Property-Variational Autoencoder (pVAE)

• pVAE is a family of variational autoencoders that incorporates explicit property information via conditioning, regression, or graph-based isometric regularization.
• It uses mechanisms like conditional priors, auxiliary regression on latent means, and property-structured graph smoothing to organize the latent space.
• Applications include design in porous metamaterials, DNA nanoclusters, and antimicrobial peptides, with strong performance in reconstruction and latent neighborhood purity.

Searching arXiv for the specified pVAE-related papers to ground the article in current literature. Property-Variational Autoencoder (pVAE) denotes a class of variational autoencoder frameworks in which explicit property information governs latent-variable modeling, generation, or latent-space organization. In the cited literature, the term covers at least three closely related constructions: a conditional VAE specialized to generation from a property vector $c$, following the arbitrary-conditioning VAE formulation; a VAE augmented with a regressor that predicts target properties from the latent mean; and a geometry-preserving sequence model, called PrIVAE, that aligns latent neighborhoods with a property manifold through graph-based smoothing and an isometric regularizer (Ivanov et al., 2018, Nguyen et al., 23 Jul 2025, Sadeghi et al., 16 Sep 2025).

1. Terminological scope and canonical variants

The literature does not use pVAE for a single canonical architecture. Rather, the name is attached to different mechanisms for coupling a VAE to target properties. One line treats the property vector $c$ as always observed and trains the model to generate the remaining data $x$ consistently with those properties. A second line retains a standard VAE backbone and adds a regression head $f_\psi$ operating on the latent mean $\mu$, so that property supervision appears as an auxiliary loss. A third line, PrIVAE, uses experimentally measured high-dimensional property vectors $y_i$ to define a property nearest-neighbor graph and then organizes latent codes to preserve local property geometry (Ivanov et al., 2018, Nguyen et al., 23 Jul 2025, Sadeghi et al., 16 Sep 2025).

Formulation	Property mechanism	Representative paper
Conditioning-based pVAE	$c$ is always observed; prior and decoder are conditioned on $c$	(Ivanov et al., 2018)
Regressor-augmented pVAE	Add $f_\psi(\mu)$ and a regression loss on true properties $P^t$	(Nguyen et al., 23 Jul 2025)
Property-isometric pVAE / PrIVAE	Property graph, GCN encoder layers, and isometric regularizer	(Sadeghi et al., 16 Sep 2025)

A common misconception is that pVAE necessarily means direct conditioning on a low-dimensional property vector. The published formulations show a broader landscape: conditioning, auxiliary regression, and latent-geometry alignment all fall under the label in different application domains. This suggests that pVAE is best understood as a property-aware VAE family rather than a uniquely specified model class.

2. Probabilistic formulations

In the conditioning-based formulation derived from arbitrary-conditioning VAE, the full probabilistic model consists of a latent prior conditioned on the property vector, a decoder conditioned on both the latent code and the property vector, and an approximate posterior conditioned on observed features and properties: $c$0

$c$1

$c$2

Training introduces a binary mask $c$3, where $c$4 means $c$5 is to be reconstructed, with

$c$6

The variational lower bound is

$c$7

In the pure generate-from-$c$8 mode, one sets $c$9 for all $x$0, so the ELBO reduces to

$x$1

At evaluation time, one fixes $x$2, sets all $x$3-dimensions to missing, samples $x$4, then samples $x$5 (Ivanov et al., 2018).

In the regressor-augmented pVAE for porous metamaterials, the probabilistic core remains that of a conventional VAE,

$x$6

with ELBO

$x$7

In practice the model minimizes the negative ELBO,

$x$8

and adds a property regression term

$x$9

The combined objective is

$f_\psi$0

or, in the $f_\psi$1-VAE variant,

$f_\psi$2

A notable design choice is that the regressor operates on $f_\psi$3 rather than sampled $f_\psi$4, so that the regressor sees a noise-free latent representation (Nguyen et al., 23 Jul 2025).

