Molecular representation shapes the balance between target fidelity and exploration in flow based polymer generation

Tianren Zhang

physics.chem-phSep 10, 2026

Molecular representation shapes the balance between target fidelity and exploration in flow based polymer generation

Authors: Tianren Zhang

Organizations: Department of Materials Science and Engineering, University of Delaware, Newark, Delaware 19716, United States

Abstract

Designing polymers with targeted properties requires navigating vast chemical spaces from limited labeled data. Here we introduce PolyLatentFlow, a framework based on continuous-time flow matching in latent space for unconditional and conditional polymer generation, together with LlamaUni, a multimodal representation combining polymer sequence and 3D structural information. In unconditional generation, PolyLatentFlow with LlamaUni produced the largest yield of valid candidates novel relative to PolyInfo among the evaluated unconditional generators while maintaining high diversity. For $T_g$ conditioning, generated property distributions shifted systematically across a 200 °C target range. In multi-property tasks, molecular representations showed similar surrogate target fidelity but differed markedly in validity, training-set replay, and structural proximity to labeled polymers. PolyLatentFlow with LlamaUni consistently combined high validity with low replay and achieved the largest per-attempt yield of nonreplayed target hits for CO $_2$ /N $_2$ conditioning. These results demonstrate latent space flow matching for polymer inverse design and identify molecular representation as a key determinant of target control and exploration beyond labeled chemistry.

Figures & tables

Figure 1: PolyLatentFlow workflow. (a) LlamaUni combines a frozen Llama 3 embedding of the pSMILES with a frozen Uni-Mol embedding of the corresponding capped polymer repeat unit dimer. A VAE compresses the fused representation into a latent representation with 512 dimensions and reconstructs the fused representation. A separately trained autoregressive decoder learns to generate pSMILES token by token from the latent representation. (b) Conditional flow matching transforms Gaussian noise at $t=0$ into a compact latent representation $z^{*}$ at $t=1$ conditioned on the requested property target. The frozen autoregressive decoder converts $z^{*}$ into a pSMILES string representing the generated polymer. The datasets used for property conditioning contain 6,668 polymers with $T_{g}$ measurements, 3,690 polymers with paired $T_{g}$ and $T_{d}$ measurements, and 623 polymers with paired CO 2 and N 2 permeability measurements.

Figure 2: LlamaUni produces the largest yield of valid polymers novel relative to PolyInfo while maintaining broad chemical space coverage. (a) Polymer validity. (b) Novelty relative to PolyInfo among unique valid polymer outputs. (c) Internal diversity calculated from pairwise Tanimoto distances using Morgan fingerprints. (d) Fraction of valid polymer outputs with an SA score below 5. (e) Fréchet ChemNet Distance (FCD) relative to PolyInfo. (f) FCD versus mean nearest neighbor similarity (SNN) to PolyInfo, comparing global distributional agreement with proximity to individual reference polymers. (g) UMAP projection in Morgan fingerprint space showing PolyInfo reference polymers and valid LlamaUni generations. Each model contributed 10,000 generation attempts.

Figure 3: Similar $T_{g}$ target responsiveness conceals representation dependent differences in exact-match replay and analogue proximity. (a) Predicted $T_{g}$ distributions of valid polymer outputs from 2,000 generation attempts per representation and requested target. Black horizontal marks indicate the means, and colored dashed lines indicate the requested targets. (b) Pooled $T_{g}$ hit rate within $\pm 25~^{\circ}\mathrm{C}$ versus exact-match replay rate among valid polymer outputs. Bubble area represents polymer validity across all generation attempts. (c) Cumulative distribution of Tanimoto distance to the nearest polymer in the labeled $T_{g}$ set. The vertical dashed line marks the analogue novelty threshold of 0.4. (d) Two representative valid polymer structures per representation at a requested $T_{g}$ of 250 ${}^{\circ}\mathrm{C}$ , selected from within the target window and with an SA score below 5, and annotated with predicted $T_{g}$ and Tanimoto distance to the nearest labeled polymer. Per target hit and replay rates are reported in Table S2 and Fig.S1.

Figure 4: Coupled property conditioning reveals a representation dependent balance between target fidelity and exact-match replay. (a) Joint $T_{g}$ and $T_{d}$ target hit rate heatmaps using a $\pm 25~^{\circ}\mathrm{C}$ window across the displayed $3\times 4$ grid of three $T_{g}$ targets and four $T_{d}$ targets. Cellwise hit rates were calculated among outputs for which both thermal surrogate models returned predictions. (b) Surrogate predicted CO 2 permeability and CO 2 /N 2 selectivity for valid polymer outputs with paired predictions. The dashed curve indicates the Robeson 2008 upper bound, $P_{\mathrm{CO_{2}}}=30{,}967{,}000\,\alpha_{\mathrm{CO_{2}/N_{2}}}^{-2.888}$ , with $P_{\mathrm{CO_{2}}}$ in Barrer. Legend percentages indicate the fraction of valid polymer outputs predicted to lie above the bound. (c) Joint $T_{g}$ and $T_{d}$ target hit rate versus exact-match replay rate $M$ , pooled across all 12 target combinations. (d) Joint CO 2 /N 2 target hit rate versus $M$ , pooled across all nine target combinations. A gas hit requires both predicted log permeabilities to fall within $\pm 0.5$ log units of their respective requested values. In (c,d), pooled rates use all valid polymer outputs, and bubble area represents polymer validity across all generation attempts. Per target combination and aggregate conditional metrics are reported in Table S2, and conditional polymer-validity rates for all three regimes are shown in Fig. S4 and Table S4.

