Lie-RMSD: Differentiable Protein Alignment

• The paper demonstrates that a differentiable, Lie algebra-based framework can recover RMSD minima equivalent to the analytical Kabsch solution.
• It reframes protein structural alignment as a gradient optimization problem over a 6D Lie algebra parameter space using automatic differentiation in PyTorch.
• Benchmarking on Adenylate Kinase highlights its precision and potential to extend to custom, differentiable structural loss functions beyond plain RMSD.

Searching arXiv for the specified paper and closely related work on RMSD, Lie-group formulations, and symmetry-aware structural alignment. Lie-RMSD is a gradient-based framework for protein structural alignment that represents rigid-body transformations in the Lie algebra $\mathfrak{se}(3)$ of the special Euclidean group $\mathrm{SE}(3)$. Instead of solving alignment analytically, as in the Kabsch algorithm, it expresses alignment as a differentiable loss and uses modern optimizers to minimize it. In the formulation introduced in "Lie-RMSD: A Gradient-Based Framework for Protein Structural Alignment using Lie Algebra" (Hu et al., 23 Aug 2025), the method is fully differentiable, benchmarked on two allosteric conformations of Adenylate Kinase, and shown to converge to the same global minimum as the analytical Kabsch solution with precision effectively identical to it. Its stated significance is not superior speed for plain RMSD minimization, but the establishment of a robust foundation for differentiable structural objectives for which no analytical solution exists.

1. Definition and problem setting

Lie-RMSD treats protein structural alignment as minimization of the root mean square deviation between two sets of corresponding atomic coordinates after a rigid-body transformation. Given corresponding coordinates $\{x_i\}$ and $\{y_i\}$, RMSD is written as

$$\mathrm{RMSD} = \sqrt{\frac{1}{N} \sum_{i=1}^N \lVert x_i^{aligned} - y_i \rVert^2}.$$

In structural biology, this minimization is used to quantify geometric similarity between protein conformations, compare predicted versus experimental structures, track conformational changes, validate modeling and docking, and analyze evolutionary relationships (Hu et al., 23 Aug 2025).

The immediate motivation for Lie-RMSD is the limitation of traditional analytical methods such as Kabsch. The Kabsch algorithm gives a closed-form solution for the rigid transformation that minimizes RMSD between paired points, but that derivation is tailored specifically to RMSD. The 2025 Lie-RMSD formulation therefore reframes alignment as a differentiable optimization problem: once a forward loss can be computed and backpropagated, the same machinery can in principle optimize arbitrary differentiable structural scores rather than only the quadratic RMSD objective.

A common misconception is that Lie-RMSD is primarily a replacement for Kabsch in routine RMSD alignment. The benchmarked formulation does not support that interpretation. For standard RMSD alignment, the analytical method remains the exact and much faster tool; Lie-RMSD is presented instead as a general differentiable alignment framework whose value appears when the target objective ceases to admit a closed-form solution (Hu et al., 23 Aug 2025).

2. Lie-algebra parameterization of rigid-body motion

The core parameterization uses a 6-dimensional vector

$$\mathbf{p} = (\boldsymbol{\omega}, \mathbf{u}) \in \mathfrak{se}(3),$$

where $\boldsymbol{\omega} \in \mathbb{R}^3$ parameterizes rotation and $\mathbf{u} \in \mathbb{R}^3$ parameterizes translation. In homogeneous coordinates, a rigid transform in $\mathrm{SE}(3)$ is written as

$$T = \begin{pmatrix} R & \mathbf{t} \ 0 & 1 \end{pmatrix}, \quad R \in SO(3),\; \mathbf{t} \in \mathbb{R}^3,$$

and an element of $\mathrm{SE}(3)$0 is represented by the twist matrix

$\mathrm{SE}(3)$1

with

$\mathrm{SE}(3)$2

The rigid-body transform is obtained through the matrix exponential,

$\mathrm{SE}(3)$3

In the practical representation emphasized in the paper, the rotational part is

$\mathrm{SE}(3)$4

the translation is treated as a 3D vector, and the transformed centered coordinate is

$\mathrm{SE}(3)$5

This mapping is smooth and differentiable with respect to $\mathrm{SE}(3)$6 and $\mathrm{SE}(3)$7 (Hu et al., 23 Aug 2025).

