betaFeedback

WorksAgricultural and Biological Sciences

Discrete Backbone Elasticity and Curvature-Dependent Torsion Statistics in Proteins

Abstract

Coarse-grained models of protein conformational dynamics use Cα\mathrm{C}_α geometry to describe the probabilities of regular helices, ordered turns and disordered conformations. To test how much of this order arises from local chain elasticity, we connect discrete Hasimoto frame transport to neighboring bending mismatch, deriving a curvature-product torsion interaction and a six-parameter generalized von Mises density. Fitted on 194 chains from the ATLAS molecular dynamics database and tested on 1744 sequence-nonredundant proteins, this baseline reduces negative log-density loss from 1.490 to 1.160 nats, torsion-sign error from 12.26 to 2.15 percentage points, and helix-fraction error from 21.84 to 10.41 points. Ablations and circular-moment projections against reference generative models motivate extensions to anisotropic stiffness and two partition-function-weighted elastic states, further lowering test loss to 0.790 nats and helix error to 1.65 points. The fixed hierarchy also improves predictions across 72 trajectories of peptides Fs\mathrm{Fs} and (AAQAA)3(\mathrm{AAQAA})_3 and miniprotein HP35\mathrm{HP35} spanning varied starting conformations and force fields. Remaining angular errors in these trajectories are associated with atomic context at shared bends, as bend-matched peptide frames adopt opposite torsion signs with distinct hydrogen-bond patterns and a proline-containing HP35\mathrm{HP35} turn remains at negative torsion. Introducing a bilinearly coupled Cα\mathrm{C}_α–N\mathrm{N} orientation angle largely reparameterizes the existing bend-dependent harmonics, with an empirical L2L^2 Gram projection placing 96.7% to 99.6% of its added coefficient-function squared norm within the original six-function span. These results link discrete Hasimoto elasticity to statistical backbone potentials and clarify the scope and limits of local Cα\mathrm{C}_α geometry.

Introduction

Protein conformational dynamics contributes to molecular recognition, catalysis and allosteric regulation through changes in the populations and accessibility of functional states [1, 2]. Characterizing these processes benefits from a statistical description of the conformations available to proteins. Coarse-grained models reduce atomic detail to explore these conformational ensembles at a lower computational cost [3], and recent models learned from atomistic data extend this approach to conformational landscapes across protein sequences [4]. In representations based on CαC_{\alpha} atoms, effective interactions aim to capture the conformational preferences of protein backbones across regular helices, ordered turns and disordered regions.

Four consecutive CαC_{\alpha} atoms form a local unit with two adjacent bending angles and one signed torsion angle. Both helical and nonhelical units can have concentrated torsion distributions, so a departure from helical geometry does not by itself quantify local disorder. Existing models retain different sources of conformational information. The five-parameter Fisher–Bingham hidden Markov model, FB5-HMM, combines continuous CαC_{\alpha} angular distributions with correlated hidden conformational states [5]. The toroidal dynamic Bayesian network, TorusDBN, models atomic backbone dihedral angles and peptide-bond states conditional on sequence [6]. Lifson–Roig theory represents helix–coil cooperativity through nucleation and propagation weights for residue-level helical states [7].

Local angular preferences also affect the organization of longer chain segments. Multitorsional potentials in the united-residue (UNRES) force field couple successive virtual torsions and improve the modeled geometry of helices and their connecting regions [8]. In static protein structures, changes in torsion sign between adjacent backbone positions are associated with spatial variations in curvature and with nonlocal contacts [9]. These findings motivate testing a direct elastic interaction between neighboring bends as a source of local torsional order.

The Hasimoto representation combines bending amplitudes with accumulated torsional phases [10, 11], while discrete frame transport relates these variables to backbone rotations [12, 13]. Existing Hasimoto formulations describe static backbone geometry through discrete solitons and spectral boundaries [14, 15], but they provide deterministic curves rather than statistical distributions for thermal conformational ensembles. We bridge this geometric representation to molecular dynamics by treating the rotational mismatch between neighboring bending generators in so(3)\mathfrak{so}(3) as an elastic energy. In the shared normal plane, this mismatch yields an analytical Hamiltonian where the curvature product modulates torsional stiffness and the Lie commutator determines chirality. This transforms a deterministic geometric identity into a rigorous statistical mechanical framework, generating a continuous six-parameter conditional probability density for fluctuating protein trajectories. We evaluate this local elastic interaction on 1744 ATLAS protein chains after fitting on a separate set of 194 chains [16]. Published FB5-HMM and TorusDBN parameters provide reference predictions, and standard Lifson–Roig theory is evaluated using its residue-level helical states. The six-parameter model improves direction probabilities over a curvature-independent density while leaving helical peaks too broad. To address this limitation, we examine two elastic states with different bending preferences and anisotropic responses whose relative weights combine bend energies with the angular partition function integrating the torsional Boltzmann weight of each state. Fixed ATLAS parameters are then tested on 72 controlled trajectories of the 21-residue alanine-rich Fs peptide, the 15-residue peptide (AAQAA)3_3 and the 35-residue villin headpiece HP35. Four starting conformations, three random repeats and two force-field protocols test transfer across sampled backgrounds. Geometry-matched atomic comparisons and elimination of an atomic orientation around the central virtual bond relate the remaining angular errors to structural context.

Theory

For consecutive Cα_\alpha positions ri\mathbf{r}_i, define bond lengths ℓi=∣ri+1−ri∣\ell_i = \lvert\mathbf{r}_{i+1} - \mathbf{r}_i \rvert, tangents ti=(ri+1−ri)/ℓi\mathbf{t}_i = (\mathbf{r}_{i+1} - \mathbf{r}_i)/\ell_i and turns κi=arccos⁡(ti⋅ti+1)\kappa_i = \arccos(\mathbf{t}_i \cdot\mathbf{t}_{i+1}). At regular bends, 0<κi<π0 < \kappa_i < \pi, the unit binormal is bi=ti×ti+1\mathbf{b}_i = \mathbf{t}_i \times\mathbf{t}_{i+1} and the signed torsion is

τi=atan2⁡[ti+1⋅(bi×bi+1),bi⋅bi+1].(1)\tau_i = \operatorname{atan2}\left[\mathbf{t}_{i+1} \cdot(\mathbf{b}_i \times\mathbf{b}_{i+1}), \mathbf{b}_i \cdot\mathbf{b}_{i+1}\right]. \tag*{(1)}

A four-Cα_\alpha unit has normalized bends xL=κL/πx_L = \kappa_L/\pi, xR=κR/πx_R = \kappa_R/\pi and product r=xLxRr = x_Lx_R. Its two binormals lie in the plane normal to the shared tangent t\mathbf{t}, with dot and directed cross products cos⁡τ\cos\tau and sin⁡τ\sin\tau.

The normalized generators of the minimal bending rotations are aL=xLbL\mathbf{a}_L = x_L\mathbf{b}_L and aR=xRbR\mathbf{a}_R = x_R\mathbf{b}_R. We assume a quadratic mismatch and a directed pairing,

Epair=K2∣aR−aL∣2−Jst⋅(aL×aR)=K2(xL2+xR2)−r(Kcos⁡τ+Jssin⁡τ).(2)\begin{aligned} E_{\mathrm{pair}} = \frac{K}{2}\lvert\mathbf{a}_R - \mathbf{a}_L\rvert^2 - J_s\mathbf{t}\cdot(\mathbf{a}_L \times\mathbf{a}_R) \\ &= \frac{K}{2}(x_L^2 + x_R^2) - r(K\cos\tau+ J_s\sin\tau). \tag*{(2)} \end{aligned}

Viewed as elements of the rotation Lie algebra so(3)\mathfrak{so}(3) in the plane normal to t\mathbf{t}, aL⋅aR=rcos⁡τ\mathbf{a}_L \cdot\mathbf{a}_R = r\cos\tau is an invariant inner product proportional to the negative Killing form and t⋅(aL×aR)=rsin⁡τ\mathbf{t}\cdot(\mathbf{a}_L \times\mathbf{a}_R) = r\sin\tau is the projection of the Lie commutator along t\mathbf{t}. The commutator supplies the directed geometric pairing; a nonzero chiral coefficient JsJ_s introduces the parity-breaking energetic preference. Writing Kcos⁡τ+Jssin⁡τ=K~cos⁡(τ−τ0)K\cos\tau+ J_s\sin\tau= \widetilde{K}\cos(\tau-\tau_0) with K~=K2+Js2\widetilde{K} = \sqrt{K^2 + J_s^2} defines the preferred phase τ0=atan2⁡(Js,K)\tau_0 = \operatorname{atan2}(J_s,K) between neighboring bending planes (τ0=49.8∘\tau_0 = 49.8^\circ in the ATLAS fit, close to the canonical right-handed α\alpha-helical torsion). At fixed bends, the diagonal term cancels from the normalized torsion density.

Discrete frame transport gives the corresponding Hasimoto construction. For the column frame Fi=(ni,bi−1,ti)F_i = (\mathbf{n}_i,\mathbf{b}_{i-1},\mathbf{t}_i) with ni=bi−1×ti\mathbf{n}_i = \mathbf{b}_{i-1} \times\mathbf{t}_i and right-handed coordinate rotations RαR_\alpha,

Fi+1=FiR3(τi−1)R2(κi).(3)F_{i+1} = F_i R_3(\tau_{i-1})R_2(\kappa_i). \tag*{(3)}

With Pauli matrices σα\sigma_\alpha, the corresponding SU(2) lift is Ui=e−iτi−1σ3/2e−iκiσ2/2U_i = e^{-i\tau_{i-1}\sigma_3/2}e^{-i\kappa_i\sigma_2/2}. Defining the discrete Hasimoto complex field ζi=xieiϕi\zeta_i = x_i e^{i\phi_i} with xi=κi/πx_i = \kappa_i/\pi and accumulated phase difference ϕi+1−ϕi=τi\phi_{i+1} - \phi_i = \tau_i yields the adjacent-difference identity

∣ζi+1−ζi∣2=xi2+xi+12−2xixi+1cos⁡τi.(4)\lvert\zeta_{i+1} - \zeta_i\rvert^2 = x_i^2 + x_{i+1}^2 - 2x_ix_{i+1}\cos\tau_i. \tag*{(4)}

Frame transport uses the torsion about ti\mathbf{t}_i, whereas the Hasimoto phase difference pairs bends about ti+1\mathbf{t}_{i+1}. Equation (4) expresses the isotropic bend mismatch as a Hasimoto phase-difference interaction, while the constant U(1) transport link Wi=eiτ0W_i = e^{i\tau_0} combines both angular terms into 12K~∣ζi+1−eiτ0ζi∣2\frac{1}{2}\widetilde{K}\lvert\zeta_{i+1} - e^{i\tau_0}\zeta_i\rvert^2 up to the bend-only term 12(K−K~)(xi2+xi+12)\frac{1}{2}(K-\widetilde{K})(x_i^2+x_{i+1}^2), which cancels from the normalized torsion density at fixed bends.

We represent the remaining angular background by its first two circular harmonics. For torsion angle t∈[−π,π)t \in[-\pi,\pi) conditional on a scalar bending variable ss, the maximum-entropy generalized von Mises density [17] takes the form p(t∣s)=Z(s)−1eg(t,s)p(t\mid s)=Z(s)^{-1}e^{g(t,s)}, with normalization Z(s)=∫−ππeg(t,s) dtZ(s)=\int_{-\pi}^{\pi}e^{g(t,s)}\,\mathrm{d}t and angular potential

g(t,s)=(a0+a1s)cos⁡t+(b0+b1s)sin⁡tg(t,s) = (a_0+a_1s)\cos t + (b_0+b_1s)\sin t
+c0cos⁡2t+d0sin⁡2t.(5){}+c_0\cos2t+d_0\sin2t. \tag*{(5)}

The single-state elastic model sets s=rs=r, with a1=βKa_1=\beta K and b1=βJsb_1=\beta J_s under Eq. (2) for β=(kBT)−1\beta=(k_BT)^{-1}, and estimates the six coefficients from molecular dynamics trajectories. The second harmonic already allows multiple torsional peaks within one elastic state. Discrete elastic-rod descriptions distinguish centerline geometry from a material orientation around the tangent [18]. In a four-Cα_\alpha representation, atomic orientation around the central virtual bond is absorbed into the effective harmonic coefficients unless an additional orientation variable is introduced. Setting a1=b1=0a_1=b_1=0 gives a curvature-independent control, whereas replacing s=rs=r with xL+xRx_L+x_R or min⁡(xL,xR)\min(x_L,x_R) tests alternative bend-coupling choices.

