MoMPNN: Protein Design & Graph Neural Networks

• The paper on protein design introduces a post-training ProtAlign method that fine-tunes ProteinMPNN to boost developability metrics while preserving designability.
• The paper on graph learning presents a many-body message passing framework that uses curvature-weighted motif Laplacians and spectral filters to model higher-order node interactions.
• The dual usage of MoMPNN highlights the necessity for contextual disambiguation, as it serves distinct applications in protein inverse folding and graph-based learning.

Searching arXiv for papers on “MoMPNN” to disambiguate the term and ground the article in the relevant literature. arXiv search query: "MoMPNN OR ProtAlign OR many-body Message Passing Neural Networks" “MoMPNN” is an overloaded acronym in the arXiv literature. In one usage, it denotes the model obtained by applying the ProtAlign multi-objective preference alignment framework to ProteinMPNN for property-driven protein inverse folding; in that setting the name is explicitly glossed as “Multi-objective MPNN” (Hou et al., 6 Mar 2026). In a separate usage, it abbreviates “Many-body Message Passing Neural Networks,” a graph-learning framework that models higher-order node interactions through tree-shaped motifs, localized spectral filters, and curvature-weighted motif Laplacians (Han, 2024). The two lines of work are technically unrelated beyond the shared acronym: the former is a post-training alignment method for protein sequence design, whereas the latter is a theoretical and architectural generalization of message passing on graphs.

1. Dual usage and scope

The protein-design MoMPNN is built on the original ProteinMPNN architecture and is positioned as a fine-tuned inverse folding model that balances designability with developability properties such as solubility, thermostability, evolutionary plausibility, and structural fidelity (Hou et al., 6 Mar 2026). The many-body MoMPNN, by contrast, is formulated as a higher-order MPNN that models interactions among $\ge 2$ nodes by enumerating tree-shaped motifs centered at a node and filtering them spectrally (Han, 2024).

This terminological overlap matters because the acronym alone does not identify a single method family. A plausible implication is that discussions of “MoMPNN” require immediate contextual disambiguation: protein inverse folding and multi-objective preference alignment point to the ProtAlign-derived model, whereas motif Laplacians, Ricci curvature, and many-body interactions point to the graph-theoretic framework.

2. MoMPNN as a protein inverse folding model

In the protein-design setting, MoMPNN uses the original ProteinMPNN architecture intact, with no graph-neural or decoder modifications (Hou et al., 6 Mar 2026). ProteinMPNN is described as an order-agnostic autoregressive MPNN that, given a backbone graph $x$, computes

$$\pi_\theta(y\mid x,\sigma)=\prod_{i=1}^{L}\pi_\theta(y_{\sigma(i)}\mid x,y_{\sigma(<i)}),$$

where $\sigma$ is a random residue permutation and the node and edge features encode $C_\alpha/N/O$ geometry (Hou et al., 6 Mar 2026). The extension consists not in changing layers but in treating the pretrained ProteinMPNN as both the reference policy $\pi_{\mathrm{ref}}$ and the policy $\pi_\theta$ to be fine-tuned.

Because the model is permutation-invariant, log-ratios in the training loss are estimated by averaging over $K$ sampled permutations:

$$\hat p_\theta(y\mid x)=\frac{1}{K}\sum_{k=1}^{K}\pi_\theta(y\mid x,\sigma_k), \qquad \hat p_{\mathrm{ref}}(y\mid x)=\frac{1}{K}\sum_{k=1}^{K}\pi_{\mathrm{ref}}(y\mid x,\sigma_k).$$

The stated objective is to enhance developability without compromising designability across sequence design for CATH 4.3 crystal structures, de novo generated backbones, and real-world binder design scenarios (Hou et al., 6 Mar 2026). The developability properties used for preference construction are solubility via Protein-Sol predictor, thermostability via TemBERTure, and evolutionary plausibility via ESM-2 pseudo-perplexity; auxiliary designability properties are TM-score computed by TMalign on ESMFold-predicted structure and pTM or pLDDT confidence from AlphaFold initial guess (Hou et al., 6 Mar 2026).

This usage of MoMPNN is therefore best understood not as a new neural architecture, but as a fine-tuned ProteinMPNN checkpoint produced by a multi-objective alignment procedure.

