Aqueous dAMP occupies several conformations at once
The central result is that a single free nucleotide in water is conformationally heterogeneous. Analysis of the free-energy landscape together with the vibrational spectra indicates that dAMP populates a mixture of anti/gauche+, anti/trans and syn/gauche+ conformations, and the authors state that the experimental spectrum emerges from their combined contributions.
The two coordinates used to describe this landscape are the dihedral angles gamma and chi. Gamma describes rotation about the backbone bond near the phosphate, and chi describes the orientation of the adenine base relative to the sugar. Their labels combine base orientation (anti or syn) with backbone orientation (gauche+ near +60 degrees, trans near 180 degrees).
This matters because a reader who assumes one average structure would misread the spectrum. The paper's own framing is that conformational heterogeneity must be accounted for when interpreting mid-IR spectra of nucleotides.
Active learning transfers accuracy to higher theory levels
The authors built committees of neural network potentials, which are machine-learned models that predict energies and forces for a given atomic configuration. Training began with a single dAMP molecule in a cubic box of explicit water with two Na+ counterions, and an active-learning protocol sampled the free-energy landscape at the GGA level.
The GGA-level data set was then reduced with SOAP descriptors and used for transfer learning to reach higher levels of theory, specifically SCAN-D3(BJ) and revPBE0-D3(BJ). The paper reports that this markedly reduces the number of required active-learning cycles and therefore the computational cost of force and energy evaluations at the meta-GGA and hybrid levels.
The resulting potentials describe dAMP in water over nanosecond timescales. The authors report that they reproduce forces, radial distribution functions and vibrational densities of states in close agreement with ab initio molecular dynamics using the corresponding functionals.
Different functionals shift the same basins
The free-energy surfaces were computed as a function of gamma and chi for three functionals, each from 1 ns of metadynamics with 5 walkers, for a total of 5 ns per surface. All three predict two main gamma minima, gauche+ around +60 degrees and trans around 180 degrees, with an extension of the trans basin toward gauche- near -60 degrees.
For chi, positive basins around +70 degrees represent syn, while the broad region from about -180 to -50 degrees covers anti and high-anti. The breadth of the gamma minima and the low barriers between them are attributed by the authors to the lack of strong steric constraints in a single free nucleotide.
The three surfaces are not identical. With BLYP-D3(BJ), the syn/gauche+ minimum shifts toward higher gamma values and the syn/gauche- minimum is deeper than expected. With SCAN-D3(BJ), the syn/gauche+ minimum shifts toward higher chi and nearly merges with the anti/gauche+ basin. The revPBE0-D3(BJ) committee comes closest to previous literature, recovering all expected minima with positions in agreement with earlier work, and its deeper minimum appears to be anti/gauche+, followed by anti/trans, while the syn basins stay higher in energy.
Sodium coordination reshapes the chi basins
The authors note a limitation of the two-variable picture: gamma and chi are not the only variables describing the free-energy surface of aqueous dAMP. Their simulations show that interaction of the phosphate group with Na+ ions plays a role.
Specifically, the relative stability of minima at negative chi values can vary with this coordination. A higher coordination number seems to favor the anti conformation, while high-anti is more stable at lower ion coordination numbers. The landscapes in the figures are averaged over different coordination values, and the authors state that future work will include counterion coordination as a third collective variable.
This is a caveat on the reported surfaces rather than a separate result: the mixture of conformations is real in the simulations, but the precise weighting of anti versus high-anti depends on a variable the two-dimensional surfaces average out.
Learning dipole moments connects structures to spectra
To compare with experiment, the authors trained a neural network to predict dipole moments. This allowed calculation of the full IR spectrum at meta-GGA and hybrid levels and a detailed analysis of vibrational properties, rather than only structural observables.
Comparison with experimental mid-IR absorption data shows that the models accurately capture key spectral signatures of solvated dAMP. These include phosphate and nucleobase stretching modes, NH2 scissoring modes, and stretching vibrations of the glycosidic bond between adenine and deoxyribose.
The simulations also give a molecular-level account of what determines these features, with the authors highlighting a central role for the gamma and chi dihedral angles. In other words, the same coordinates that organize the conformational landscape also organize the spectral response.
The IR comparison rests on a solvent-subtracted measurement
The experimental side is an attenuated total reflectance measurement converted to absorbance and then to an absorption coefficient using the Beer-Lambert law, with penetration depth computed from wavelength, refractive indices of sample and diamond crystal, and an incident angle of pi/4.
The contribution of the solute was isolated by subtracting the absorption coefficient of the corresponding solvent from that of the solution in the same experimental scenario. This difference spectrum is what the simulated spectra are compared against.
That detail matters for reading the agreement claim: the comparison is between computed spectra and a solvent-subtracted mid-IR absorption measurement, not between computed spectra and a raw instrument output.
Heterogeneity is the interpretive priority
The paper's practical emphasis is not that one conformation wins, but that several coexist and jointly produce the observed spectrum. The authors state that the free-energy landscape analysis and the IR spectral data together indicate a mixture of anti/gauche+, anti/trans and syn/gauche+ conformations, with the experimental spectrum emerging from their combined contributions.
The authors also argue that conventional force fields are typically tailored to torsional distributions from X-ray crystallography or parameterized for polymeric DNA and RNA forms, and so often lack the flexibility to capture the conformational variability of free nucleotides in solution. That is the gap the neural-network potentials are meant to fill.
The scope is a single dAMP molecule in explicit water with two Na+ counterions, simulated over nanosecond timescales, with landscapes averaged over counterion coordination. Within that scope, the most useful takeaway for a reader comparing methods or interpreting nucleotide spectra is this: when a mid-IR band of a solvated nucleotide is assigned, the assignment should be tested against a population of conformations, and the anti versus high-anti balance should be treated as sensitive to phosphate-sodium coordination rather than fixed.
Sources used in this answer
Conformational heterogeneity and vibrational signatures of dAMP in aqueous solution revealed by high-accuracy machine-learning potentials
Neural-network potentials trained by active learning and transfer learning show that aqueous dAMP populates a mixture of anti/gauche+, anti/trans and syn/gauche+ conformations whose combined contributions produce the experimental mid-IR spectrum.