PrIVAE adopts a $f_\psi$5-VAE backbone and adds a geometry-preserving penalty on latent codes. With

$f_\psi$6

and prior $f_\psi$7, the VAE term is

$f_\psi$8

The isometric regularization is

$f_\psi$9

The total loss per minibatch is

$\mu$0

Equivalently,

$\mu$1

(Sadeghi et al., 16 Sep 2025).

3. Architectural patterns

The conditioning-based pVAE uses three networks: an encoder $\mu$2, a conditional prior $\mu$3, and a decoder $\mu$4. The encoder takes $\mu$5 as input, passes it through a few fully-connected or convolutional layers, and outputs $\mu$6 and $\mu$7. The prior network takes $\mu$8, or simply $\mu$9 if the mask always indicates all-missing $y_i$0, and outputs $y_i$1 and $y_i$2. The decoder takes $y_i$3, often omitting $y_i$4 in the pure pVAE case, and outputs either mean/variance parameters for real-valued features or logits for categorical features. The latent variable is sampled via the reparameterization

$y_i$5

The formulation also allows optional skip-connections from encoder to decoder for image inpainting (Ivanov et al., 2018).

In the porous-metamaterials pVAE, the encoder $y_i$6 receives a binary 3D microstructure tensor, for example $y_i$7 voxels. It applies four 3D-convolutional blocks with strided convolution and no pooling:

• Conv3D(filters = 48, kernel = $y_i$8, stride = 2) $y_i$9 ReLU
• Conv3D(filters = 48, kernel = $c$0, stride = 2) $c$1 ReLU
• Conv3D(filters = 192, kernel = $c$2, stride = 2) $c$3 ReLU
• Conv3D(filters = 128, kernel = $c$4, stride = 2) $c$5 ReLU

The feature map is flattened to a vector of length 128 000 and mapped by a dense layer to $c$6 outputs representing $c$7. The decoder mirrors this architecture: Dense $c$8 ReLU, reshape to $c$9, then four Conv3DTranspose blocks ending in SteepSigmoid$c$0. The regressor $c$1 takes the latent mean $c$2 and applies Dense(16)$c$3ReLU $c$4 Dense(16)$c$5ReLU $c$6 Dense($c$7)$c$8ReLU, where $c$9 for real foams and $f_\psi(\mu)$0 for the synthetic dataset (Nguyen et al., 23 Jul 2025).

PrIVAE is an encoder-decoder with two geometry-aware enhancements: a GNN layer over a Property Nearest-Neighbor Graph in the encoder and an isometric regularizer on the latent codes. Input $f_\psi(\mu)$1 is a length-$f_\psi(\mu)$2 sequence over an alphabet of size $f_\psi(\mu)$3. In the DNA nanocluster task, $f_\psi(\mu)$4 and one-hot encoding in $f_\psi(\mu)$5 is used. In the antimicrobial peptide task, pretrained residue embeddings (ESM-2) map each amino acid to a vector in $f_\psi(\mu)$6. Token embeddings are processed either by a bidirectional LSTM with two layers and hidden size $f_\psi(\mu)$7 per direction followed by a dense layer, or by multi-head self-attention layers with tuned head count. The resulting sequence representation $f_\psi(\mu)$8 is smoothed over the property graph by one or more GCN layers, after which a small MLP parameterizes the Gaussian posterior

$f_\psi(\mu)$9

The decoder is either an LSTM unrolled for $P^t$0 steps or an attention-based autoregressive network trained with tokenwise cross-entropy reconstruction loss (Sadeghi et al., 16 Sep 2025).