3. Multi-objective preference alignment in ProtAlign

The training principle is a semi-online Direct Preference Optimization strategy. The paper defines a multi-objective policy objective that maximizes expected weighted rewards while penalizing deviation from the reference policy:

$$\arg\max_\theta L(\pi_\theta)=\mathbb{E}_{x,y\sim\pi_\theta}\left[\sum_k w_k r_k(x,y)\right]-\beta\,\mathrm{KL}(\pi_\theta(\cdot\mid x)\Vert \pi_{\mathrm{ref}}(\cdot\mid x)).$$

Preference comparisons are modeled with a Bradley–Terry preference model,

$x$0

and combined into a flexible-margin multi-objective DPO loss. For each property $x$1, the loss term is

$x$2

with adaptive margin

$x$3

The paper states that the margin $x$4 reduces the strength of preference for property $x$5 when $x$6 performs worse on other objectives $x$7 (Hou et al., 6 Mar 2026). The overall training loss is

$x$8

Preference pairs are generated semi-online. For each backbone $x$9, the method samples $$\pi_\theta(y\mid x,\sigma)=\prod_{i=1}^{L}\pi_\theta(y_{\sigma(i)}\mid x,y_{\sigma(<i)}),$$0 sequences under $$\pi_\theta(y\mid x,\sigma)=\prod_{i=1}^{L}\pi_\theta(y_{\sigma(i)}\mid x,y_{\sigma(<i)}),$$1 at rollout temperature $$\pi_\theta(y\mid x,\sigma)=\prod_{i=1}^{L}\pi_\theta(y_{\sigma(i)}\mid x,y_{\sigma(<i)}),$$2, computes property values $$\pi_\theta(y\mid x,\sigma)=\prod_{i=1}^{L}\pi_\theta(y_{\sigma(i)}\mid x,y_{\sigma(<i)}),$$3, ranks the candidates, and forms $$\pi_\theta(y\mid x,\sigma)=\prod_{i=1}^{L}\pi_\theta(y_{\sigma(i)}\mid x,y_{\sigma(<i)}),$$4 pairs by pairing the $$\pi_\theta(y\mid x,\sigma)=\prod_{i=1}^{L}\pi_\theta(y_{\sigma(i)}\mid x,y_{\sigma(<i)}),$$5-th best with the $$\pi_\theta(y\mid x,\sigma)=\prod_{i=1}^{L}\pi_\theta(y_{\sigma(i)}\mid x,y_{\sigma(<i)}),$$6-th; pairs are retained only when $$\pi_\theta(y\mid x,\sigma)=\prod_{i=1}^{L}\pi_\theta(y_{\sigma(i)}\mid x,y_{\sigma(<i)}),$$7 (Hou et al., 6 Mar 2026). The training pseudocode is organized into a rollout phase, in which sequences are sampled and scored, and a training phase, in which parameters are updated using $$\pi_\theta(y\mid x,\sigma)=\prod_{i=1}^{L}\pi_\theta(y_{\sigma(i)}\mid x,y_{\sigma(<i)}),$$8.

For the reported MoMPNN[Sol+IG+EP] example, the stated hyperparameters are Adam$$\pi_\theta(y\mid x,\sigma)=\prod_{i=1}^{L}\pi_\theta(y_{\sigma(i)}\mid x,y_{\sigma(<i)}),$$9, learning rate $\sigma$0, $\sigma$1 rounds, approximately $\sigma$2 steps per round, batch size $\sigma$3 across $\sigma$4 GPUs, DPO strength $\sigma$5, margin scale $\sigma$6, and objective weights $\sigma$7, $\sigma$8, $\sigma$9, $C_\alpha/N/O$0, $C_\alpha/N/O$1 (Hou et al., 6 Mar 2026). Conflict mitigation is attributed to uniform sampling across property datasets and adaptive margins that down-weight incompatible pairs.

4. Empirical behavior of the protein-design MoMPNN

The reported results span CATH 4.3 crystal backbones, de novo backbones, and de novo binder design (Hou et al., 6 Mar 2026). On CATH 4.3 crystal backbones, the model is reported to preserve designability relative to ProteinMPNN while improving developability: RMSD is approximately $C_\alpha/N/O$2, TM is approximately $C_\alpha/N/O$3, and pLDDT is approximately $C_\alpha/N/O$4; solubility increases from $C_\alpha/N/O$5 to $C_\alpha/N/O$6 for MoMPNN[Sol+TM], and thermostability increases from $C_\alpha/N/O$7 to $C_\alpha/N/O$8 for MoMPNNThermo+IG. The same section reports that it outperforms subset-trained SolubleMPNN, with Sol approximately $C_\alpha/N/O$9, and HyperMPNN, with Thermo approximately $\pi_{\mathrm{ref}}$0, without designability loss.

