTPs and is strongly reduced in λ tR2 recognition at 1µM NTPs. F773V alters RNAP pausing and backtracking and favors misincorporation. By contrast, the adjacent Y772A substitution increases fidelity and exhibits other transcriptional defects generally opposite to those of F773V. All atom molecular dynamics simulation revealed two separate functional connections emanating from YFI explaining the distinct effects of substitutions: Y772 communicates with the active site through the link domain in the ß subunit, whereas F773 communicates through the fork domain in the ß subunit. I774 interacts with the F-loop, which also contacts the glycine hinge of the bridge helix. These results identified negative and positive circuits coupled at YFI and employed for regulation of catalysis, elongation, termination and translocation.

Solvation and evolution dynamics of an excess electron in supercritical CO2.

We present an ab initio molecular dynamics simulation of the dynamics of an excess electron solvated in supercritical CO2. The excess electron can exist in three types of states: CO2-core localized, dual-core localized, and diffuse states. All these states undergo continuous state conversions via a combination of long lasting breathing oscillations and core switching, as also characterized by highly cooperative oscillations of the excess electron volume and vertical detachment energy. All of these oscillations exhibit a strong correlation with the electron-impacted bending vibration of the core CO2, and the core-switching is controlled by thermal fluctuations.