4. Property manifolds and latent geometry

PrIVAE makes the strongest geometric claim among the pVAE variants. Each sequence $P^t$1 is paired with a high-dimensional, experimentally measured property vector

$P^t$2

such as an emission spectrum described by up to four Gaussian-like peaks $P^t$3, or a set of log-MIC values against several bacteria. Although $P^t$4 lies in $P^t$5, feasible property profiles are assumed to lie on a lower-dimensional smooth manifold $P^t$6. To preserve the local geometry of $P^t$7, the method constructs a Property Nearest-Neighbor Graph (PNNG): pairwise property distances are computed, a $P^t$8-NN graph $P^t$9 is built, and the distances are converted into similarity weights with an RBF kernel

$c$00

With weighted adjacency $c$01, degree matrix $c$02, and combinatorial Laplacian $c$03, the graph supplies both the GCN propagation structure and the isometric penalty (Sadeghi et al., 16 Sep 2025).

The choice of property metric is domain-specific. For Ag$c$04–DNA emission spectra, each spectrum is fit as a mixture of Gaussians $c$05, and the closed-form Cauchy–Schwarz divergence is used. For antimicrobial peptide MIC profiles, the Manhattan $c$06 distance on the log-MIC vector is used. GCN smoothing then updates node features according to

$c$07

encouraging each sequence representation to look at other sequences with similar properties (Sadeghi et al., 16 Sep 2025).

The porous-metamaterials pVAE does not impose graph-isometric structure, but it also treats latent space as property-structured. After training, the latent means $c$08 form a structured, approximately Gaussian space. PCA on $c$09 reveals that the first principal components correlate strongly, with $c$10–0.7, with porosity $c$11 and permeability $c$12, and samples cluster by property such as compression level in real foams. This suggests a weaker but still explicit structure-property organization, achieved through supervised regression rather than graph Laplacian regularization (Nguyen et al., 23 Jul 2025).

5. Training protocols and property-guided generation

The conditioning-based pVAE is trained by stochastic variational Bayes under randomly sampled masks. On each SGD step, one samples minibatches of $c$13, draws masks $c$14, forms $c$15, infers $c$16, samples $c$17, and computes the reconstruction term only on indices with $c$18. To specialize the model to generate $c$19 from $c$20 alone, the masking distribution includes the all-missing case for $c$21. The practical significance is that a single model can support both full generation from properties and partial imputation under arbitrary conditioning patterns (Ivanov et al., 2018).

The porous-metamaterials pVAE is trained on two datasets: a synthetic academic dataset of 48 831 samples of size $c$22 voxels, generated by placing 10–40 nonoverlapping square pores of size $c$23 pixels in each 2D slice, and approximately 8965 subvolumes from $c$24CT scans of real open-cell foams at five compression levels. Intrinsic permeability is computed by LBM. A surrogate 3D-CNN is first trained to predict $c$25 on the synthetic dataset to accelerate property labeling as dataset size grows. Hyperparameters include latent dimension $c$26, optimized via Optuna, batch size 16, Adam optimization, pretraining learning rates $c$27 for the encoder-decoder and $c$28 for the regressor, followed by joint fine-tuning with lower learning rate or an adaptive scheduler, early stopping, learning-rate reduction on plateau, and model checkpoints (Nguyen et al., 23 Jul 2025).

In this regressor-augmented formulation, inverse design proceeds by solving

$c$29

by gradient descent in latent space. Initialization uses the nearest $c$30 in the training set, or many nearest neighbors to capture non-uniqueness, and the resulting $c$31 is decoded into a generated microstructure $c$32. Validation is performed with the surrogate CNN or direct LBM. The same trained model supports prior sampling $c$33, property-guided sampling near latent regions associated with a desired property band via KDE on $c$34-space, and spherical interpolation

$c$35

which yields smooth microstructure families with monotonically varying predicted $c$36 and $c$37 (Nguyen et al., 23 Jul 2025).

PrIVAE employs a different generation strategy. Graph $c$38 and all distance computations are fixed prior to training. Minibatches contain 32 “core” sequences plus their 1-hop neighbors so that the GCN can access full neighborhoods, and minibatches are preconstructed by grouping sequences with the same pseudo-labels to improve homogeneity. Training uses Adam with learning rate approximately $c$39, and models train for 2–5 hours on a Tesla V100. Hyperparameters $c$40, $c$41, neighborhood size $c$42, bandwidth $c$43, latent dimension $c$44, LSTM or attention sizes, and dropout are tuned via Optuna (TPESampler) on a multi-objective criterion balancing reconstruction accuracy against latent purity (Sadeghi et al., 16 Sep 2025).

