Geometric Vector Perceptron (GVP)

• GVP is a neural module that extends dense layers to process paired scalar and vector features, preserving rotation invariance and equivariance in 3D space.
• Its design enables effective geometric reasoning in tasks like protein design, model quality assessment, and atomic-scale structural prediction by propagating orientation-dependent information.
• Enhanced with vector gating, GVP overcomes limitations of earlier versions by enabling scalar-to-vector information flow, matching or outperforming higher-order equivariant architectures.

Searching arXiv for the cited GVP papers and closely related work to ground the article. Attempting arXiv lookup for "Geometric Vector Perceptron" and the specified arXiv ids. The Geometric Vector Perceptron (GVP) is a neural module for learning on 3D biomolecular structure that extends a standard dense layer to operate on paired scalar features and geometric vector features. In the GVP formulation, scalar channels are invariant under rotation, while vector channels are constructed to be equivariant under 3D rotations and reflections, so that geometric information can be propagated through a graph neural network without discarding orientation-dependent structure. GVPs were introduced for protein-structure learning and then embedded in message-passing architectures, often termed GVP-GNNs, for tasks including computational protein design, model quality assessment, atomic-scale structural biology benchmarks, descriptor-free collective-variable discovery, and RNA inverse folding (Jing et al., 2020).

1. Formal definition and layer computation

A GVP acts on an input pair $(\mathbf{s}, \mathbf{V})$, where $\mathbf{s} \in \mathbb{R}^n$ denotes scalar features and $\mathbf{V} \in \mathbb{R}^{\nu \times 3}$ denotes $\nu$ geometric vectors in $\mathbb{R}^3$. It produces an output pair $(\mathbf{s}', \mathbf{V}')$ with $\mathbf{s}' \in \mathbb{R}^m$ and $\mathbf{V}' \in \mathbb{R}^{\mu \times 3}$ (Jing et al., 2020).

In the original formulation, the computation proceeds by linear transformation of vector channels, extraction of row-wise $L_2$ norms, concatenation of those norms with the input scalars, and separate nonlinear updates of scalar and vector outputs. The equations given for this construction are

$$\begin{align*} \mathbf{V}_h &= \mathbf{W}_h \mathbf{V} \ \mathbf{V}_\mu &= \mathbf{W}_\mu \mathbf{V}_h \ \vec{s}_h &= \|\mathbf{V}_h\|_{row-wise} \ \vec{v}_\mu &= \|\mathbf{V}_\mu\|_{row-wise} \ \vec{s}_{h+n} &= \text{concat}(\vec{s}_h, \mathbf{s}) \ \vec{s}_m &= \mathbf{W}_m \vec{s}_{h+n} + b \ \mathbf{s}' &= \sigma(\vec{s}_m) \ \mathbf{V}' &= \sigma^{+}(\vec{v}_\mu) \odot \mathbf{V}_\mu . \end{align*}$$

Here $\mathbf{s} \in \mathbb{R}^n$0 is a scalar nonlinearity such as ReLU or $\mathbf{s} \in \mathbb{R}^n$1, while $\mathbf{s} \in \mathbb{R}^n$2 is a positive nonlinearity such as softplus or ReLU that scales each output vector channel by a function of its norm. The result is a module in which scalar and vector channels interact through invariant norms, but vector outputs retain their geometric transformation behavior (Jing et al., 2020).

A later “modified GVP” introduces vector gating, in which scalar information modulates vector channels through a gate computed from the updated scalar representation. In the form summarized for the 2021 macromolecular-structure paper, the vector update is

$\mathbf{s} \in \mathbb{R}^n$3

with $\mathbf{s} \in \mathbb{R}^n$4 a sigmoid-like function applied row-wise. This modification explicitly permits scalar-to-vector information flow while maintaining equivariance (Jing et al., 2021).

2. Symmetry properties and approximation results

The central design criterion of the GVP is compatibility with the symmetries of macromolecular geometry. Scalar channels are invariant, whereas vector channels are equivariant: if the input vectors are rotated or reflected, the output vectors transform in the same way. In the original presentation, this is expressed as

$\mathbf{s} \in \mathbb{R}^n$5

because the only vector operations are linear combinations, scalar multiplications, and norm computations, with norms supplying invariant quantities to the scalar pathway (Jing et al., 2020).