For centered mobile coordinates $\mathrm{SE}(3)$8 and centered reference coordinates $\mathrm{SE}(3)$9, the loss is

$\{x_i\}$0

Under the Lie parameterization, $\{x_i\}$1 and $\{x_i\}$2 are functions of $\{x_i\}$3, so RMSD becomes a differentiable scalar objective over $\{x_i\}$4. The paper does not derive explicit closed-form gradients; instead, it implements the exponential map, matrix operations, and RMSD in PyTorch and lets automatic differentiation compute derivatives through the chain rule (Hu et al., 23 Aug 2025).

3. Optimization framework and implementation

Lie-RMSD casts alignment as

$\{x_i\}$5

and applies standard deep-learning-style optimizers to the same 6D parameter vector. The benchmark uses SGD, Adam, AdamW, and Sophia. In the stated implementation, the framework runs in PyTorch on CPU, uses a learning rate of $\{x_i\}$6, and performs 1000 optimization steps for all optimizers (Hu et al., 23 Aug 2025).

The implementation emphasis is not on deriving a specialized optimizer for rigid alignment, but on validating that a standard autodiff stack is sufficient. The Lie algebra parameterization avoids direct optimization over rotation matrices, which would have to satisfy orthogonality and determinant constraints. By optimizing an unconstrained 6D vector and mapping it into $\{x_i\}$7 through the exponential map, the resulting rotation is guaranteed to be a valid element of $\{x_i\}$8. This eliminates the need for direct orthonormality enforcement or quaternion normalization.

The reported numerical behavior is correspondingly straightforward. All operations are handled within PyTorch; the matrix exponential and matrix operations are inherited from the framework; and the optimization landscape is described as well behaved for RMSD. Finite-precision effects remain visible only as tiny differences, on the order of $\{x_i\}$9 Å in the final RMSD relative to Kabsch. The benchmark further reports that no optimizer gets stuck in poor local minima, which the paper relates to the convex nature of the RMSD objective under rigid-body motion (Hu et al., 23 Aug 2025).

4. Benchmark on Adenylate Kinase and comparison with Kabsch

The empirical evaluation aligns two Adenylate Kinase conformations: 4AKE chain A in an open conformation and 1AKE chain A in a closed conformation. Both chains contain 214 C$\{y_i\}$0 atoms, so no residue mapping or truncation is needed. This benchmark is structurally relevant because Adenylate Kinase is a classic allosteric system whose open and closed states involve substantial domain motions while preserving the underlying fold (Hu et al., 23 Aug 2025).

The Kabsch baseline gives a final RMSD of 7.130699 Å in 0.51 ms. The gradient-based methods converge to essentially the same result: Adam reaches 7.130700 Å in 557.67 ms, SGD 7.130702 Å in 549.55 ms, Sophia 7.130710 Å in 587.31 ms, and AdamW 7.130717 Å in 582.88 ms. The differences from Kabsch span $\{y_i\}$1 to $\{y_i\}$2 Å, i.e. $\{y_i\}$3–$\{y_i\}$4 Å, and are attributed to floating-point precision and stopping criteria. Visual overlays are reported as indistinguishable across all methods (Hu et al., 23 Aug 2025).

The comparison clarifies the method’s scope. For pure RMSD alignment, Kabsch retains decisive advantages: exact analytic solution, deterministic behavior, no hyperparameter tuning, and runtime that is roughly three orders of magnitude lower in the reported CPU benchmark. Lie-RMSD’s contribution is therefore methodological rather than competitive in speed. It validates that a Lie algebra-based gradient descent procedure can recover the same optimum as the closed-form solution, which is the prerequisite for replacing the loss while keeping the same rigid-motion parameterization.

5. Generalization beyond plain RMSD

The principal extension claimed for Lie-RMSD is the ability to replace RMSD with any differentiable structural loss while keeping the same $\{y_i\}$5 parameterization and gradient-based optimization machinery. The paper explicitly lists biophysical and energetic terms such as steric clash penalties, electrostatic or hydrophobic interaction scores, and docking energy functions; topological or global similarity measures such as smooth approximations or differentiable proxies to TM-score or GDT-TS; and data-driven scores such as learned contact or interface scores or losses derived from networks that evaluate nativeness or structure quality (Hu et al., 23 Aug 2025).