6. Empirical behavior, case studies, and interpretation

PrIVAE evaluates latent-space quality with both reconstruction and neighborhood consistency. Reconstruction Accuracy per sequence is defined by

$c$45

where $c$46 is the Hamming distance. The reported performance is typically greater than $c$47 on both tasks. Purity$c$48 measures the Jaccard overlap between the pseudo-labels of a latent point’s $c$49 nearest neighbors in latent space and the point’s own pseudo-labels. On Ag$c$50–DNA, $c$51 on validation, compared to 0.40 for a vanilla VAE; on peptides, $c$52 versus 0.25 for the baseline. Three-dimensional PCA projections of latent means show semantically faithful geometry: for DNA spectra, single-peak clusters G, R, F, N occupy the corners of a tetrahedral pattern and dual-peak combinations lie between them; for peptides, single-species and multi-species activity nodes similarly interpolate (Sadeghi et al., 16 Sep 2025).

The Ag$c$53–DNA case study uses $c$54 sequences of length 10 with up to 4 emission peaks, Cauchy–Schwarz distance on Gaussian-mixture spectra, and a Bi-LSTM encoder plus GCN with latent dimension $c$55, $c$56, $c$57, and $c$58. The reported validation metrics are ACC $c$59 and Purity$c$60. For generative design, the method targets NIR emitters, which constitute only $c$61 of the training set, by locating the top-100 NIR-pure training codes, fitting group-specific Gaussians in latent space for labels $c$62, sampling 1 000 codes per group, decoding, filtering duplicates, and ranking by “NIR purity” in latent space. Wet-lab synthesis of the top 90 per group yields up to $c$63 truly NIR sequences, corresponding to 14.2-fold enrichment over a $c$64 baseline, and up to 16.1-fold enrichment when sampling from the NR region (Sadeghi et al., 16 Sep 2025).

The antimicrobial peptide case study uses 2 503 peptides of length 20 with ESM-2 embeddings and log-MIC values against E coli, S aureus, and P aeruginosa. Property distance is $c$65 distance on the 3-dimensional MIC vector. The model uses Bi-LSTM $c$66 GCN $c$67 latent dimension $c$68, with $c$69, $c$70, and $c$71. Validation metrics are ACC $c$72 and Purity$c$73. Design samples 1 000 sequences from each of 7 activity-profile regions, ranks by purity, selects the top 100, and evaluates them with a MIC-prediction oracle identified as the DBAASP model. Reported fully-active hit rates reach $c$74 for single-target groups, versus $c$75 for a vanilla VAE, while broad-spectrum success for the ESP group is $c$76 versus a $c$77 baseline (Sadeghi et al., 16 Sep 2025).

Ablation studies in PrIVAE separate the roles of graph smoothing and isometric regularization. GCN smoothing alone, denoted “VAE+GCN,” raises purity but slightly cuts reconstruction. The isometric regularizer alone, denoted “VAE+Reg,” helps alignment. PrIVAE, combining both mechanisms, achieves the best trade-off. In the metamaterials setting, the latent space also supports interpolation, clustering by property, and inverse design, but the mechanism is auxiliary supervision rather than graph-isometric preservation. The conditioning-based formulation, by contrast, emphasizes flexible probabilistic conditioning and one-shot generation from observed properties. Taken together, these results indicate that pVAE methods differ primarily in how property information enters the model: through conditional priors and decoders, through explicit regression on latent means, or through direct preservation of property-space geometry in latent space (Ivanov et al., 2018, Nguyen et al., 23 Jul 2025, Sadeghi et al., 16 Sep 2025).