Product and sum inputs are related by

xLxR=xˉ(xL+xR)−xˉ 2+(xL−xˉ)(xR−xˉ).(6)x_Lx_R = \bar{x}(x_L+x_R)-\bar{x}^{\,2}+(x_L-\bar{x})(x_R-\bar{x}). \tag*{(6)}

Near (xˉ,xˉ)(\bar{x},\bar{x}), the affine part is absorbed into the coefficients of the sum model, leaving a quadratic difference in the bend deviations that accounts for their similar predictions on folded proteins. Because torsion is defined only when both bends are nonzero, the product r=xLxRr=x_Lx_R makes the bend-dependent interaction vanish continuously as either bend straightens, whereas xL+xRx_L+x_R remains finite at that boundary—even though the background harmonics still require a defined torsion.

A unit is geometrically helical when its minimum root mean square deviation from a right-handed four-Cα_\alpha template is at most 0.5 Å after translation and proper rotation [19, 20]. The template has radius 2.3 Å, rise 1.5 Å per residue and rotation +100∘+100^\circ per residue. Keeping observed bond lengths and bends fixed and replacing torsion by tt gives the indicator Hi(t)H_i(t) and predicted occupancy

h^i=∫−ππHi(t)p(t∣si) dt.(7)\widehat{h}_i = \int_{-\pi}^{\pi} H_i(t)p(t\mid s_i)\,\mathrm{d}t. \tag*{(7)}
Geometry of neighboring backbone bends

Figure 1. Geometry of neighboring backbone bends. Panel a shows an Fs peptide conformation with residues 16–19 highlighted in purple. Panel b shows the corresponding four-Cα unit with normalized bends xLx_L and xRx_R and central tangent t\mathbf{t}. Panel c expresses the bending generators in their shared normal plane, relating the curvature product and signed torsion through aL⋅aR=xLxRcos⁡τ\mathbf{a}_L \cdot\mathbf{a}_R = x_L x_R \cos\tau.

An isolated gap has predicted probability hi,Lhi,R(1−hˉi)h_{i,L}h_{i,R}(1-\bar{h}_i) conditional on observed neighboring helix labels hi,L,hi,R∈{0,1}h_{i,L},h_{i,R} \in\{0,1\}.

Anisotropic bending and conformational states

A single bend response can allocate probability between torsion signs while giving an inaccurate width within either population. We allow different stiffnesses along two directions of the shared normal plane, replacing the scalar stiffness in Eq. (2) by a symmetric matrix. With q=(xL,xR)\mathbf{q}=(x_L,x_R) and g6g_6 denoting the exponent in Eq. (5), the eight-coefficient angular potential is

g8(t,q)=g6(t,r)+xR2(c2cos⁡2t+d2sin⁡2t).(8)g_8(t,\mathbf{q})=g_6(t,r)+x_R^2(c_2\cos2t+d_2\sin2t). \tag*{(8)}

where inverse temperature is absorbed into the positive-semidefinite normal-plane stiffness matrix

K=(a1−2d2−2d2a1+4c2)⪰0.(9)\mathbf{K}=\begin{pmatrix}a_1 & -2d_2 \\ -2d_2 & a_1+4c_2\end{pmatrix}\succeq0. \tag*{(9)}

Its eigenvalues are a1+2c2±2c22+d22a_1+2c_2\pm2\sqrt{c_2^2+d_2^2}, so a nonnegative mismatch energy requires a1+2c2≥2c22+d22a_1+2c_2\ge2\sqrt{c_2^2+d_2^2}. In coordinates aL=(xL,0)\mathbf{a}_L=(x_L,0) and aR=xR(cos⁡t,sin⁡t)\mathbf{a}_R=x_R(\cos t,\sin t), defining the torsion-independent bend energy C(q)=12a1(xL2+xR2)+c2xR2C(\mathbf{q})=\frac{1}{2}a_1(x_L^2+x_R^2)+c_2x_R^2 gives the quadratic mismatch expansion

12(aR−aL)TK(aR−aL)=C(q)−a1rcos⁡t+2d2rsin⁡t−xR2(c2cos⁡2t+d2sin⁡2t).(10)\frac{1}{2}(\mathbf{a}_R-\mathbf{a}_L)^{\mathsf{T}}\mathbf{K}(\mathbf{a}_R-\mathbf{a}_L)=C(\mathbf{q})-a_1r\cos t+2d_2r\sin t-x_R^2(c_2\cos2t+d_2\sin2t). \tag*{(10)}

Subtracting the directed pairing (b1+2d2)rsin⁡t(b_1+2d_2)r\sin t and the angular background g0(t)=a0cos⁡t+b0sin⁡t+c0cos⁡2t+d0sin⁡2tg_0(t)=a_0\cos t+b_0\sin t+c_0\cos2t+d_0\sin2t gives C(q)−g8(t,q)C(\mathbf{q})-g_8(t,\mathbf{q}). Directional stiffness therefore changes the second harmonic through the right bend as well as restricting its coefficients. This anisotropic response is retained in each component of the two-state elastic model. The two states represent backgrounds with different preferred bends and elastic responses. Their relative bend energy is U−(x)=0U_-(x)=0 and U+(x)=δ0+δ1x+δ2x2U_+(x)=\delta_0+\delta_1x+\delta_2x^2; any common bend potential cancels at fixed q\mathbf{q}. For u,v∈{−1,+1}u,v\in\{-1,+1\}, independent state priors wu=ehu/(2cosh⁡h)w_u=e^{hu}/(2\cosh h) and wv=ehv/(2cosh⁡h)w_v=e^{hv}/(2\cosh h) are set by a global field hh, giving the unnormalized bend weight for an adjacent state pair

ξuv(q)=wuwve−Cv(q)−Uu(xL)−Uv(xR).(11)\xi_{uv}(\mathbf{q})=w_uw_v e^{-C_v(\mathbf{q})-U_u(x_L)-U_v(x_R)}. \tag*{(11)}

Integrating each state potential over torsion defines the state angular partition function Zv(q)=∫−ππeg8v(t,q) dtZ_v(\mathbf{q})=\int_{-\pi}^{\pi}e^{g_{8v}(t,\mathbf{q})}\,dt, and summing over the adjacent states yields the conditional torsion density

p(t∣q)=∑uvξuv(q)eg8v(t,q)∑uvξuv(q)Zv(q).(12)p(t\mid\mathbf{q})=\frac{\sum_{uv}\xi_{uv}(\mathbf{q})e^{g_{8v}(t,\mathbf{q})}}{\sum_{uv}\xi_{uv}(\mathbf{q})Z_v(\mathbf{q})}. \tag*{(12)}

The probability of a state pair at the observed bends is proportional to ξuv(q)Zv(q)\xi_{uv}(\mathbf{q})Z_v(\mathbf{q}). A favorable bend energy and a larger angular partition function both increase its contribution. The constants Cv(q)C_v(\mathbf{q}) cancel within an isolated component but contribute to the relative weights when comparing different components. The two-state elastic model thus changes which elastic background contributes at each bend pair, including the width and location of its torsion peaks.

Results

Baseline local elasticity and torsional order

The analyzed ATLAS collection contains 1938 protein chains with three 100 ns trajectories per chain [16]. A one-to-nine fitting-to-test partition keeps sequence-related chains, common-PDB sources and all repeats together, giving 194 fitting chains and 1744 test chains. The test set contains 1317 sequence groups. Predictions are averaged equally over chains within a group and then over groups, with 95% intervals from 2000 paired group bootstrap resamples. All elastic models use the same fitting observations and are evaluated on the same test observations. Relative to the curvature-independent control, including the curvature-product interaction reduces the continuous negative log-density loss from 1.490 to 1.160 nats, the mean absolute chain error in negative-torsion fraction from 12.26 to 2.15 percentage points, and geometric helix-fraction error from 21.84 to 10.41 points. These reductions indicate that local bending carries substantial information about torsion direction and helical geometry. Replacing the product by the sum or minimum of the two bends yields smaller changes than introducing bend dependence itself (Table S1).

ModelCoefficientsNegative log densityDirection errorHelix error
Curvature-independent control41.490 (1.476, 1.504)12.26 (11.80, 12.73)21.84 (21.20, 22.50)
Single-state elastic model61.160 (1.143, 1.176)2.15 (2.06, 2.24)10.41 (10.06, 10.76)
Bend-sum control61.171 (1.155, 1.187)2.24 (2.15, 2.34)10.99 (10.62, 11.35)
Minimum-bend control61.209 (1.192, 1.226)2.77 (2.65, 2.89)11.15 (10.78, 11.52)

Table S1. Performance on 1744 ATLAS test chains. Entries are group-weighted means with 95% group-bootstrap intervals. Direction and helix errors are absolute chain-fraction errors for negative torsion and geometric helicity, respectively, in percentage points.

The remaining structural error depends on how probability is distributed within each torsion direction. For an angular region IσI_\sigma, let wσ(r)w_\sigma(r) be its probability and qσ(t∣r)q_\sigma(t \mid r) its normalized density. Splitting the circle into negative and nonnegative torsion gives the exact identity

−log⁡p(τ∣r)=−log⁡wσ(r)−log⁡qσ(τ∣r).(13)-\log p(\tau\mid r) = -\log w_\sigma(r) - \log q_\sigma(\tau\mid r). \tag*{(13)}

The first term measures allocation between directions and the second measures the angular shape within the observed direction. Circular concentration within a region is R=∣E[eiτ∣τ∈Iσ]∣R = \left|\mathbb{E}[e^{i\tau} \mid\tau\in I_\sigma]\right|. Two models can assign the same directional probability while assigning different probabilities to a narrow helical range.

Figure 2 resolves this difference across six fitting-defined curvature-product ranges. In the lowest range, the single-state elastic model predicts 86.64% negative torsion against 74.34% observed. In the intermediate range 0.13302≤r<0.182680.13302 \le r < 0.18268, these fractions are 44.73% and 51.90%. In the two highest ranges, directional errors are about 0.68 percentage points, yet observed geometric helicities of 90.64% and 94.09% are predicted as 66.36% and 69.01%. The broader predicted positive-torsion peaks account for the appreciable helix error that remains after directional allocation improves.

Observed and predicted Cα virtual-torsion densities across six ranges of the normalized curvature product

Figure 2. Observed and predicted Cα_\alpha virtual-torsion densities across six ranges of the normalized curvature product r=xLxRr = x_L x_R. The ranges are defined from the fitting data, and predictions use the single-state elastic model. Each chain is normalized within a range before averaging equally over chains within sequence groups and over contributing groups. Observations within each chain and range have equal weight. Low and intermediate ranges show differences in direction allocation. At high curvature, the predicted positive-torsion peaks are broader than the observed peaks. Densities are per radian.

Two-state elastic predictions and reference backbone models

The two-state elastic model allows the observed bend pair to select between different bend preferences and anisotropic responses through Eq. (12). Figure 3 compares it with the single-state elastic model and the published FB5-HMM and TorusDBN references. All four models use a common grid of 100 bending cells and 36 torsion intervals on the 1744 test chains, with mean common bending coverage of 99.81% and a minimum of 97.47% for an individual chain. Angular interval loss is 2.899 nats for the single-state elastic model, 2.581 for the two-state elastic model, 2.444 for FB5-HMM and 2.565 for TorusDBN. The corresponding mean absolute chain helix-fraction errors, obtained by summing probabilities within each chain before comparison, are 10.43, 1.65, 1.64 and 3.81 percentage points (Table S3).

Two-state elastic predictions and reference backbone models

Figure 3. Two-state elastic predictions and reference backbone models on 1744 ATLAS test chains. Panel a shows negative log probability in 36 torsion intervals and panel b shows absolute chain geometric helix-fraction error. All models use the same 100-cell bending support, with equal chains within sequence groups and equal groups. Intervals are 95% limits from 2000 paired sequence-group bootstrap resamples. Panels c and d show distributions at 0.13302≤r<0.182680.13302 \le r < 0.18268 and 0.23214≤r<0.246940.23214 \le r < 0.24694, respectively, using native-coordinate range assignment and common within-range support. Curves join the centers of 64 angular bins without fitted smoothing. The inset in panel d enlarges the positive-torsion peak using the same bin values. Range conditioning follows cohort averaging, so each chain contributes in proportion to its frequency within the range, unlike the chain-normalized densities in Figure 2. The retained observed mass within the two ranges is 99.98% and 99.99%. Densities are per radian.

Table S3. Four-model common-support evaluation on 1744 ATLAS chains in 1317 sequence groups. Angular loss uses 36 torsion intervals, and helix error is the absolute chain-fraction difference in percentage points. Parentheses give 95% limits from 2000 paired group bootstrap resamples. All models use the same 100 bending cells and observed weights.

ModelAngular loss (nat)Chain helix error
Single-state elastic model2.899 (2.884, 2.915)10.43 (10.07, 10.78)
Two-state elastic model2.581 (2.558, 2.604)1.65 (1.58, 1.72)
FB5-HMM, published2.444 (2.423, 2.465)1.64 (1.57, 1.72)
TorusDBN, published2.565 (2.547, 2.582)3.81 (3.68, 3.94)

Table S3.