This symmetry handling distinguishes GVPs from architectures that reduce geometry to invariant scalars alone. The original paper states that the GVP can approximate any continuous rotation- and reflection-invariant scalar-valued function of its vector inputs. The 2021 extension further states that stacked GVP layers, with the gating modification, can approximate any continuous rotation- and reflection-equivariant function of their inputs. The theoretical distinction is important: the original result concerns invariant scalar-valued functions, whereas the later result addresses equivariant mappings more directly (Jing et al., 2021).

A common misconception is that equivariant molecular learning necessarily requires higher-order tensor representations or spherical harmonic convolutions. The GVP literature positions the module differently: it uses first-order vector representations and nevertheless remains competitive with equivariant networks using higher-order representations and spherical harmonic convolutions. This suggests that substantial geometric expressiveness can be obtained without the computational and conceptual overhead associated with higher-order $\mathbf{s} \in \mathbb{R}^n$6-equivariant constructions (Jing et al., 2021).

3. GVP-GNN as a message-passing architecture

GVPs are typically used as the core nonlinear blocks inside a graph neural network. In this setting, node and edge states remain tuples of scalar and vector features throughout message passing, aggregation, and node updates, enabling simultaneous relational and geometric reasoning (Jing et al., 2020).

In the original protein-structure formulation, nodes are amino acid residues. Node scalar features include dihedral angles encoded as $\mathbf{s} \in \mathbb{R}^n$7 and $\mathbf{s} \in \mathbb{R}^n$8, with amino acid identity optionally included, while node vector features include forward and reverse backbone directions and an imputed C$\mathbf{s} \in \mathbb{R}^n$9 direction. Edges connect each node to its $\mathbf{V} \in \mathbb{R}^{\nu \times 3}$0 nearest sequence or spatial neighbors, typically $\mathbf{V} \in \mathbb{R}^{\nu \times 3}$1; scalar edge features include distance as RBF and sequence separation, and vector edge features encode direction from one residue to another. At each propagation step, messages are produced by a sequence of GVPs applied to concatenated node and edge features, aggregated over neighbors, normalized, and followed by a GVP-based feed-forward update (Jing et al., 2020).

The 2021 atomic-scale extension applies the same design principle to a wider range of macromolecular tasks. There, nodes are atoms and edges are spatially proximate atom pairs; node features include scalar atom-type information and vector features that can be directional, while edge features combine a direction vector with distance encoding. The paper emphasizes that no explicit encoding of bond/types is used to keep the architecture general and uniform. It also identifies the “modified GVP” with vector gating as a key architectural change relative to the original formulation (Jing et al., 2021).

In both versions, the GVP replaces the role ordinarily played by an MLP inside message-passing GNNs. The resulting architecture preserves graph structure, propagates vector-valued geometric information directly, and avoids the need to collapse all geometry into handcrafted invariant descriptors at the input stage.

4. Protein and macromolecular benchmarks

The original GVP paper evaluates the architecture on two protein-structure tasks: computational protein design (CPD) and model quality assessment (MQA). For CPD, the reported datasets are CATH 4.2 and TS50, with perplexity and sequence recovery as the metrics. On the CATH 4.2 full test set, the reported GVP-GNN results are Perplexity 5.29 and Recovery 40.2\%, compared with 6.55, 37.3\% for Structured GNN and 6.85, 36.4\% for Structured Transformer. On TS50, the paper reports 44.9\% recovery. For MQA on CASP 11/12/13, the paper reports that GVP-GNN is the best-performing structure-only method on all datasets; on CASP 11 Stage 2 it gives Global 0.87 and Per-target 0.45, and on CASP 13 it reports Global 0.888 and Per-target 0.671 (Jing et al., 2020).

The atomic-scale extension evaluates GVP-GNN on the ATOM3D benchmark, described as comprising 8 structural biology tasks on atomic-level 3D data. The paper states that the method outperforms all reference architectures on three out of eight tasks in the ATOM3D benchmark, is tied for first on two others, and is competitive with equivariant networks using higher-order representations and spherical harmonic convolutions. In the task-level summary provided, GVP-GNN is best on Protein-Protein Interface (PPI), Residue Identity (RES), and Mutation Stability Prediction (MSP), and tied for first on Small Molecule (SMP) and Protein/RNA Structure Ranking (PSR/RSR). The listed metric values include AUROC 0.866 for PPI, Acc 0.527 for RES, and AUROC 0.680 for MSP (Jing et al., 2021).