This generality is the basis for the paper’s description of Lie-RMSD as a proof-of-concept. It is intended as a validated scaffold for differentiable docking, flexible alignment, multi-domain alignment with multiple transforms, integration into neural networks for structure prediction or refinement, and custom scoring functions for tasks such as interface optimization and motif alignment. A plausible implication is that Lie-RMSD is best understood not as an endpoint algorithm for RMSD itself, but as an alignment layer that can be embedded in broader differentiable structural pipelines.

The same section of the work also situates the method within current automatic-differentiation practice. Because rigid-body alignment becomes “just another differentiable layer or module,” the approach is naturally compatible with end-to-end training in deep learning settings. This suggests a direct route from classical structural superposition toward learned objectives that are geometrically constrained by $\{y_i\}$6 but no longer analytically reducible to Kabsch-style linear algebra (Hu et al., 23 Aug 2025).

6. Position within the broader Lie-group and symmetry-aware RMSD literature

The 2025 Lie-RMSD paper uses the term in a specific sense: a differentiable protein alignment framework built on a 6D $\{y_i\}$7 parameterization and optimized by backpropagation. Related literature uses Lie-group structure more broadly. "Similarity of particle systems using an invariant root mean square deviation measure" (Bulin et al., 2021) defines an invariant RMSD by minimizing over permutations, orthogonal transformations, and translations, i.e. over $\{y_i\}$8 together with $\{y_i\}$9. "Metrics for measuring distances in configuration spaces" (Sadeghi et al., 2013) studies global RMSD minimized over translations, rotations in $$\mathrm{RMSD} = \sqrt{\frac{1}{N} \sum_{i=1}^N \lVert x_i^{aligned} - y_i \rVert^2}.$$0, and permutations, and introduces invariant fingerprint metrics that correlate with that global RMSD. "Geometrical comparison of two protein structures using Wigner-D functions" (Fathi et al., 2013) places comparison in the representation space of $$\mathrm{RMSD} = \sqrt{\frac{1}{N} \sum_{i=1}^N \lVert x_i^{aligned} - y_i \rVert^2}.$$1 through Wigner–D expansions and a reciprocal-space RMSD-type similarity value. "Canonized then Minimized RMSD for Three-Dimensional Structures" (Li et al., 2024) combines symmetry-aware canonization with minimized RMSD over rigid motions and chemically permissible permutations. "OTMol: Robust Molecular Structure Comparison via Optimal Transport" (Wei et al., 1 Sep 2025) separates atom correspondence from rigid alignment by using fused supervised Gromov–Wasserstein matching followed by Kabsch alignment in $$\mathrm{RMSD} = \sqrt{\frac{1}{N} \sum_{i=1}^N \lVert x_i^{aligned} - y_i \rVert^2}.$$2, with reflections disallowed.

This broader context suggests that Lie-RMSD belongs to a larger family of group-aware structural comparison methods. What distinguishes the 2025 formulation is not merely the use of $$\mathrm{RMSD} = \sqrt{\frac{1}{N} \sum_{i=1}^N \lVert x_i^{aligned} - y_i \rVert^2}.$$3, which is common to rigid alignment, but the decision to optimize directly in Lie algebra coordinates inside an autodiff framework. Invariant-RMSD methods minimize over group actions analytically or combinatorially; Wigner–D methods move comparison into representation space; canonization and optimal-transport methods focus on the discrete correspondence problem; Lie-RMSD, in the narrow 2025 sense, centers the continuous rigid-motion component as a differentiable optimization primitive.

The reference implementation is available at $$\mathrm{RMSD} = \sqrt{\frac{1}{N} \sum_{i=1}^N \lVert x_i^{aligned} - y_i \rVert^2}.$$5 and is described as containing code for the $$\mathrm{RMSD} = \sqrt{\frac{1}{N} \sum_{i=1}^N \lVert x_i^{aligned} - y_i \rVert^2}.$$4 representation and exponential mapping, RMSD loss implementation in PyTorch, and benchmark scripts for Adenylate Kinase alignment with different optimizers (Hu et al., 23 Aug 2025). In practice, its reported workflow is to prepare paired coordinates, center them, initialize the 6D parameter vector, define RMSD or another differentiable loss, and optimize with a standard optimizer. For the benchmarked RMSD case, the central conclusion remains precise: Lie-RMSD matches Kabsch almost exactly, while its real utility lies in differentiable structural objectives for which Kabsch has no analogue.