On de novo backbones, the baseline ProteinMPNN metrics are listed as TM approximately $\pi_{\mathrm{ref}}$1, EP approximately $\pi_{\mathrm{ref}}$2, Sol approximately $\pi_{\mathrm{ref}}$3, and Thermo approximately $\pi_{\mathrm{ref}}$4 (Hou et al., 6 Mar 2026). MoMPNN[Sol+IG+EP] is reported at TM approximately $\pi_{\mathrm{ref}}$5, EP approximately $\pi_{\mathrm{ref}}$6, Sol approximately $\pi_{\mathrm{ref}}$7, and Thermo approximately $\pi_{\mathrm{ref}}$8, while MoMPNN[Thermo+IG] reaches TM approximately $\pi_{\mathrm{ref}}$9, EP approximately $\pi_\theta$0, and Thermo approximately $\pi_\theta$1 (Hou et al., 6 Mar 2026).

In de novo binder design, the reported sequence success rate increases from approximately $\pi_\theta$2 for ProteinMPNN to approximately $\pi_\theta$3 for MoMPNN[Sol+IG+EP], and backbone success increases from approximately $\pi_\theta$4 to approximately $\pi_\theta$5 (Hou et al., 6 Mar 2026). EP is reduced, Sol is increased, and pLDDT and inter-chain PAE are maintained.

These results are presented as evidence that a post-training preference-alignment procedure can alter the property profile of inverse-folded sequences while preserving structural recovery metrics. This suggests that MoMPNN, in this sense, functions as a deployable adaptation of ProteinMPNN rather than a replacement for it.

5. MoMPNN as many-body message passing on graphs

In the graph-learning literature, MoMPNN denotes a many-body Message Passing Neural Network framework that explicitly models higher-order node interactions (Han, 2024). The central construct is the tree-shaped motif of order $\pi_\theta$6: for a central node $\pi_\theta$7 and a choice of $\pi_\theta$8 neighbors $\pi_\theta$9, the motif is the induced undirected tree on nodes $K$0.

Edges in a motif are weighted by their global Ricci curvatures $K$1, yielding a motif Laplacian $K$2 with entries

$K$3

On each motif Laplacian, the framework applies a Chebyshev-polynomial spectral filter. If $K$4 is the diagonal matrix of eigenvalues of $K$5 and $K$6 its largest eigenvalue, the rescaled spectrum is

$K$7

and the filter is

$K$8

where $K$9 is the $$\hat p_\theta(y\mid x)=\frac{1}{K}\sum_{k=1}^{K}\pi_\theta(y\mid x,\sigma_k), \qquad \hat p_{\mathrm{ref}}(y\mid x)=\frac{1}{K}\sum_{k=1}^{K}\pi_{\mathrm{ref}}(y\mid x,\sigma_k).$$0-th Chebyshev polynomial (Han, 2024).

The two-body graph convolution is written as

$$\hat p_\theta(y\mid x)=\frac{1}{K}\sum_{k=1}^{K}\pi_\theta(y\mid x,\sigma_k), \qquad \hat p_{\mathrm{ref}}(y\mid x)=\frac{1}{K}\sum_{k=1}^{K}\pi_{\mathrm{ref}}(y\mid x,\sigma_k).$$1

while the $$\hat p_\theta(y\mid x)=\frac{1}{K}\sum_{k=1}^{K}\pi_\theta(y\mid x,\sigma_k), \qquad \hat p_{\mathrm{ref}}(y\mid x)=\frac{1}{K}\sum_{k=1}^{K}\pi_{\mathrm{ref}}(y\mid x,\sigma_k).$$2-body message at node $$\hat p_\theta(y\mid x)=\frac{1}{K}\sum_{k=1}^{K}\pi_\theta(y\mid x,\sigma_k), \qquad \hat p_{\mathrm{ref}}(y\mid x)=\frac{1}{K}\sum_{k=1}^{K}\pi_{\mathrm{ref}}(y\mid x,\sigma_k).$$3 is

$$\hat p_\theta(y\mid x)=\frac{1}{K}\sum_{k=1}^{K}\pi_\theta(y\mid x,\sigma_k), \qquad \hat p_{\mathrm{ref}}(y\mid x)=\frac{1}{K}\sum_{k=1}^{K}\pi_{\mathrm{ref}}(y\mid x,\sigma_k).$$4