Figure 5: Structural exploration and decoder capture. (a) Analogue novel fraction among nonreplayed valid polymer outputs in the CO 2 /N 2 task versus the yield of distinct analogue novel clusters per 1,000 generation attempts. Bubble area represents polymer validity across all attempts. (b) Probability that stochastic decoding returns a canonical pSMILES matching any polymer in the gas labeled corpus of 623 polymers after perturbing 96 labeled latent representations by the indicated radius. Perturbation radii were normalized by the representation specific median leave one out nearest neighbor spacing among labeled latents. Four stochastic decodes were performed for each source latent and perturbation radius. Shading denotes 95% normal intervals calculated from standard errors across source latents.

Explore similar work

Mar 27, 2026

Property-Guided Molecular Generation and Optimization via Latent Flows

Molecular discovery is increasingly framed as an inverse design problem: identifying molecular structures that satisfy desired property profiles under feasibility constraints. While recent generative models provide continuous latent representations of chemical space, targeted optimization within these representations often leads to degraded validity, loss of structural fidelity, or unstable behavior. We introduce MoltenFlow, a modular framework that combines property-organized latent representations with flow-matching generative priors and gradient-based guidance. This formulation supports both conditioned generation and local optimization within a single latent-space framework. We show that guided latent flows enable efficient multi-objective molecular optimization under fixed oracle budgets with controllable trade-offs, while a learned flow prior improves unconditional generation quality.

Alexander Arjun Lobo, Urvi Awasthi, Leonid Zhukov

BCG X AI Science Institute

Molecular Generation Molecular Optimization Latent Space Optimization Conditional Generative Modeling Flow-Based Generative Modeling

Ask PDF

Aug 2, 2026

PolymerGPT: Multi-property Optimization with a Decoder-Based GPT Model for Generative Polymer Design

Polymer property prediction and inverse generative design targeting desired properties are two crucial tasks in machine learning-assisted polymer design. While the former has received considerable attention, there have been limited methods developed for the latter. Existing methods focus on single-property optimization in the generative process, whereas accurate prediction of macroscopic material behavior requires simultaneous control of multiple physical properties. In this paper, we provide a transformative framework for direct optimization of a large collection of polymer properties. We propose PolymerGPT, a decoder-based GPT model that incorporates up to 37 commonly used polymer properties into the generative process via learned conditioning prefixes. It also supports a scaffold condition that specifies a desired scaffold for predicted structures. Our experimental results demonstrate that PolymerGPT achieves exceptional performance for unconditional and conditional generation while maintaining high validity, uniqueness, and novelty. Conditioning on five key properties yields generated structures whose predicted values closely match all target properties simultaneously.

Charlie Pyle, Adarsh Gadari, C. Adrian Figg +3

Texas A&M University · University of Pittsburgh · Department of Chemistry, Virginia Tech +2

Molecular Optimization Molecular Generation Conditional Generative Modeling

Ask PDF

Sep 8, 2026

Fixed-Dimensional Latent Flow for Generating Variable-Size 3D Molecules

Molecular size is coupled to composition, structure, and function, yet most 3D molecular generators require a predefined atom count. We introduce Equivariant-Free Transformer-Autoencoded Latent Flow Matching, a two-stage framework that samples a fixed-dimensional latent vector using flow matching and uses an autoregressive Transformer to determine molecular size, atom types, coordinates, and chemical attributes. Canonical atom ordering and rigid-pose alignment enable Transformers without equivariant layers, while decoded attributes guide bond reconstruction. On PCQM4Mv2, unconditional generation yields 87.9% unique, novel molecules passing sanitization and PoseBusters checks, exceeding baselines with lower end-to-end training and sampling time and higher end-to-end throughput. Across ten target HOMO-LUMO gaps, internal ranking retains 30% of screened candidates and increases the density functional theory-verified hit rate within 0.1 eV from 25.0% to 52.4%, while largely preserving novelty and diversity. These results demonstrate fixed-dimensional latent generation with autoregressive decoding as a practical approach to molecular design without prespecifying size.

Weichi Yao, Cameron Gruich, Bryan R. Goldsmith +1

Michigan Institute for Data & AI in Society University of Michigan, Ann Arbor, USA · Department of Chemical Engineering University of Michigan, Ann Arbor, USA · Department of Statistics University of Michigan, Ann Arbor, USA

3D Molecular Generation Molecular Generation Flow-Based Generative Modeling

Ask PDF