Panels c and d of Figure 3 resolve this helical improvement across intermediate and high curvature-product ranges assigned from the native coordinates of each sample. All curves use the same observed bending weights and common support. In the high range, observed helicity is 93.74%, compared with 66.44% for the single-state elastic model, 93.10% for the two-state elastic model, 92.57% for FB5-HMM and 90.62% for TorusDBN. The two-state elastic model places substantially more probability mass in the narrow helical region while preserving a normalized angular distribution.

Direct continuous evaluation across all observed targets without a bending-support restriction is consistent with this pattern (Table S2). Extending the single-state model to two states lowers continuous density loss from 1.160 to 0.790 nats, reduces mean absolute chain helix-fraction error from 10.41 to 1.65 percentage points, and lowers the unit-level helix Brier score, the mean squared error of predicted occupancy, from 0.0708 to 0.0432. The two-state elastic model thus approaches the 1.56-point chain helix error and the 0.0366 Brier score of published FB5-HMM while retaining a larger continuous angular loss than its 0.448 nats.

ModelSizeATLAS Density lossATLAS Helix errorControlled trajectories Density lossControlled trajectories Helix error
Single-state elastic model61.16010.411.23110.73
Anisotropic elastic model81.0566.161.1567.11
Unconstrained anisotropic control81.0164.701.1535.87
Linear second-harmonic control81.0183.741.1445.11
Two-state elastic model200.7901.650.8922.87
Sequence-associated two-state model400.7801.400.8932.86
FB5-HMM, published75 states0.4481.560.6583.34

Table S2. Elastic extensions and published FB5-HMM on the complete observed targets. Sizes denote elastic coefficients or reference hidden states. Density losses are continuous per-radian scores. Helix errors are absolute chain or trajectory fraction errors in percentage points. ATLAS means give equal weight to sequence groups; controlled means give equal weight to the 24 system, protocol and start combinations, each with three repeats.

At fixed bend, each Fisher–Bingham component of FB5-HMM has the same first two log-density harmonics as the elastic density, whereas its bend response and mixture weights vary with hidden state and sequence. Frozen-reference ablations indicate that neighboring-state dependence and sequence information each contribute to this difference, as removing adjacent-state dependence and marginalizing amino-acid information raise the fitting-set FB5-HMM loss of 0.430 nats by 0.283 and 0.167 nats, respectively. Circular-moment projections further indicate that restricting the coefficients to the global six-coefficient form accounts for a larger divergence than angular structure outside the two-harmonic family. Single-state anisotropic controls and a sequence-associated two-state variant separate the effects of directional stiffness, the positive-stiffness constraint and sequence-dependent state priors (Table S2). These comparisons support distinct elastic backgrounds and directional stiffness without identifying a fitted state with a unique atomic structure.

Standard Lifson–Roig theory addresses residue-level helical populations and cooperativity using its atomic ϕ/ψ\phi/\psi state definition, which differs from the four-Cα_\alpha geometric criterion used for the angular models. Under that same definition for prediction and observation, it gives a mean helical fraction of 41.48% against 41.22% observed (Table 1). This agreement in the ensemble mean coexists with an 11.23-point mean absolute chain-fraction error and a 34.01-point residue occupancy error. The mean chain run-length total variation, half the summed absolute probability difference, is 0.275. The comparison distinguishes agreement in total helical population from agreement in its distribution among chains and along a sequence (Table S7).

ObservableEstimate
Mean helix fraction, predicted / observed41.48% / 41.22%
Absolute chain-fraction error11.23 (10.77, 11.67)
Residue occupancy error34.01 (33.74, 34.27)
Chain run-length total variation0.275 (0.270, 0.281)

Table 1. Standard Lifson–Roig predictions using identical atomic ϕ/ψ\phi/\psi residue states for model and observation. Errors are in percentage points. Intervals are 95% sequence-group bootstrap limits.

ObservableMean95% intervalUnit
Absolute chain helix-fraction error11.23(10.77, 11.67)percentage points
Residue occupancy error34.01(33.74, 34.27)percentage points
Helix-start probability error12.29(12.21, 12.38)percentage points
Helix-end probability error11.44(11.36, 11.52)percentage points
Mean chain run-length TV0.275(0.270, 0.281)probability
Binary-state log loss0.382(0.379, 0.386)nat per observed residue

Table S7. Standard residue-level Lifson–Roig predictions on 1744 test chains. Errors compare the identical atomic ϕ/ψ\phi/\psi state definition. Means and 95% intervals use equal sequence-group weights. The binary-state log loss is evaluated in residue-state probability mass.

Transfer across systems, starting conformations and simulation protocols

We test transfer from ATLAS native-state ensembles to isolated peptides and a folded miniprotein using fixed ATLAS parameters. The systems are the 21-residue alanine-rich Fs peptide, the 15-residue peptide (AAQAA)3_3 [21], and the 35-residue villin headpiece HP35 with a proline-containing turn and a packed three-helix structure [22]. Each system has four starting conformations (Table S5) and three independent repeats under each of two protocols, giving 72 trajectories of 500 ns and 36 μ\mus in total. The controlled protocols use AMBER99SB-ILDN/TIP3P and CHARMM36m with matched water, whereas ATLAS uses CHARMM36m/TIP3P.

Table S5. Four starting conformations per system. Start numbers follow the order shown. Each start has three repeats under both simulation conditions.

SystemResiduesStarting structures
FsF_{\mathrm{s}}21Archive-derived conformations with initial hydrogen-bond-based DSSP [39] helicities of 0%, 61.9%, 90.5% and 33.3%
(AAQAA)3(\mathrm{AAQAA})_{3}15Regular helix, local relaxation, central bend and extended chain
HP35 N27H35Crystal structure (PDB ID 1YRF) and deterministic backbone perturbations of 1°, 1.5° and 2°

Table S5.

Figure 4 evaluates the fixed ATLAS parameters without refitting on all internal four-Cα_\alpha units of Fs, (AAQAA)3_3 and HP35 at 0.1 ns intervals over 5<t≤5005 < t \le500 ns. Averaged equally over the 24 system, protocol and start combinations, the two-state extension reduces continuous density loss from 1.231 to 0.892 nats and mean absolute trajectory helix-fraction error from 10.73 to 2.87 percentage points, compared with 0.658 nats and 3.34 points for published FB5-HMM (Table S2). The lower FB5-HMM angular loss indicates that further conditional angular structure remains unresolved. Across the 72 individual trajectories, the two-state elastic model lowers continuous angular loss in 64 and trajectory helix-fraction error in 48, suggesting that angular-density gains transfer more consistently than gains in total local helicity.

Transfer of fixed ATLAS parameters across systems, starting conformations and simulation protocols

Figure 4. Transfer of fixed ATLAS parameters across systems, starting conformations and simulation protocols. Rows show Fs, (AAQAA)3_3 and HP35. The left column, Angular prediction, compares continuous negative log-density loss for the single-state elastic model, two-state elastic model and published FB5-HMM. The right column, Local helix fraction, compares observed geometric helix fractions with the same three predictions. Each protocol contains four starts with three repeats per start. Different symbols identify the three repeats, and vertical bars mark the mean for each model within each start and protocol group. Both columns use the same 72 trajectories, all internal four-Cα_\alpha units and every tenth saved frame at 0.1 ns intervals over 5<t≤5005 < t \le500 ns. The controlled protocols are AMBER99SB-ILDN/TIP3P and CHARMM36m with matched water.

Atomic context at matched bends and material orientation

Comparing the two force-field protocols indicates that transfer accuracy depends on the sampled conformational regime. Observed Fs helicity changes from 13.03% under AMBER to 53.67% under CHARMM, compared with two-state elastic predictions of 17.63% and 51.81%, respectively. In the coil-rich AMBER ensembles, replacing the single-state elastic model with the two-state elastic model increases mean absolute trajectory helix-fraction error from 2.49 to 4.63 percentage points for Fs and from 3.14 to 6.33 points for (AAQAA)3_3. Under CHARMM, the two-state elastic model improves helix-fraction predictions for all three systems (Figure 4). The excess predicted helicity in the two AMBER peptide ensembles points to a protocol-dependent bias in the fixed two-state elastic model, even when the aggregate angular score improves.

The controlled trajectories retain differences associated with their starting structures and simulation protocols. The comparison therefore tests conditional predictions in the sampled ensembles rather than an equilibrium folding population (Table S8 and Table S9).

ModelAngular loss (nat)Direction errorHelix errorGenerator batch variation
Curvature-independent control3.247 ± 0.02037.41 ± 0.2722.81 ± 0.80n/a
Single-state elastic model2.948 ± 0.02216.25 ± 0.2511.74 ± 0.47n/a
Bend-sum control2.946 ± 0.02116.33 ± 0.2212.30 ± 0.48n/a
Minimum-bend control2.992 ± 0.02318.62 ± 0.3012.43 ± 0.49n/a
FB5-HMM, published2.600 ± 0.0359.67 ± 0.325.28 ± 0.290.0003
FB5-HMM, adapted2.505 ± 0.0327.95 ± 0.335.25 ± 0.290.0016
TorusDBN, published2.679 ± 0.0299.67 ± 0.356.44 ± 0.280.0045
TorusDBN, adapted2.606 ± 0.0298.02 ± 0.325.28 ± 0.290.0064

Table S8. Common conditional comparison across all internal units of the 72 controlled trajectories. Entries are equal-design-cell means with standard errors from the three molecular dynamics repeats per cell. Direction and helix errors are in percentage points on shared bending support. The last column separately gives half the absolute angular-loss difference between the two generator batches, in nat; its two-batch estimate has one degree of freedom.

SystemProtocolObserved helical fractionPredicted helical fractionResidue error
(AAQAA)3(\mathrm{AAQAA})_3AMBER11.1479.3668.23
(AAQAA)3(\mathrm{AAQAA})_3CHARMM36.5679.3644.70
Fs_{\mathrm{s}}}AMBER13.4785.1871.70
Fs_{\mathrm{s}}}CHARMM50.3885.1837.42
HP35AMBER55.7541.7633.68
HP35CHARMM66.1841.7640.91

Table S9. Standard Lifson–Roig residue-state predictions in the controlled systems over 5 < tt ≤ 500 ns. Observed and predicted helical fractions are percentages. Residue occupancy errors are in percentage points. Each entry weights four starts equally and three repeats within each start. The same fixed sequence parameters are used in both protocols.

To examine the structural information remaining at similar bends, we use Fs residues 16 to 19, (AAQAA)3_3 residues 12 to 15 and HP35 residues 20 to 23. The latter two windows maximize shared two-bend coverage across the eight start and protocol combinations. Negative-torsion peptide frames are matched to their nearest nonnegative controls using both bends. At tolerance 0.002, each bend differs by at most 0.36°, yet median circular torsion separations range from 92.7° to 118.2°. Negative populations have circular concentration R=0.795R = 0.795 to 0.853 and matched controls have R=0.470R = 0.470 to 0.537. Similar bending therefore permits distinct, internally ordered torsional populations.

The matched CHARMM examples in Figure 5a and c show the corresponding atomic arrangements. Circular means of ϕ18\phi_{18} differ by 38° to 40° in Fs and ϕ14\phi_{14} by 48° to 64° in (AAQAA)3_3 across the two protocols. Backbone hydrogen-bond occupancies also differ. Fs O14–HN18 occupancy is 4.90 percentage points lower in negative frames under AMBER and 6.10 points lower under CHARMM than in their matched controls. Panels b and d resolve these contact-specific differences, indicating that opposite torsional populations at matched bends are associated with distinct backbone hydrogen-bond patterns (Table S13).

Atomic context at matched bends and the ordered HP35 turn

Figure 5. Atomic context at matched bends and the ordered HP35 turn. Panels a and c superpose actual bend-matched CHARMM frames of Fs residues 16–19 and (AAQAA)3_3 residues 12–15 rendered in ChimeraX. Purple and warm gold identify negative and nonnegative torsions, respectively; neighboring backbone atoms are gray. Panels b and d give backbone hydrogen-bond occupancy differences between negative frames and matched controls in percentage points. The difference is negative minus nonnegative, and Oi_i–HNj_j identifies the acceptor and donor. Warm orange denotes AMBER99SB-ILDN and blue-green denotes CHARMM36m. Panel e enlarges the Pro21 environment in the HP35 20–23 turn; gray dashed bonds satisfy the hydrogen-bond criterion in the displayed frame. Panel f compares the same CHARMM HP35 unit using observed density in black, the single-state elastic model in light blue, the two-state elastic model in dark blue and published FB5-HMM in brick red. Figure S3 gives the corresponding AMBER distributions.

Table S13. Matching at tolerance 0.002 and material-direction diagnostics. Controls may be reused. HP35 has 594000 negative observations per condition and no nonnegative control. RdR_d is material-angle concentration and 100ϵd2100\epsilon_d^2 is the coefficient-function squared-norm residual.