The same paper also reports that transfer learning can further improve performance on certain downstream tasks and describes this as the first demonstration for 3D molecular learning. Within the bounds of the reported experiments, this positions GVP-GNN not only as a task-specific architecture but also as a transferable structural representation learner (Jing et al., 2021).

5. Extensions beyond protein learning

Subsequent work applies GVPs outside the original protein-design and protein-quality settings while preserving the same scalar/vector decomposition.

In descriptor-free collective-variable learning for enhanced sampling, a geometric GNN based on GVP is used to learn collective variables directly from atomic coordinates. The construction uses atoms as nodes, atom–atom neighborhoods as edges, one-hot atom types as initial scalar node features, zeros for initial vectors, expanded interatomic distances as scalar edge features, and orientation versors between atoms as vector edge features. Global pooling over graph node outputs yields the final collective variable, and the paper states that the framework is invariant to permutation, overall rotation, and translation. It further introduces node sensitivity analysis and LASSO regression as interpretation tools for the learned collective variables (Zhang et al., 2024).

In RNA inverse folding, a 2025 framework integrates GVP layers with a Transformer architecture for end-to-end RNA design. The workflow described in the paper is: extract key atom coordinates for each nucleotide; compute per-nucleotide structural features; encode each nucleotide as a graph node with scalar features such as torsion angles and vector features such as displacement vectors in a local reference frame; connect each nucleotide to its 5 spatially nearest neighbors; process the graph through 6 GVP layers; and pass the resulting embeddings to a Transformer encoder-decoder for sequence generation. The paper reports state-of-the-art performance, with RNA-Puzzles recovery and TM-scores of 0.481 and 0.332, standard-benchmark values of 0.413 and 0.281, and masked family-level validation values of 0.36 and 0.254. It further states that inverse-folded sequences, when refolded using AlphaFold3, closely resemble native structures (Yao, 31 Dec 2025).

These applications show that GVP is not restricted to a single biological scale. The same module appears in residue-level protein graphs, atomic-level macromolecular benchmarks, atomistic simulation pipelines, and nucleotide-level sequence-design systems.

6. Relation to other geometric architectures and recurring design themes

The GVP literature repeatedly contrasts the module with two prevailing alternatives: graph models that use only scalar features or indirect geometric encodings, and equivariant neural networks based on spherical harmonic convolutions or higher-order tensors. Conventional GNNs are described as excelling at relational reasoning but usually encoding geometry indirectly or as scalar features, while 3D CNNs encode geometry through grids or voxels but do not exploit the natural graph structure of proteins as directly. By comparison, GVP-GNNs are presented as combining geometric and relational reasoning in a single message-passing framework (Jing et al., 2020).

Relative to higher-order equivariant networks such as Cormorant and Tensor Field Networks, the specific claim made in the macromolecular-structure paper is one of simplicity and first-order vector representations rather than tensor orders up to $\mathbf{V} \in \mathbb{R}^{\nu \times 3}$2. The same paper also identifies a limitation of the original GVP: vector outputs depended only on vector inputs, so information flow from scalar to vector channels was missing. The vector-gating modification is introduced precisely to remove that restriction (Jing et al., 2021).

Several design themes recur across the cited applications. First, GVPs maintain a dual channel paradigm in which invariant scalar features and equivariant vector features coexist at every layer. Second, geometric information is preserved throughout message passing rather than compressed into descriptors before learning begins. Third, graph-level outputs are typically obtained through pooling over scalar node outputs, which supports permutation-invariant readout in tasks such as collective-variable construction (Zhang et al., 2024).

A plausible implication is that the enduring role of GVP across these domains derives less from any one benchmark result than from a particular architectural compromise: it preserves explicit 3D vector structure, remains compatible with graph-based relational reasoning, and avoids the full machinery of higher-order equivariant tensor networks. Within the reported literature, that compromise has made GVP a reusable building block for a broad class of macromolecular and atomistic learning problems.