The node update is residual:

$$\hat p_\theta(y\mid x)=\frac{1}{K}\sum_{k=1}^{K}\pi_\theta(y\mid x,\sigma_k), \qquad \hat p_{\mathrm{ref}}(y\mid x)=\frac{1}{K}\sum_{k=1}^{K}\pi_{\mathrm{ref}}(y\mid x,\sigma_k).$$5

The stated practical motivation is that small tree-motif enumeration together with localized spectral filtering allows $$\hat p_\theta(y\mid x)=\frac{1}{K}\sum_{k=1}^{K}\pi_\theta(y\mid x,\sigma_k), \qquad \hat p_{\mathrm{ref}}(y\mid x)=\frac{1}{K}\sum_{k=1}^{K}\pi_{\mathrm{ref}}(y\mid x,\sigma_k).$$6-body, $$\hat p_\theta(y\mid x)=\frac{1}{K}\sum_{k=1}^{K}\pi_\theta(y\mid x,\sigma_k), \qquad \hat p_{\mathrm{ref}}(y\mid x)=\frac{1}{K}\sum_{k=1}^{K}\pi_{\mathrm{ref}}(y\mid x,\sigma_k).$$7-body, and up to $$\hat p_\theta(y\mid x)=\frac{1}{K}\sum_{k=1}^{K}\pi_\theta(y\mid x,\sigma_k), \qquad \hat p_{\mathrm{ref}}(y\mid x)=\frac{1}{K}\sum_{k=1}^{K}\pi_{\mathrm{ref}}(y\mid x,\sigma_k).$$8-body interactions to be captured in one layer, enlarging the receptive field without stacking many layers (Han, 2024).

6. Curvature weighting, theory, and experiments for the many-body MoMPNN

The curvature term is Balanced Forman curvature of the global edge $$\hat p_\theta(y\mid x)=\frac{1}{K}\sum_{k=1}^{K}\pi_\theta(y\mid x,\sigma_k), \qquad \hat p_{\mathrm{ref}}(y\mid x)=\frac{1}{K}\sum_{k=1}^{K}\pi_{\mathrm{ref}}(y\mid x,\sigma_k).$$9, following Topping et al. (ICLR’22) as stated in the summary (Han, 2024). When inserted into motif Laplacians, the off-diagonal entry $$\arg\max_\theta L(\pi_\theta)=\mathbb{E}_{x,y\sim\pi_\theta}\left[\sum_k w_k r_k(x,y)\right]-\beta\,\mathrm{KL}(\pi_\theta(\cdot\mid x)\Vert \pi_{\mathrm{ref}}(\cdot\mid x)).$$0 alters the spectral weighting of edges; the summary notes that large-magnitude negative Ricci, described there as a “bottleneck” edge, becomes a large positive weight and causes the filter to pay more attention to signals through that edge.

Three theoretical guarantees are highlighted. First, permutation invariance: for the message construction $$\arg\max_\theta L(\pi_\theta)=\mathbb{E}_{x,y\sim\pi_\theta}\left[\sum_k w_k r_k(x,y)\right]-\beta\,\mathrm{KL}(\pi_\theta(\cdot\mid x)\Vert \pi_{\mathrm{ref}}(\cdot\mid x)).$$1, the framework satisfies

$$\arg\max_\theta L(\pi_\theta)=\mathbb{E}_{x,y\sim\pi_\theta}\left[\sum_k w_k r_k(x,y)\right]-\beta\,\mathrm{KL}(\pi_\theta(\cdot\mid x)\Vert \pi_{\mathrm{ref}}(\cdot\mid x)).$$2

for any permutation matrix $$\arg\max_\theta L(\pi_\theta)=\mathbb{E}_{x,y\sim\pi_\theta}\left[\sum_k w_k r_k(x,y)\right]-\beta\,\mathrm{KL}(\pi_\theta(\cdot\mid x)\Vert \pi_{\mathrm{ref}}(\cdot\mid x)).$$3 (Han, 2024). Second, the sensitivity bound for graph-distance $$\arg\max_\theta L(\pi_\theta)=\mathbb{E}_{x,y\sim\pi_\theta}\left[\sum_k w_k r_k(x,y)\right]-\beta\,\mathrm{KL}(\pi_\theta(\cdot\mid x)\Vert \pi_{\mathrm{ref}}(\cdot\mid x)).$$4 between nodes $$\arg\max_\theta L(\pi_\theta)=\mathbb{E}_{x,y\sim\pi_\theta}\left[\sum_k w_k r_k(x,y)\right]-\beta\,\mathrm{KL}(\pi_\theta(\cdot\mid x)\Vert \pi_{\mathrm{ref}}(\cdot\mid x)).$$5 and $$\arg\max_\theta L(\pi_\theta)=\mathbb{E}_{x,y\sim\pi_\theta}\left[\sum_k w_k r_k(x,y)\right]-\beta\,\mathrm{KL}(\pi_\theta(\cdot\mid x)\Vert \pi_{\mathrm{ref}}(\cdot\mid x)).$$6 is