SystemConditionWindowQueriesControlsMedian Δτ\Delta\tauRdR_d100ϵd2100\epsilon_d^2
FsAMBER16–1928806276260105.0°0.9863.33
FsCHARMM16–1919829243813118.2°0.9950.73
(AAQAA)3_3AMBER12–153214627287392.7°0.9931.74
(AAQAA)3_3CHARMM12–1527803042104105.4°0.9921.28
HP35AMBER20–2300n/a0.9970.38
HP35CHARMM20–2300n/a0.9960.69

Table S13.

The HP35 window contains Pro21 and forms an ordered nonhelical turn that remains at negative torsion in all 594 000 densely saved frames per protocol, with R≃0.993R \simeq0.993 and zero geometric helicity. On the common 0.1 ns target, the single-state elastic model assigns 33.00% negative probability under AMBER and 31.16% under CHARMM.

The two-state elastic model increases these probabilities to 49.35% and 48.02%, whereas direct FB5-HMM gives 87.67% and 87.26% (Table S14). The state mixture reduces a local directional error but leaves a substantial discrepancy. Figure 5e shows the atomic environment of this turn and panel f resolves its CHARMM angular density. The proline-containing turn and surrounding contacts suggest that this unresolved directional preference reflects specific atomic constraints.

Table S14. Negative-torsion probabilities at the representative four-residue units, in percent. Predictions use all observed geometries from the matched 0.1 ns target over 5<t≤5005 < t \le500 ns, averaged equally over twelve trajectories per system and protocol. Each external entry is direct conditional FB5-HMM evaluation.

SystemProtocolObservedSingle-state elasticTwo-state elasticSequence-associatedFB5-HMM
FsAMBER56.1060.8658.8958.6647.06
FsCHARMM37.2136.8736.0035.3831.68
(AAQAA)3_3AMBER58.4964.4863.2062.1453.17
(AAQAA)3_3CHARMM49.9350.6249.8548.6943.10
HP35AMBER100.0033.0049.3549.5687.67
HP35CHARMM100.0031.1648.0248.2387.26

Table S14.

A projected Cα_\alpha–N direction provides one possible additional local variable. Its material angle dd measures the nitrogen direction in the second residue around the shared virtual bond, relative to nL=bL×t\mathbf{n}_L = \mathbf{b}_L \times\mathbf{t}. For fixed bends q=(xL,xR)\mathbf{q} = (x_L,x_R), a bilinear coupling to the right normal nR=bR×t\mathbf{n}_R = \mathbf{b}_R \times\mathbf{t} gives the coupled energy

E(q,τ,d)=E0(q,τ)+Em(q,d)−Jrcos⁡(τ−d),(14)E(q,\tau,d) = E_0(q,\tau) + E_m(q,d) - Jr \cos(\tau-d), \tag*{(14)}

where the baseline E0(q,τ)E_0(q,\tau) corresponds to the single-state elastic density p0(τ∣q)=Z0(q)−1e−βE0(q,τ)p_0(\tau\mid q) = Z_0(q)^{-1}e^{-\beta E_0(q,\tau)} in Eq. (5). When the material orientation is frozen at a common angle d∗d_*, the coupling reparameterizes the existing curvature slopes according to

(a1eff,b1eff)=(a1,b1)+βJ(cos⁡d∗,sin⁡d∗).(15)(a_1^{\mathrm{eff}},b_1^{\mathrm{eff}})=(a_1,b_1)+\beta J(\cos d_*,\sin d_*). \tag*{(15)}

In the controlled trajectories, the material-angle concentration Rd=∣⟨eid⟩∣R_d=\left|\langle e^{\mathrm{i}d}\rangle\right| ranges from 0.986 to 0.997. An empirical L2L^2 Gram projection places 96.7% to 99.6% of the added coefficient-function squared norm within the original six-function span. The projection measures redundancy on these sampled geometries, consistent with the reparameterization at a fixed material angle.

For a fluctuating direction, define w0(d∣q)∝e−βEm(q,d)w_0(d\mid q)\propto e^{-\beta E_m(q,d)}, z=βJrz=\beta Jr, mn=Ew0eindm_n=\mathbb{E}_{w_0}e^{\mathrm{i}nd} and Λ=Ew0ezcos⁡(τ−d)\Lambda=\mathbb{E}_{w_0}e^{z\cos(\tau-d)}. Then p∝p0Λp\propto p_0\Lambda, with

log⁡Λ=zRe⁡(m1eiτ)+z24{1−∣m1∣2+Re⁡[(m2∗−(m1∗)2)e2iτ]}+O(∣z∣3).(16)\begin{aligned} \log\Lambda={}&z\operatorname{Re}(m_1e^{\mathrm{i}\tau})\\ &+\frac{z^2}{4}\{1-|m_1|^2+\operatorname{Re}[(m_2^*-(m_1^*)^2)e^{2\mathrm{i}\tau}]\}+O(|z|^3). \tag*{(16)} \end{aligned}

For this coupling, a fixed material direction shifts the existing first-harmonic coefficients. Fluctuations generate a second-harmonic correction whose leading amplitude scales as r2r^2 when the material moments remain fixed across bends. The tested anisotropic mismatch contributes a term proportional to xR2x_R^2. The material reduction thus specifies a distinct possible correction. Its magnitude is not determined by the observed material-angle concentration alone because the fluctuation term depends on the uncoupled moments m1m_1 and m2m_2, which satisfy ∣m2−m12∣≤1−∣m1∣2|m_2-m_1^2|\le1-|m_1|^2 and therefore suppress the correction only as the uncoupled distribution becomes fully concentrated.

Discussion

Formulating discrete Hasimoto frame transport as an elastic Hamiltonian bridges the geometric description of static protein backbones with the statistical mechanics of fluctuating ensembles. Whereas discrete nonlinear Schrödinger models describe individual conformations as deterministic discrete solitons, evaluating the rotational mismatch between neighboring bending generators in so(3)\mathfrak{so}(3) directly yields an analytical statistical potential for conformational trajectories. While curvature–torsion order parameters and local entropy capture overall conformational diversity [23, 24], this conditional formulation resolves a further distinction between torsion-direction probability and angular concentration within that direction. On ATLAS, the single-state elastic baseline improves direction allocation while leaving helical peaks too broad, and its residual helix-fraction error indicates that improved direction probabilities can still leave structural populations inaccurately described.

Different bend preferences and anisotropic elastic responses improve prediction when their contributions are weighted by both bend energy and the angular partition function through ξuv(q)Zv(q)\xi_{uv}(q)Z_v(q). A state with lower bend energy need not dominate if the competing state has greater integrated angular weight. This dependence of effective interactions on conformational background is also central to statistical-mechanical treatments of multiple coarse-grained free-energy surfaces [25].

interactions on conformational background is also central to statistical-mechanical treatments of multiple coarse-grained free-energy surfaces [25]. In the two-state elastic model, the two states supply different elastic responses to observed bends, while each state can already support multiple torsional peaks. The close helix-fraction accuracy of the two-state elastic model and FB5-HMM, together with their different angular losses, indicates that matching one structural population can leave further angular information unresolved.

Helical population accuracy also leaves correlations between neighboring backbone units unresolved. Coarse-graining studies of peptide ensembles distinguish local structural agreement from helix–coil cooperativity [26], and multis tate UNRES potentials improve modeled helix and linker geometry through interactions between successive torsions [8]. The reference models represent these correlations at different resolutions. FB5-HMM uses sequence and neighboring hidden states to vary continuous angular responses [5]. TorusDBN predicts atomic backbone variables before conversion to CαC_\alpha geometry [6]. Lifson–Roig theory describes residue-level helix–coil cooperativity [7, 27]; in the present comparison it reproduces the mean residue-state helical population, with remaining differences in its allocation among chains, residue positions and run lengths.

Across the controlled trajectories, the two-state elastic model lowers the aggregate angular loss and mean absolute helix-fraction error relative to the single-state elastic model using fixed ATLAS parameters, whereas the helix-fraction error increases for both peptides under AMBER. Peptide force-field benchmarks report that conformational populations and folding behavior depend on the force field, sequence and starting structure [28, 29], consistent with this uneven transfer of the fixed interaction across the sampled conformational ensembles. Start-dependent populations restrict this interpretation to conditional local predictions in those ensembles, without determining equilibrium folding free energies or transition rates.

The atomic comparisons identify structural differences at similar bends. FsF_s and (AAQAA)3(\mathrm{AAQAA})_3 retain opposite torsion populations with different atomic dihedrals and hydrogen-bond occupancies. Probabilistic CαC_\alpha backmapping samples possible atomic structures from reduced coordinates [30], and recent reconstruction benchmarks locate deviations in proline geometry and cooperative secondary structure when atomic ensembles are recovered from single-bead simulations [31]. Consistent with the geometric restrictions imposed by proline on the polypeptide backbone [32], the negative-torsion HP35 turn remains underestimated after the two-state extension. Static-structure analyses also associate torsion-sign changes between adjacent positions with curvature variation and nonlocal contacts [9], motivating an examination of neighboring geometry and contact environment alongside the local bends. Separating the contributions of Pro21, hydrogen bonding and tertiary packing in HP35 would require controlled comparisons across sequence variants or contact environments at matched geometry.

Eliminating the material angle specifies how an additional atomic orientation can enter the local potential. Averaging peptide-group orientations has previously yielded multitorsional interactions through cumulant expansions [33]. In the present reduction, a common fixed direction shifts the existing first-harmonic slopes, and fluctuations introduce a second-harmonic term that scales as r2r^2 when the uncoupled moments are bend-independent. An empirical Gram projection onto the original six-function span recovers most of the squared L2L^2 norm of the added orientation-dependent coefficient functions. The fixed-angle identity supplies a possible explanation for this redundancy. The concentration measured from the coupled trajectories does not determine the material moments of the uncoupled distribution, including m1m_1, which are needed to evaluate the fluctuation correction. The predictive value of the additional variable depends on the information and couplings it contributes beyond the retained bend-dependent angular functions, connecting this example to the broader question of representability in coarse-graining [34]. More broadly, transforming a discrete geometric identity into an analytical statistical potential demonstrates how exact differential-geometric relations can be generalized to describe conformational thermodynamics in fluctuating biomolecules.

Methods

Dataset partition and molecular dynamics protocols

The ATLAS comparison uses sequence groups formed from MMseqs2 relationships at 30% identity, 80% bidirectional coverage and an E value at most 0.001, together with identical sequences and common PDB sources. All three molecular dynamics repeats of a chain remain in the same partition. The fit contains 194 chains in 139 sequence groups, and evaluation contains 1744 chains in 1317 groups. All elastic forms, including the extensions, are fitted to the same 9894 observations comprising 51 equally weighted units dispersed across positions, repeats and times after 5 ns from each fitting chain. Each test chain supplies 4000 saved evaluation observations restricted to times after 5 ns. Model forms and numerical settings were chosen in an earlier grouped exploration with 1165 fitting, 388 selection and 385 evaluation chains; the present comparison refits these fixed forms on the 194-chain subset and reports all evaluated elastic variants. Full-model adaptation of external generative references uses twelve complete atomic conformations per fitting chain, with four times from each repeat after 5 ns, assigning equal total weight to each sequence group, equal weight to chains within each group and equal weight to sampled conformations within each chain, divided by chain length.

For the controlled simulations, FsF_s contains 21 residues with sequence AAAAAAAAAAAAAAAAAAAAAAA and ACE/NME caps, (AAQAA)3(\mathrm{AAQAA})_3 contains 15 residues with ACE and amide caps, and HP35 is the 35-residue N27H variant derived from PDB ID 1YRF with free termini and neutral epsilon-protonated histidine [22]. Table S5 summarizes the four starting conformations per system. All three repeats of a fixed starting structure use separately recorded solvent, velocity and thermostat random states; the first two FsF_s repeats per start and condition retain their existing 40 ns production and native random states when available. The FsF_s starts use the conformations at 249.95 ns from archive trajectories 5, 21, 27 and 1, respectively [35]. The regular (AAQAA)3(\mathrm{AAQAA})_3 helix has (ϕ,ψ)=(−57∘,−47∘)(\phi,\psi)=(-57^\circ,-47^\circ) and trans peptide bonds; the locally relaxed start substitutes (−75∘,−25∘)(-75^\circ,-25^\circ) at residues 4, 5, 11 and 12; the central-bend start substitutes (−80∘,−20∘)(-80^\circ,-20^\circ), (−100∘,10∘)(-100^\circ,10^\circ) and (−75∘,−30∘)(-75^\circ,-30^\circ) at residues 7, 8 and 9; and the extended start uses (−135∘,135∘)(-135^\circ,135^\circ) throughout. HP35 start 1 retains the experimental heavy atoms with the missing terminal oxygen added. For starts 2 to 4, indexed by j=1,2,3j=1,2,3 with amplitudes (A1,A2,A3)=(1∘,1.5∘,2∘)(A_1,A_2,A_3)=(1^\circ,1.5^\circ,2^\circ), backbone increments at residue ii are

δθijc=Ajsin⁡(φi+0.73j+c),(17)\delta\theta_{ijc}=A_j\sin(\varphi_i+0.73j+c), \tag*{(17)}

where φ=(1+5)/2\varphi=(1+\sqrt{5})/2, c=0c=0 for ϕ\phi and c=1.1c=1.1 for ψ\psi. Undefined terminal torsions and proline ϕ\phi rotations are omitted, and hydrogens are rebuilt with heavy atoms fixed. The two controlled protocols use AMBER99SB-ILDN/TIP3P and CHARMM36m protein parameters [36] from the OpenMM files derived from the July 2024 CHARMM release, with matched CHARMM-modified TIP3P water. Both use 1.2 nm solvent padding and neutralizing ions without added salt. Preparation comprises minimization, 0.1 ns restrained constant-volume equilibration, 0.5 ns restrained constant-pressure equilibration and 1 ns unrestrained constant-pressure equilibration. Production uses OpenMM 8.3.1 at 300 K and 1 bar, a 2 fs step, Langevin-middle friction of 1 ps−1^{-1}, particle-mesh Ewald electrostatics, a 1.2 nm cutoff and a potential switch from 1.0 to 1.2 nm, saving coordinates every 10 ps. ATLAS uses CHARMM36m/TIP3P with 150 mM salt in GROMACS [16].

Regularized estimation and two-state elastic inference

The single-state elastic model and its three curvature controls minimize the mean negative log density plus a coefficient-function penalty with regularizer λ=10−4\lambda= 10^{-4}. The curvature-independent control sets a1=b1=0a_1 = b_1 = 0, while the sum and minimum controls replace s=rs = r by s=xL+xRs = x_L + x_R and s=min⁡(xL,xR)s = \min(x_L,x_R), respectively. With η=(a0+a1s,b0+b1s,c0,d0)\boldsymbol{\eta} = (a_0 + a_1s, b_0 + b_1s, c_0, d_0),

L=−1m∑ilog⁡p(τi∣si)+λ4m∑i∥η(si)∥2.(18)\mathcal{L}=-\frac{1}{m}\sum_i \log p(\tau_i\mid s_i)+\frac{\lambda}{4m}\sum_i\lVert\boldsymbol{\eta}(s_i)\rVert^2. \tag*{(18)}

For η(s)=f(s)Tβ\boldsymbol{\eta}(s)=\mathbf{f}(s)^{\mathsf{T}}\boldsymbol{\beta} and G=m−1∑if(si)f(si)TG=m^{-1}\sum_i\mathbf{f}(s_i)\mathbf{f}(s_i)^{\mathsf{T}}, training-only Gram whitening uses β=G−1/2θ\boldsymbol{\beta}=G^{-1/2}\boldsymbol{\theta}, so the penalty is λ∥θ∥2/4\lambda\lVert\boldsymbol{\theta}\rVert^2/4 on the nonsingular feature space. Circular fitting and prediction use 512 and 1024 equally spaced midpoint nodes, subtracting the exponent maximum before normalization. Table S6 records the fixed parameters, and xˉ=0.4157\bar{x}=0.4157 is the fitting reference in Eq. (6).

The anisotropic energy expansion is given in Eq. (10). Its constrained single-state fit is convex in the raw coefficients with the convex matrix constraint in Eq. (9). The free anisotropic control uses the identical eight angular coefficients without that constraint. The linear second-harmonic control instead replaces xR2x_R^2 by rr in Eq. (8). Each state of the two-state elastic model retains the constrained anisotropic response and parameterizes its stiffness by a Cholesky factor with positive diagonal.

The main two-state elastic model has 20 fitted coefficients, comprising two eight-coefficient angular responses, three relative bend-energy coefficients and one global state field. Its prior factorizes as wuwvw_u w_v. The optional sequence-associated variant has 40 coefficients and replaces that prior by adjacent-state probabilities from an open-chain Ising distribution,

P(z∣A)∝exp⁡(∑ihAizi+Jseq∑izizi+1).(19)P(z\mid A)\propto\exp\left(\sum_i h_{A_i}z_i+J_{\mathrm{seq}}\sum_i z_i z_{i+1}\right). \tag*{(19)}

where AA is the full amino-acid sequence and zi∈{−1,+1}z_i\in\{-1,+1\}. Twenty amino-acid fields and one neighbor coupling replace the global field. Forward–backward messages give Wuv(A)W_{uv}(A) for the central residue pair before its observed bends enter. In this variant,

ξuv(q,A)=Wuv(A)e−Cv(q)−Uu(xL)−Uv(xR),(20)\xi_{uv}(q,A)=W_{uv}(A)e^{-C_v(q)-U_u(x_L)-U_v(x_R)}, \tag*{(20)}

which replaces Eq. (11) in the density and normalization. Each prediction uses the full sequence and its own bend pair; other observed torsions do not enter it. The local model is obtained with Jseq=0J_{\mathrm{seq}}=0 and constant fields, giving Wuv=wuwvW_{uv}=w_u w_v.

For the two-state model, the dimensionless edge energy is

Euv(q,t)=Cv(q)+Uu(xL)+Uv(xR)−g8v(t,q).(21)E_{uv}(q,t)=C_v(q)+U_u(x_L)+U_v(x_R)-g_8v(t,q). \tag*{(21)}

Summing wuwve−Euvw_u w_v e^{-E_{uv}} over the states and integrating over tt gives Eq. (12), with predictive edge probabilities proportional to ξuvZv\xi_{uv}Z_v. Replacing wuwvw_u w_v by Wuv(A)W_{uv}(A) gives the sequence-associated variant. This reduction retains the state-dependent constants in both CvC_v and UvU_v; if g8vg_{8v} is common to the states, it factors from the sum and recovers its normalized single-state density. The open-chain Ising recursion begins with α1(v)=hA1v\alpha_1(v)=h_{A_1}v and propagates

αi+1(v)=hAi+1v+log⁡∑ueαi(u)+Jsequv,(22)\alpha_{i+1}(v)=h_{A_{i+1}}v+\log\sum_u e^{\alpha_i(u)+J_{\mathrm{seq}}uv}, \tag*{(22)}

while the backward recursion includes fields to the right of the queried edge to give WuvW_{uv} before the local bend update. Normalized log messages and their derivatives with respect to the 20 fields and JseqJ_{\mathrm{seq}} are propagated together and checked against short-chain enumeration and finite differences. In the 20-coefficient local model, the constant field hh and relative energy constant δ0\delta_0 enter through 2h−δ02h-\delta_0, with regularization selecting a representative. The fitting objective adds λ/8\lambda/8 times the mean squared four-coefficient norm summed over states (λ=10−4\lambda=10^{-4}) and λ/4\lambda/4 times the mean squared centered edge energy, the mean squared fields and Jseq2J_{\mathrm{seq}}^2 when present. Coefficients are optimized with bounded limited-memory quasi-Newton optimization (L-BFGS) from separate sign-based and identical-component perturbation starts, restricting sequence fields to [−4,4][-4,4] and JseqJ_{\mathrm{seq}} to [0,3][0,3] and retaining the lower converged training objective.

Common bending support and structural probabilities

The comparison in Figure 3a and Figure 3b uses the 100 finer cells obtained by splitting each interval of the fitting-defined 5×55\times5 bending grid at its midpoint. A cell is retained when both sampling batches of every published and fitting-set-adapted reference generator contain native-coordinate observations in that cell. The same retained observations are used for all four models. Both elastic models are evaluated at the individual observed bend pairs and integrated using 1024 angular nodes. The external distributions reuse two independent batches of 512 complete conformations per test sequence. Native generated coordinates determine their bends, torsions and geometric helix labels. The 36-interval probabilities use a symmetric pseudocount of one half per interval, equivalent to total uniform prior mass 18. All four predictions are mixed with the same observed fine-cell weights within each chain. Summing the predicted and observed helix contributions before taking their absolute difference gives the chain-fraction error. Averaging this error equally over chains within each group and then over groups gives Figure 3b. The cellwise absolute errors in Table S4 instead take the absolute difference before summing over bending cells.

Table S4. Common conditional comparison on all 1744 ATLAS test chains. Angular loss is the negative log probability of a torsion interval. TV is total variation between the 36-interval distributions. Direction and helix entries are cellwise absolute errors in percentage points. Entries give group-weighted means with 95% group-bootstrap intervals on shared bending support.

ModelAngular loss (nat)TVDirection errorHelix error
Single-state elastic model2.899 (2.885, 2.914)0.471 (0.468, 0.474)14.50 (14.32, 14.70)11.23 (10.90, 11.55)
FB5-HMM, published2.444 (2.424, 2.464)0.222 (0.220, 0.224)5.05 (4.96, 5.14)2.69 (2.63, 2.75)
TorusDBN, published2.565 (2.548, 2.582)0.294 (0.292, 0.296)6.15 (6.06, 6.24)4.94 (4.83, 5.05)

Table S4.

The two displayed ranges in Figure 3c and Figure 3d are 0.13302≤r<0.182680.13302\le r<0.18268 and 0.23214≤r<0.246940.23214\le r<0.24694, with full-precision boundaries retained in the analysis files. Each observed or generated sample is assigned using rr from its own native coordinates. For a given range and fine cell, both independent batches of each native method must contain positive sample counts. All observed and predicted curves then retain the same range-specific bending support and observed cell weights. The angular display uses 64 bins with total uniform prior mass 18 for the external estimates, and integrates internal predictions on 1024 nodes. Lines join bin centers without a fitted smoothing function. The supported joint angular mass is first aggregated with equal chain weights within sequence groups and equal group weights, and only then normalized by the cohort mass in the range. This ordering retains the contribution of each chain to that range instead of assigning equal weights to separately normalized regional curves. The retained observed masses are 99.98% and 99.99%, respectively. The 36-interval scores, chain errors and bootstrap limits are recorded in Table S3.

For all interval scores, a torsion observation contributes the negative logarithm of its interval probability. Continuous scores use density with respect to dtdt in radians.

External generative models and Lifson–Roig evaluation

FB5-HMM uses 75 hidden states, Markov transitions and five-parameter Fisher–Bingham distributions on the sphere of virtual backbone angles [5]. TorusDBN uses 55 hidden states with bivariate von Mises distributions for the backbone dihedral angles ϕ\phi and ψ\psi and a categorical peptide-bond cis/trans variable [6]. Our implementation uses the published BackboneDBN 0.24 model definitions and parameter files, retaining state counts, amino-acid emissions, angular families and chain dependencies while stabilizing hidden-path backtracking, integrating concentrated angular densities over their full support and checking atomic reconstruction across residue types. Complete-chain sampling is conditioned on the amino-acid sequence with secondary-structure variables marginalized. A secondary sensitivity analysis adapts both model families on the fitting conformations using weighted forward–backward expectation steps and numerical maximization of angular emission objectives; the FB5-HMM and TorusDBN adaptations stopped after 203 and 214 accepted updates, respectively, at the prescribed iteration budget.

The generator comparison in Table S4 converts two independent batches of 512 complete conformations per test sequence into four-Cα_\alpha geometry and divides torsion into 36 equal intervals. Splitting each interval of the fitting-defined 5×55 \times5 bending grid at its midpoint yields 100 finer cells in which generated angular probabilities are estimated with a symmetric Dirichlet pseudocount of one half per torsion interval and mixed using the observed fine-cell proportions within each protein. Elastic densities are integrated over the same angular intervals at each observed bend pair, yielding a discrete probability score on shared bending support. Internal positions are weighted equally within each protein, followed by equal chains within a sequence group and equal groups, with 95% confidence intervals from 2000 paired sequence-group resamples. Circular concentration within region IσI_\sigma is Rσ=∣⟨eit⟩τ∈Iσ∣R_\sigma= \left|\langle e^{it}\rangle_{\tau\in I_\sigma}\right|, computed by dividing the joint complex moment by the probability of that region before taking its magnitude. For a generative model, the proper-rotation RMSD helicity criterion is applied to generated coordinates before bending standardization. Elastic predictions integrate each observed bond-length-specific mask Hi(t)H_i(t) against its angular density.

Figure 4, Figure S2 and the complete-target elastic comparison use direct conditional evaluation of the published FB5-HMM on the complete observed targets (4000 saved units per ATLAS chain and every tenth frame at 0.1 ns intervals over 5<t≤5005 < t \le500 ns in the controlled trajectories), where full-sequence amino-acid messages and the two observed bends determine state weights and numerical integration of the Kent components gives the continuous torsion density without discarding bending regions. Figure S2 displays the corresponding observed and predicted probabilities in 36 equal angular intervals. Within each trajectory, internal four-Cα_\alpha windows have equal weight, with equal selected valid frames within each window. The resulting trajectory distributions are averaged equally over the twelve trajectories of each system and protocol. A position-resolved sampling comparison reuses the two 512-conformation generator batches across all 1744 ATLAS test chains (retaining 94.00% mean group-weighted support) and 72 controlled trajectories (57.99% mean support), as summarized in Tables S11 and S12.