$$\arg\max_\theta L(\pi_\theta)=\mathbb{E}_{x,y\sim\pi_\theta}\left[\sum_k w_k r_k(x,y)\right]-\beta\,\mathrm{KL}(\pi_\theta(\cdot\mid x)\Vert \pi_{\mathrm{ref}}(\cdot\mid x)).$$7

which the summary contrasts with standard MPNNs that have only the factor $$\arg\max_\theta L(\pi_\theta)=\mathbb{E}_{x,y\sim\pi_\theta}\left[\sum_k w_k r_k(x,y)\right]-\beta\,\mathrm{KL}(\pi_\theta(\cdot\mid x)\Vert \pi_{\mathrm{ref}}(\cdot\mid x)).$$8 (Han, 2024). Third, the learned Dirichlet energy $$\arg\max_\theta L(\pi_\theta)=\mathbb{E}_{x,y\sim\pi_\theta}\left[\sum_k w_k r_k(x,y)\right]-\beta\,\mathrm{KL}(\pi_\theta(\cdot\mid x)\Vert \pi_{\mathrm{ref}}(\cdot\mid x)).$$9 is bounded as

$x$00

The summary states that higher correlation order $x$01 strictly increases the upper bound, allowing MoMPNN to generate more energy than ChebNet with $x$02 in the same architecture (Han, 2024).

The experiments include graph-energy regression on $x$03 random Erdős–Rényi graphs with $x$04–$x$05 nodes and edge probability $x$06, heterophilic node classification on a single synthetic graph of $x$07 nodes and $x$08 classes, and an efficiency benchmark on an NVIDIA RTX 2080 Ti (Han, 2024). The reported observations are that deeper many-body networks outperform GCN and ChebNet when the regression target depends on global distances; wider many-body networks perform best when the target depends on local clustering; test accuracy is on par with or above standard GCN and ChebNet on the heterophilic graph; and Dirichlet energy grows much higher during training (Han, 2024). For efficiency, the many-body MPNN at $x$09 layers and $x$10 is reported as approximately $x$11 slower than ChebNet, while still scaling linearly with layers as predicted (Han, 2024).

A plausible implication is that this MoMPNN is aimed at the classical graph-learning problems of over-squashing and over-smoothing, rather than at domain-specific property optimization.

7. Limitations, applications, and disambiguation in practice

For the protein-design MoMPNN, the reported advantages are plug-in post-training on any pretrained inverse-folding model with no architectural change, unified treatment of arbitrary properties, and semi-online DPO that is described as stable and efficient because rollout and training are decoupled (Hou et al., 6 Mar 2026). The stated limitations are that all results are in silico, with no wet-lab validation yet, and that the focus is on monomer properties rather than complex-specific objectives such as binding affinity (Hou et al., 6 Mar 2026). Potential applications explicitly listed are large-scale protein sequence design pipelines requiring high developability, functional design tasks by adding further objectives such as binding energy or surface hydrophobicity, and extension to other backbones or language-model-based inverse folding frameworks (Hou et al., 6 Mar 2026).

For the many-body MoMPNN, the practical advantages stated in the summary are its ability to avoid over-squashing, avoid over-smoothing, remain permutation-invariant, and maintain a controlled energy range, together with empirical benefits on regression and heterophilic node classification (Han, 2024). The efficiency benchmark also notes that future implementations can reduce overhead close to two-body complexity, citing Proposition 5.1 in the paper summary (Han, 2024).

Because the acronym now names two unrelated constructs, precise usage is essential. In protein design, “MoMPNN” refers to a ProteinMPNN model fine-tuned by ProtAlign’s flexible-margin multi-objective DPO. In graph representation learning, “MoMPNN” refers to a many-body message passing architecture built from curvature-weighted motif Laplacians and localized spectral filters. The overlap is nominal rather than methodological.