The representative-window comparison in Figure S1a to d uses two independent batches of 65 536 complete conformations per sequence and reference model. Each trajectory uses the same retained observations and observed bending-cell weights for all angular predictions. The single-state and two-state elastic densities are evaluated at the observed bend pairs, integrated with 1024 angular nodes and projected into the same 36 torsion intervals. The resulting distributions are averaged equally over the 24 trajectories of each system. The mean shared bending coverage is 99.63%, 99.70% and 100% for Fs, (AAQAA)3_3 and HP35, respectively. Each generated batch is averaged separately over the same trajectories before its pointwise range is displayed. Table S10 reports the corresponding sampling variation.

SystemPublished FB5-HMMAdapted FB5-HMMPublished TorusDBNAdapted TorusDBN
Fs_{\mathrm{s}}}0.420.790.480.62
(AAQAA)3(\mathrm{AAQAA})_30.520.380.350.43
HP350.280.480.320.36

Table S10. Largest two-batch standard error among the 36 angular-bin probabilities at each representative window, in percentage points. Each of the two independent batches contains 65 536 complete sequence conformations. Each batch curve is averaged over the same 24 molecular dynamics trajectories before its contrast is computed, leaving one degree of freedom for the sampling estimate.

Standard Lifson–Roig theory is evaluated in residue-level variables [7, 27], assigning a helical state when ϕ∈[−90∘,−30∘]\phi\in[-90^\circ,-30^\circ] and ψ∈[−77∘,−17∘]\psi\in[-77^\circ,-17^\circ] [37] and a unit-weight coil state otherwise. A helical residue has weight wiw_i when both neighbors are helical and viv_i otherwise, with exterior coil states, so a maximal helical run of length n≥2n \ge2 has homogeneous weight v2wn−2v^2w^{n-2} and an isolated helical residue has weight vv. Exact finite-chain transfer calculations give the partition function, occupancy, run-length and boundary probabilities, with 40 amino-acid-dependent log weights fitted by maximum likelihood on the 194 fitting chains and evaluated on 48 complete conformations per ATLAS test chain (Table S7) and across the controlled trajectories (Table S9).

Conditional Kent mixtures, moment projections and ablations

For a Kent state ss with orthonormal axes γ1s,γ2s,γ3s\gamma_{1s}, \gamma_{2s}, \gamma_{3s}, concentration κs\kappa_s and ovalness βs\beta_s, write hs=κsγ1s\mathbf{h}_s = \kappa_s\gamma_{1s} and Qs=βs(γ2sγ2sT−γ3sγ3sT)Q_s = \beta_s(\gamma_{2s}\gamma_{2s}^{\mathsf{T}}-\gamma_{3s}\gamma_{3s}^{\mathsf{T}}). With internal virtual bond angle θ=π−κ\theta= \pi- \kappa, c=cos⁡θc = \cos\theta and u=sin⁡θu = \sin\theta, substituting z=(c,ucos⁡τ,usin⁡τ)\mathbf{z} = (c, u\cos\tau, u\sin\tau) into the spherical exponent hsTz+zTQsz\mathbf{h}_s^{\mathsf{T}}\mathbf{z} + \mathbf{z}^{\mathsf{T}}\mathbf{Q}_s\mathbf{z} gives the unnormalized angular log density

ℓs(τ;θ)=ascos⁡τ+bssin⁡τ+cscos⁡2τ+dssin⁡2τ,(23)\ell_s(\tau;\theta) = a_s\cos\tau+ b_s\sin\tau+ c_s\cos2\tau+ d_s\sin2\tau, \tag*{(23)}

where the first-harmonic coefficients are as=uhs2+2cuQs12a_s = uh_{s2} + 2cuQ_{s12} and bs=uhs3+2cuQs13b_s = uh_{s3} + 2cuQ_{s13}, and the second-harmonic coefficients are

cs=12u2(Qs22−Qs33),ds=u2Qs23.(24)c_s = \frac{1}{2}u^2(Q_{s22} - Q_{s33}), \qquad d_s = u^2Q_{s23}. \tag*{(24)}

The omitted constant and the sphere-area Jacobian sin⁡θ\sin\theta cancel in ps(τ∣θ)=eℓs/∫eℓs dtp_s(\tau\mid\theta) = e^{\ell_s}/\int e^{\ell_s}\,dt while remaining in the bend marginal ms(θ)m_s(\theta). Letting quv(a)q_{uv}(\mathbf{a}) denote the posterior probability of two adjacent hidden states given the amino-acid sequence a\mathbf{a}, conditional independence of emissions expresses the conditional torsion density as the state mixture p(τ∣a,θL,θR)=∑vwvpv(τ∣θR)p(\tau\mid\mathbf{a}, \theta_L, \theta_R) = \sum_v w_v p_v(\tau\mid\theta_R), with state weights

wv=mv(θR)∑uquv(a)mu(θL)∑u,kquk(a)mu(θL)mk(θR).(25)w_v = \frac{m_v(\theta_R)\sum_u q_{uv}(\mathbf{a})m_u(\theta_L)}{\sum_{u,k}q_{uk}(\mathbf{a})m_u(\theta_L)m_k(\theta_R)}. \tag*{(25)}

Both bends influence the mixture weights, and the right bend also changes the component shapes. Equal components proportional to e±acos⁡τe^{\pm a\cos\tau} give a mixture proportional to cosh⁡(acos⁡τ)\cosh(a\cos\tau), whose log density contains −a4cos⁡4τ/96+O(a6)-a^4\cos4\tau/96 + O(a^6) for small aa. A mixture therefore generally exceeds the two-harmonic log-density family even though each component belongs to it.

For a frozen FB5-HMM conditional density ff at one input, let T(t)=(cos⁡t,sin⁡t,cos⁡2t,sin⁡2t)\mathbf{T}(t) = (\cos t, \sin t, \cos2t, \sin2t). Its information projection onto the local generalized von Mises family has coefficients η∗\boldsymbol{\eta}^* satisfying ∇log⁡Z(η∗)=EfT\nabla\log Z(\boldsymbol{\eta}^*) = \mathbb{E}_f\mathbf{T}, and moment matching gives

D(f∥pη)=D(f∥pη∗)+D(pη∗∥pη).(26)D(f\mathbin{\|}p_\eta) = D(f\mathbin{\|}p_{\eta^*}) + D(p_{\eta^*}\mathbin{\|}p_\eta). \tag*{(26)}

The equality follows from integrating the affine difference of the two log densities against ff and pη∗p_{\eta^*}, which share the same four moments. A global six-coefficient projection and two nested polynomial families of 16 coefficients on (1,r,r2,r3)(1,r,r^2,r^3) and 32 coefficients adding (xL,xR,xL2,xR2)(x_L,x_R,x_L^2,x_R^2) are fitted by convex optimization without penalty on the 9894 training inputs. The four frozen-model ablations retain or marginalize amino-acid emission factors and retain or remove adjacent-state dependence by replacing quv(a)q_{uv}(\mathbf{a}) with the product of its node marginals before conditioning on either bend.

Gauge invariance and material-angle reduction

In the shared normal plane, bL⋅bR=cos⁡τ\mathbf{b}_L \cdot\mathbf{b}_R = \cos\tau and t⋅(bL×bR)=sin⁡τ\mathbf{t}\cdot(\mathbf{b}_L \times\mathbf{b}_R) = \sin\tau, yielding Eq. (2) for the chosen generator metric. Whereas the generators in Eq. (2) describe minimal bending rotations between adjacent tangents, Eq. (3) gives the full frame transport including torsion. Under a change of local normal basis by αi\alpha_i, a normal coordinate ziz_i and its transport link WiW_i transform as zi′=e−iαiziz_i' = e^{-i\alpha_i}z_i and Wi′=e−iαi+1WieiαiW_i' = e^{-i\alpha_{i+1}}W_ie^{i\alpha_i}, preserving the covariant difference

∣zi+1′−Wi′zi′∣2=∣zi+1−Wizi∣2.(27)\lvert z_{i+1}' - W_i'z_i'\rvert^2 = \lvert z_{i+1} - W_iz_i\rvert^2. \tag*{(27)}

Setting the links to unity along the open chain gives the parallel-frame field and adjacent-difference identity in Eq. (4). Setting Wi=eiτ0W_i = e^{i\tau_0} with τ0=atan2⁡(Js,K)\tau_0 = \operatorname{atan2}(J_s,K) gives the linked difference ∣zi+1−eiτ0zi∣2=xi2+xi+12−2xixi+1cos⁡(τi−τ0)\lvert z_{i+1} - e^{i\tau_0}z_i\rvert^2 = x_i^2 + x_{i+1}^2 - 2x_ix_{i+1}\cos(\tau_i-\tau_0). With K~=K2+Js2\widetilde{K} = \sqrt{K^2 + J_s^2}, the complete pair energy is

Epair=K~2∣zi+1−eiτ0zi∣2+K−K~2(xi2+xi+12).(28)E_{\mathrm{pair}} = \frac{\widetilde{K}}{2}\lvert z_{i+1} - e^{i\tau_0}z_i\rvert^2 + \frac{K-\widetilde{K}}{2}(x_i^2+x_{i+1}^2). \tag*{(28)}

The second term cancels from a normalized torsion density at fixed bends. State-dependent bend constants must remain in the mixture weights, as in Eq. (12). The SU(2) lifts UiU_i and −Ui-U_i describe the same vector rotation.

Let aN\mathbf{a}_N point from Cα_\alpha to N in the second residue of a window. For a nonzero projection with P⊥=I−ttTP_\perp= I - \mathbf{t}\mathbf{t}^{\mathsf{T}}, the unit normal projection mN=P⊥aN/∣P⊥aN∣\mathbf{m}_N = P_\perp\mathbf{a}_N/\lvert P_\perp\mathbf{a}_N\rvert and left normal nL=bL×t\mathbf{n}_L = \mathbf{b}_L \times\mathbf{t} define the material angle

d=atan2⁡(mN⋅bL,mN⋅nL).(29)d = \operatorname{atan2}(\mathbf{m}_N\cdot\mathbf{b}_L,\mathbf{m}_N\cdot\mathbf{n}_L). \tag*{(29)}

The right normal is nR=cos⁡τ nL+sin⁡τ bL\mathbf{n}_R = \cos\tau\,\mathbf{n}_L + \sin\tau\,\mathbf{b}_L, giving mN⋅nR=cos⁡(τ−d)\mathbf{m}_N\cdot\mathbf{n}_R = \cos(\tau-d). Rotating the fourth Cα_\alpha around the shared tangent at fixed right bend changes τ\tau while leaving mN\mathbf{m}_N and dd unchanged. For Eq. (14) at fixed bends qq and circular measure dτ ddd\tau\,dd, defining Zm(q)=∫−ππe−βEm(q,d) ddZ_m(q) = \int_{-\pi}^{\pi}e^{-\beta E_m(q,d)}\,dd and the uncoupled material weight w0(d∣q)=Zm(q)−1e−βEm(q,d)w_0(d\mid q) = Z_m(q)^{-1}e^{-\beta E_m(q,d)} yields the coupling factor

Λ(q,τ)=∫−ππw0(d∣q)eβJrcos⁡(τ−d) dd.(30)\Lambda(q,\tau) = \int_{-\pi}^{\pi}w_0(d\mid q)e^{\beta Jr\cos(\tau-d)}\,dd. \tag*{(30)}

Integrating out dd then gives the reduced conditional torsion density

p(τ∣q)=p0(τ∣q)Λ(q,τ)∫−ππp0(t∣q)Λ(q,t) dt.(31)p(\tau\mid q) = \frac{p_0(\tau\mid q)\Lambda(q,\tau)}{\int_{-\pi}^{\pi}p_0(t\mid q)\Lambda(q,t)\,dt}. \tag*{(31)}

The coupled marginal includes an additional factor M(q,d)=∫p0(t∣q)eβJrcos⁡(t−d) dt\mathcal{M}(q,d) = \int p_0(t\mid q)e^{\beta Jr\cos(t-d)}\,dt, so p(d∣q)∝w0(d∣q)M(q,d)p(d\mid q)\propto w_0(d\mid q)\mathcal{M}(q,d). Writing Cτ=cos⁡(τ−d)C_\tau= \cos(\tau-d), its uncoupled mean is ECτ=Re⁡(m1eiτ)\mathbb{E}C_\tau= \operatorname{Re}(m_1e^{i\tau}) and its uncoupled variance is

Var⁡Cτ=12{1−∣m1∣2+Re⁡[(m2∗−(m1∗)2)e2iτ]}.(32)\operatorname{Var} C_\tau= \frac{1}{2}\left\{1-\lvert m_1\rvert^2+\operatorname{Re}\left[(m_2^*-(m_1^*)^2)e^{2i\tau}\right]\right\}. \tag*{(32)}

whose first two cumulants give Eq. (16). For Z=eidZ=e^{id}, the circular moments satisfy the bound

∣m2−m12∣=∣E(Z−m1)2∣≤E∣Z−m1∣2=1−∣m1∣2.(33)\lvert m_2-m_1^2\rvert= \lvert\mathbb{E}(Z-m_1)^2\rvert\le\mathbb{E}\lvert Z-m_1\rvert^2 = 1-\lvert m_1\rvert^2. \tag*{(33)}

For a von Mises weight w0(d)∝eηcos⁡(d−μ)w_0(d)\propto e^{\eta\cos(d-\mu)}, the integral evaluates to

Λ(τ)=I0(η2+z2+2ηzcos⁡(τ−μ))I0(η).(34)\Lambda(\tau) = \frac{I_0\left(\sqrt{\eta^2+z^2+2\eta z\cos(\tau-\mu)}\right)}{I_0(\eta)}. \tag*{(34)}

where I0I_0 is the modified Bessel function. For the effective angular free energy F=−β−1log⁡∫e−βE ddF=-\beta^{-1}\log\int e^{-\beta E}\,dd and conditional material distribution Pτ∝e−βEP_\tau\propto e^{-\beta E}, differentiation at fixed qq gives F′=⟨E′⟩PτF'=\langle E'\rangle_{P_\tau} and

F′′=⟨E′′⟩Pτ−βVar⁡Pτ(E′),(35)F''=\langle E''\rangle_{P_\tau}-\beta\operatorname{Var}_{P_\tau}(E'), \tag*{(35)}

where the fluctuation term reduces curvature relative to the conditional mean direct curvature.

Atomic matching and material-subspace projection

For candidate four-residue window jj, each of eight start-condition sources pools three repeats and supplies a two-bend histogram Nj,auvN_{j,a}^{uv} with bin width 0.005. The selection score is

Sj=∑uv (xLxR<0.23214)min⁡a=1,…,8Nj,auvNj,a,(36)S_j=\sum_{uv\ (x_Lx_R<0.23214)}\min_{a=1,\ldots,8}\frac{N_{j,a}^{uv}}{N_{j,a}}, \tag*{(36)}

where Nj,aN_{j,a} counts valid frames for window jj in source aa, and the largest score selects the window with the lowest residue index breaking ties. Matching pairs each negative query with its nearest nonnegative control in the maximum norm of (xL,xR)(x_L,x_R) at tolerances 0.005 and 0.002 with control reuse permitted. Backbone angles are obtained from saved coordinates [38], a backbone hydrogen bond is occupied when the H–O distance is below 0.25 nm and the N–H–O angle exceeds 120∘120^\circ, and internal contacts use a 0.45 nm heavy-atom distance threshold between residues separated by at least six positions. Projecting the added material coefficient functions (rcos⁡d,rsin⁡d,0,0)(r\cos d,r\sin d,0,0) onto the six-function span separates two L2L^2 projections onto {1,r}\{1,r\}. With the two-by-two Gram matrix G2G_2 having entries 11, ⟨r⟩\langle r\rangle and ⟨r2⟩\langle r^2\rangle and moment vectors uc=(⟨rcos⁡d⟩,⟨r2cos⁡d⟩)T\mathbf{u}_c=(\langle r\cos d\rangle,\langle r^2\cos d\rangle)^{\mathsf{T}} and us=(⟨rsin⁡d⟩,⟨r2sin⁡d⟩)T\mathbf{u}_s=(\langle r\sin d\rangle,\langle r^2\sin d\rangle)^{\mathsf{T}}, the relative squared-norm residual is

ϵd2=1−1⟨r2⟩(ucTG2−1uc+usTG2−1us),(37)\epsilon_d^2=1-\frac{1}{\langle r^2\rangle}\left(\mathbf{u}_c^{\mathsf{T}}G_2^{-1}\mathbf{u}_c+\mathbf{u}_s^{\mathsf{T}}G_2^{-1}\mathbf{u}_s\right), \tag*{(37)}

where moments are normalized within each trajectory and averaged equally across the twelve trajectories in a group (Table S13 and Table S14). The projection uses the measured coupled trajectory distribution, and 100(1−ϵd2)100(1-\epsilon_d^2) measures the fraction of coefficient-function squared L2L^2 norm within the existing span. It is an empirical projection and does not estimate the uncoupled moments of w0w_0. The dense matching analysis retains all 0.01 ns frames. Table S14 instead uses the 0.1 ns common target and assigns negative probability by summing the 18 negative half-circle intervals. The dense HP35 single-state elastic probabilities are 32.96% under AMBER and 31.20% under CHARMM, while the common-target values are 33.00% and 31.16%.

Supporting Information

The Supporting Information retains the complete curvature controls, elastic variants, four-model common evaluation, reference ablations and moment projections, together with the starting structures, fixed coefficients, residue-state benchmarks, generator sensitivity and bend-matched atomic diagnostics. Figure S1, Figure S2 and Figure S3 present helical-run and representative-window predictions, full-system angular distributions and the AMBER HP35 turn, respectively.

Panel a: Helical run lengths
Panel b: Fs residues 16–19

(a)

Panel c: (AAQAA)3 residues 12–15

(b)

Panel d: HP35 residues 20–23

(c)

Figure S1. Helical-run and representative-window predictions. Panel a compares observed and standard Lifson–Roig run-length probabilities on 1744 ATLAS test chains in 1317 sequence groups, using the atomic ϕ/ψ\phi/\psi definition of residue helicity in Methods. Mean run counts are pooled with equal sequence-group weights before normalization. Lengths 1 to 40 are shown without renormalizing this range. Panels b to d show the Cα\alpha virtual-torsion distributions for Fs residues 16–19, (AAQAA)3_3 residues 12–15 and HP35 residues 20–23, respectively. The observed, single-state elastic, two-state elastic, FB5-HMM and TorusDBN curves use the same shared bending support and observed bending-cell weights within each trajectory, followed by equal averaging over the 24 trajectories of each system. Mean retained observations are 99.63%, 99.70% and 100%, respectively. Lines join the centers of 36 equal angular intervals, with density per radian. Shading for FB5-HMM and TorusDBN spans the predictions from two independent batches of 65 536 complete sequence conformations after each batch is averaged over the same trajectories. Table S10 reports the sampling variation.

Complete angular distributions across the controlled systems and force-field protocols

Figure S2. Complete angular distributions across the controlled systems and force-field protocols. Rows show FsF_s, (AAQAA)3_3 and HP35, and columns show AMBER99SB-ILDN and CHARMM36m. Each panel compares observations with the single-state elastic, two-state elastic and FB5-HMM predictions. The elastic parameters are fixed from ATLAS, and FB5-HMM uses its published parameters. The curves include all internal four-CαC\alpha windows and selected frames at 0.1 ns intervals over 5<t≤5005 < t \le500 ns. Each window has equal weight within a trajectory, with equal valid selected frames within each window, followed by equal averaging over twelve trajectories per system and protocol. Lines join the centers of the same 36 angular intervals for all curves, with density per radian.

Observed and predicted torsion distributions for HP35 residues 20–23 in AMBER99SB-ILDN, showing density versus Cα virtual torsion $\tau$ and integrated negative-torsion probabilities for four models.

Figure S3. Observed and predicted torsion distributions for HP35 residues 20–23 under AMBER99SB-ILDN. The elastic parameters are fixed from ATLAS, FB5-HMM uses its published parameters, and model colors match Figure 5f. Probabilities are averaged equally over four starts and three repeats, using all observed geometries at 0.1 ns intervals over 5<t≤5005 < t \le500 ns. The table gives the integrated negative-torsion probabilities. Curves join the centers of 36 angular intervals, with density per radian.

Data and Software Availability

ATLAS trajectories and the earlier Fs archive are publicly available [16, 35]. The 72 molecular dynamics trajectories generated in this study are available on Zenodo at https://doi.org/10.5281/zenodo.23895045. The simulation and analysis code, initial structures, and fitted model parameters are available on GitHub at https://github.com/wyqmatch/protein-backbone-elasticity.

Author Contributions

Yiquan Wang contributed to conceptualization, methodology, software, investigation, formal analysis, data curation, visualization, writing of the original draft, and manuscript review and editing. Mingyu Bian contributed to methodology, software, investigation, formal analysis, and visualization. Jiayuan Ma contributed to methodology, software, investigation, and formal analysis.

Supporting Information

Curvature controls and elastic variants

The curvature controls in Table S1 use the same fitting observations and matched model sizes where applicable. The original continuous baseline loss decreases from 1.490 to 1.160 nats, with a paired reduction of 0.330 nats and 95% interval [0.327, 0.334]. The curvature-product model has a smaller advantage over the sum control, with paired loss difference −0.011 nats and interval [−0.012, −0.011]. (6) explains their shared first-order bend response.

The constrained anisotropic extension lowers continuous ATLAS loss from 1.160 to 1.056 nats and absolute chain helix-fraction error from 10.41 to 6.16 percentage points. The equal-size unconstrained and linear second-harmonic controls give lower losses, preserving a predictive cost of the positive-stiffness restriction. The two-state elastic model gives 0.790 nats and 1.65 points. Its sequence-associated variant gives 0.780 nats and 1.40 points, with helix Brier score 0.0404 against 0.0432 for the local model. Sign-fraction errors remain near two points. All variants are fitted to torsions, with geometric helix labels used only for evaluation. On the controlled trajectories, the sequence-associated variant gives loss 0.893 nats, helix-fraction error 2.86 points, and helix Brier score 0.0508, compared with 0.892, 2.87 and 0.0545 for the local model. Direct FB5-HMM gives helix Brier score 0.0484. On the controlled trajectories, the sequence-associated variant slightly increases angular loss while reducing helix-fraction error and the helix Brier score.

Common angular comparison and alternative error aggregation

Table S3 records the four-model comparison used in Figure 3. The chain-fraction error sums supported probabilities within each chain before taking the absolute difference. Table S4 retains the earlier cellwise absolute direction and helix errors, total variation and bootstrap intervals. Its absolute errors are averaged over bending cells before cohort aggregation and cannot be read as chain-fraction errors.

Applying Eq. (13) to the 36-state interval probabilities gives direction loss 0.436 nats and within-direction interval loss 2.463 nats for the single-state elastic model. Published FB5-HMM reduces these components by 0.088 and 0.367 nats, and published TorusDBN by 0.078 and 0.256 nats. Most of the reference gain in this decomposition comes from within-direction shape, with an additional improvement in directional allocation. Finite-budget adaptation improves FB5-HMM interval loss to 2.416 nats and its cellwise geometric helix error to 1.91 percentage points. TorusDBN adaptation reduces interval loss to 2.548 nats and improves direction allocation while increasing cellwise geometric helix error to 5.09 points, indicating that angular-loss and structural-error changes need not coincide. The earlier controlled generator comparison has interval losses 2.600 and 2.679 nats for published FB5-HMM and TorusDBN, against 2.948 for the single-state elastic model, with mean shared bending coverage 93.85% for FsF_{\mathrm{s}}, 91.21% for (AAQAA)3(\mathrm{AAQAA})_{3} and 99.87% for HP35 (Table S8).

Reference ablations and circular-moment projections

The continuous reference diagnostics use only the 9894 fitting inputs and leave the published FB5-HMM parameters fixed. Its full conditional loss is 0.430 nats. Marginalizing amino-acid information adds 0.167 nats, removing adjacent-state dependence adds 0.283 nats, and removing both gives 0.922 nats. The nonadditive interaction is 0.043 nats, and Eqs. (23) and (25) describe how sequence and adjacent-state weights enter the reference prediction.

Matching the first two circular harmonics of the reference independently at each input leaves mean divergence 0.139 nats. Restricting the coefficients to the global six-coefficient curvature-product form adds 0.609 nats, even when optimized against the reference moments. Replacing the original coefficients by this optimum changes divergence by only 0.007 nats. Within nested coefficient spaces, a nonlinear product response removes 0.203 nats and retaining the separate bends removes another 0.117 nats. The restriction therefore involves bend response and contextual state mixing as well as angular structure outside the two-harmonic family. Equation (26) gives the exact information-projection decomposition used in this comparison.

Simulation inputs and supplementary benchmarks

Modela0a_0b0b_0c0c_0d0d_0a1a_1b1b_1
Curvature-independent control−0.1205-0.12050.44680.4468−0.1021-0.10211.35871.35870000
Single-state elastic model−2.4096-2.4096−2.2665-2.2665−0.0494-0.04941.04111.041113.896713.896716.463316.4633
Bend-sum control−4.7553-4.7553−4.8008-4.8008−0.0626-0.06261.05391.05395.71195.71196.46876.4687
Minimum-bend control−3.1141-3.1141−2.9137-2.9137−0.0726-0.07261.10321.10328.14338.14339.14129.1412

Table S6. Coefficients of the fixed ATLAS densities in Eq. (5). Columns give a0a_0, b0b_0, c0c_0, d0d_0, a1a_1, b1b_1; values are rounded to four decimal places, and full precision is retained in the parameter files.

Table S12. Sampling sensitivity on the fixed common position-by-bend support. Losses are 36-interval scores from each separate batch of 512 conformations, with the same one-half pseudocount per bin applied separately to each batch. The final column averages the absolute half-difference of the two losses over members using the cohort weights. It measures sampling sensitivity and is separate from molecular dynamics repeat error. Pooled estimates in Table S11 combine counts before regularization.

CohortModelBatch 1 lossBatch 2 lossMean absolute half-difference
ATLASFB5-HMM2.69812.69800.0071
ATLASTorusDBN2.83982.83870.0071
FsFB5-HMM2.53112.53950.0053
FsTorusDBN2.76832.74680.0120
(AAQAA)3_3FB5-HMM2.75442.75050.0055
(AAQAA)3_3TorusDBN2.92662.88620.0202
HP35FB5-HMM2.31382.29210.0109
HP35TorusDBN2.49932.52330.0120
Controlled, pooledFB5-HMM2.53312.52740.0072
Controlled, pooledTorusDBN2.73142.71880.0147

Table S12.

ModelATLAS LossATLAS BrierATLAS Helix errorControlled trajectories LossControlled trajectories BrierControlled trajectories Helix error
Single-state elastic model2.8770.071211.202.7630.122116.28
Linear second-harmonic control2.7310.04604.352.7000.09979.37
Unconstrained anisotropic control2.7300.04815.352.6890.101010.14
Anisotropic elastic model2.7720.05256.852.7050.103711.49
Two-state elastic model2.5300.04071.612.4470.10337.47
Sequence-associated two-state model2.5190.03791.362.4790.10177.78
FB5-HMM, sampled2.5450.072912.392.4440.109113.40
TorusDBN, sampled2.6940.089814.872.5960.131417.73

Table S11. Position-resolved sampling comparison over the complete test cohorts. All models use the same positions and observed bond lengths on common generated bending support. Losses score the 36 torsion intervals. Brier scores refer to the local helix label, and helix-fraction errors are in percentage points. External estimates reuse two batches of 512 complete conformations per sequence and model.

Atomic diagnostics and turn probabilities

The dense CHARMM Fs ensemble has 53.66% observed local helicity and 37.63% under the single-state elastic model despite close direction allocation. At Fs residues 16–19, dense negative-torsion fractions are 37.18% observed and 36.86% predicted, while geometric helicities are 44.84% and 32.90%. Native-coordinate sampling for HP35 residues 20–23 assigns 86.26% to 86.77% negative torsion with published FB5-HMM and 87.52% to 88.74% with published TorusDBN, using all observed bending mass. These sampled estimates and angular peaks appear in Figure S1. The reference curves in that figure estimate conditional distributions from sampled complete conformations, whereas Table S14 evaluates FB5-HMM directly at every observed bend pair using full-sequence amino-acid messages.

References

  1. [1]Katherine Henzler-Wildman and Dorothee Kern. Dynamic personalities of proteins. Nature, 450(7172):964–972, 2007.
  2. [2]Stephanie A Wankowicz and James S Fraser. Advances in uncovering the mechanisms of macromolecular conformational entropy. Nature Chemical Biology, 21(5):623–634, 2025.DOI
  3. [3]Sebastian Kmiecik, Dominik Gront, Michal Koliński, Lukasz Wieteska, Aleksandra Elzbieta Dawid, and Andrzej Kolinski. Coarse-grained protein models and their applications. Chemical Reviews, 116(14):7898–7936, 2016.DOI
  4. [4]Nicholas E Charron, Klara Bonneau, Aldo S Pasos-Trejo, Andrea Guljas, Yaoyi Chen, Félix Musil, Jacopo Venturin, Daria Gusev, Iryna Zaporozhets, Andreas Krämer, et al. Navigating protein landscapes with a machine-learned transferable coarse-grained model. Nature Chemistry, 17(8):1284–1292, 2025.
  5. [5]Thomas Hamelryck, John T Kent, and Anders Krogh. Sampling realistic protein conformations using local structural bias. PLoS Computational Biology, 2(9):e131, 2006.
  6. [6]Wouter Boomsma, Kanti V Mardia, Charles C Taylor, Jesper Færkinghoff-Borg, Anders Krogh, and Thomas Hamelryck. A generative, probabilistic model of local protein structure. Proceedings of the National Academy of Sciences, 105(26):8932–8937, 2008.DOI
  7. [7]Shneior Lifson and A Roig. On the theory of helix–coil transition in polypeptides. The Journal of Chemical Physics, 34(6):1963–1974, 1961.DOI
  8. [8]Elizaveta F Petrushevich and Adam Liwo. Recapturing cooperativity of α-helix formation and packing in coarse-grained protein structure modeling with multisortional potentials. The Journal of Physical Chemistry B, 129(28):7119, 2025.
  9. [9]Jianshi Wang and Yukio Ohsawa. Instability-driven torsion sign flips revealed by the Laplacian structure of curvature support fields in protein backbones. The Journal of Physical Chemistry B, 130(37):9406–9417, 2026.DOI
  10. [10]Hidenori Hasimoto. A soliton on a vortex filament. Journal of Fluid Mechanics, 51(3):477–485, 1972.DOI
  11. [11]Nora Molkenthin, Shuangwei Hu, and Antti J Niemi. Discrete nonlinear Schrödinger equation and polygonal solitons with applications to collapsed proteins. Physical Review Letters, 106(7):078102, 2011.pubmed.ncbi.nlm.nih.gov/21405545
  12. [12]Shuangwei Hu, Martin Lundgren, and Antti J Niemi. Discrete Frenet frame, inflection point solitons, and curve visualization with applications to folded proteins. Physical Review E, 83(6):061908, 2011.pubmed.ncbi.nlm.nih.gov/21797404
  13. [13]Ulf H Danielsson, Martin Lundgren, and Antti J Niemi. Gauge field theory of chirally folded homopolymers with applications to folded proteins. Physical Review E, 82(2):021910, 2010.
  14. [14]Yiquan Wang. Descriptive power and predictive limits of a discrete Hasimoto–DNLS model of protein backbone structure. Physica D: Nonlinear Phenomena, page 135362, 2026.DOI
  15. [15]Yiquan Wang. Spectral analysis of protein backbone geometry reveals abrupt helix–coil boundaries. The European Physical Journal E, 49(8):70, 2026.DOI
  16. [16]Yann Vander Meersche, Gabriel Cretin, Aria Gehaert, Jean-Christophe Gelly, and Tatiana Golochkina. ATLAS: protein flexibility description from atomistic molecular dynamics simulations. Nucleic Acids Research, 52(D1):D384–D392, 2024.
  17. [17]Riccardo Gatto and Sreenivasa Rao Jammalamadaka. The generalized von Mises distribution. Statistical Methodology, 4(3):341–353, 2007.DOI
  18. [18]Miklós Bergou, Max Wardetzky, Stephen Robinson, Basile Audoly, and Eitan Grinspun. Discrete elastic rods. In ACM SIGGRAPH 2008 Papers, pages 1–12. 2008.
  19. [19]Wolfgang Kabsch. A solution for the best rotation to relate two sets of vectors. Foundations of Crystallography, 32(5):922–923, 1976.
  20. [20]Wolfgang Kabsch. A discussion of the solution for the best rotation to relate two sets of vectors. Foundations of Crystallography, 34(5):827–828, 1978.
  21. [21]William Shalongo, Laxmichand Dugad, and Earle Stellwagen. Distribution of helicity within the model peptide acetyl(AQAA)3amide. Journal of the American Chemical Society, 116(18):8288–8293, 1994.DOI
  22. [22]Thang K Chiu, Jan Kubelka, Regine Herbst-Irmer, William A Eaton, James Hofrichter, and David R Davies. High-resolution X-ray crystal structures of the villin headpiece subdomain, an ultrafast folding protein. Proceedings of the National Academy of Sciences, 102(21):7517–7522, 2005.pubmed.ncbi.nlm.nih.gov/15894611
  23. [23]Paul Grassein, Patrice Delarue, Adrien Nicolaï, Fabrice Neiers, Harold A Scheraga, Gia G Maisuradze, and Patrick Senet. Curvature and torsion of protein main chain as local order parameters of protein unfolding. The Journal of Physical Chemistry B, 124(22):4391, 2020.DOI
  24. [24]Patrick Senet, Adrien Guzzo, Patrice Delarue, Christophe Laforge, Gia G Maisuradze, Jean-Marie Heydel, Fabrice Neiers, and Adrien Nicolaï. Local entropy in proteins. Chemical Science, 17(11):5482–5497, 2026.DOI
  25. [25]Jaehyeok Jin and Gregory A Voth. Statistical mechanical design principles for coarse-grained interactions across different conformational free energy surfaces. The Journal of Physical Chemistry Letters, 14(6):1354, 2023.DOI
  26. [26]Joseph F Rudzinski and William G Noid. Bottom-up coarse-graining of peptide ensembles and helix–coil transitions. Journal of Chemical Theory and Computation, 11(3):1278–1291, 2015.
  27. [27]Andreas Vitalis and Amedeo Caflisch. 50 years of Lifson–Roig models: Application to molecular simulation data. Journal of Chemical Theory and Computation, 8(1):363–373, 2012.DOI
  28. [28]Robert B Best and Gerhard Hummer. Optimized molecular dynamics force fields applied to the helix-coil transition of polypeptides. The Journal of Physical Chemistry B, 113(26):9004, 2009.DOI
  29. [29]Bhumika Singh, Yisel Martínez-Noa, and Alberto Perez. How well do molecular dynamics force fields model peptides? a systematic benchmark across diverse folding behaviors. The Journal of Physical Chemistry B, 130(16):4344, 2026.DOI
  30. [30]Michael S Jones, Kirill Shmilovich, and Andrew L Ferguson. Diamondback: Diffusion-denoising autoregressive model for non-deterministic backmapping of Cα protein traces. Journal of Chemical Theory and Computation, 19(21):7908–7923, 2023.DOI
  31. [31]Jianxiang Huang, Xin Qiao, Ning Liu, Zongtao Chai, and Shaoyang Lu. Local geometry recovers but cooperative structure does not: Residue-resolved limits of all-atom reconstruction from single-bead coarse-grained disordered protein ensembles. Journal of Chemical Information and Modeling, 66(16):10287–10301, 2026.DOI
  32. [32]Malcolm W MacArthur and Janet M Thornton. Influence of proline residues on protein conformation. Journal of Molecular Biology, 218(2):397–412, 1991.pubmed.ncbi.nlm.nih.gov/2010917
  33. [33]Celina Sikorska and Adam Liwo. Origin of correlations between local conformational states of consecutive amino acid residues and their role in shaping protein structures and in allostery. The Journal of Physical Chemistry B, 126(46):9493, 2022.DOI
  34. [34]William G Noid, Ryan J Szulako, Katherine M Kidder, and Maria C Lesniewski. Rigorous progress in coarse-graining. Annual Review of Physical Chemistry, 75(1):21–45, 2024.DOI
  35. [35]Robert T. McGibbon. Fs MD Trajectories. figshare dataset, version 1, May 2014. URL https://figshare.com/articles/dataset/Fs_MD_Trajectories/1030363.DOI
  36. [36]Jing Huang, Sarah Rauscher, Gregorz Nawrocki, Ting Ran, Michael Feig, Bert L De Groot, Helmut Grubmüller, and Alexander D MacKerell Jr. CHARMM36m: an improved force field for folded and intrinsically disordered proteins. Nature Methods, 14(1):71–73, 2017.DOI
  37. [37]Eric J Sorin and Vijay S Pande. Exploring the helix-coil transition via all-atom equilibrium ensemble simulations. Biophysical Journal, 88(4):2472–2493, 2005.pubmed.ncbi.nlm.nih.gov/15665128
  38. [38]Robert T McGibbon, Kyle A Beauchamp, Matthew P Harrigan, Christoph Klein, Jason M Swails, Carlos X Hernández, Christian R Schwantes, Lee-Ping Wang, Thomas J Lane, and Vijay S Pande. MDTraj: a modern open library for the analysis of molecular dynamics trajectories. Biophysical Journal, 109(8):1528–1532, 2015.DOI
  39. [39]Wolfgang Kabsch and Christian Sander. Dictionary of protein secondary structure: pattern recognition of hydrogen-bonded and geometrical features. Biopolymers: Original Research on Biomolecules, 22(12):2577–2637, 1983.

Paper details

vlnc:W_9dx66b4h
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY-SA 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Agricultural and Biological Sciences

Not recorded for this paper

View statement
Looking for related papers…
Contents