Source-linked AI summary
Molecular representation shapes the balance between target fidelity and exploration in flow based polymer generation
Tianren Zhang
TL;DR
Polymer inverse design must navigate vast chemical spaces with relatively few labeled measurements, while molecular representation may affect both target control and structural exploration. This paper introduces PolyLatentFlow, combining latent-space flow matching with the multimodal LlamaUni representation. LlamaUni achieved strong validity and exploration while maintaining responsive conditioning, highlighting representation as a determinant of generation outcomes.
Problem
Polymer property datasets contain only thousands of labeled examples, and the effect of molecular representation on target control and structural exploration is difficult to isolate from other design choices.
Source: §1 · pp. 1–2Method
PolyLatentFlow uses continuous-time flow matching in a compact latent space, with LlamaUni combining pSMILES sequence and capped-dimer structural embeddings for polymer generation.
Source: §1 · p. 2, §2.1 · p. 3Results
Across unconditional and conditional generation, LlamaUni combined high validity, broad exploration, and low training-set replay, producing the largest yield of valid PolyInfo-novel unconditional candidates.
Source: §1 · p. 2, §3 · p. 10Takeaways & Limitations
Molecular representation is a key determinant of target control and exploration beyond labeled chemistry, and additional labeled data do not necessarily raise nonreplayed target-hit yield.
Source: §3 · p. 11Property attainment was evaluated with surrogate predictions rather than independent prospective validation; simulation, synthesis, and experimental characterization remain needed.
Source: §3 · p. 11, §4.5 · p. 13
Abstract
from arXiv · showhide
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 {\deg}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.
1 Introduction
PolyLatentFlow pairs latent-space flow matching with LlamaUni to generate polymers, while testing how molecular representation shapes target control and exploration beyond labeled chemistry.
1 Introduction
- Property conditioning is learned from datasets of only thousands of labeled polymers, with fewer examples available for jointly specified properties.Large virtual corpora broaden structural coverage but consist primarily of hypothetical polymers rather than experimentally characterized materials.Source: §1 · p. 1
- The effect of representation on generation remains difficult to isolate because prior studies often vary representation alongside architecture, training, and decoder design.The motivating trade-off is that compact representations may aid conditioning but erase structural distinctions, whereas richer ones may support exploration at the expense of control.Source: §1 · p. 2
- The framework enables a controlled assessment of representation by comparing LlamaUni with Morgan fingerprints, PolyBERT, and pSMILES-VAE under a consistent generative and evaluation setup.LlamaUni fuses pSMILES sequence embeddings with embeddings of capped repeat-unit dimers, then reconstructs pSMILES from compact latents.Source: §1 · p. 2
- PolyLatentFlow with LlamaUni combined high validity, broad exploration, and low training-set reuse across unconditional and conditional generation.It produced the largest yield of valid, PolyInfo-novel unconditional candidates among evaluated generators.Source: §1 · p. 2
- Figure 1 depicts LlamaUni encoding and compression followed by latent flow matching from Gaussian noise and autoregressive decoding to pSMILES.The conditioning datasets contain 6,668 Tg polymers, 3,690 paired Tg/Td polymers, and 623 paired CO2/N2 permeability polymers.Source: Figure 1 · p. 3
2 Results
PolyLatentFlow applies flow matching to compact polymer latents, with LlamaUni combining sequence and structural information. Across generation tasks, LlamaUni paired strong validity and target responsiveness with low replay and broad structural exploration.
2.1 PolyLatentFlow pipeline
- PolyLatentFlow separates molecular representation, latent generation, and reconstruction; LlamaUni fuses Llama-3 sequence and UniMol structural embeddings into a 512-dimensional latent space.A frozen decoder reconstructs pSMILES, and flow matching learns unconditional or property-conditioned latent distributions.Source: §2.1 · pp. 3–4
2.2 Unconditional generation
- 9,491 valid PolyInfo-novel outputs made LlamaUni the top unconditional yield; validity was 95.8%, 9.1% above CharRNN.Internal diversity remained 0.83–0.88 across PolyLatentFlow configurations; LlamaUni outputs also extended into peripheral regions of the PolyInfo manifold.Source: §2.2 · pp. 4–5, Figure 2 · p. 5
2.3 Conditional generation for glass transition temperature
- Across Tg targets, predicted distributions shifted upward monotonically, while LlamaUni paired 32.4% hit rate with 1.2% replay and 93.0% validity.69.5% of LlamaUni valid outputs exceeded the 0.4 Tanimoto-distance threshold from the labeled Tg set, compared with approximately half for Morgan and PolyBERT.Source: §2.3 · pp. 5–7
2.4 Conditional generation for joint objectives
- Joint Tg–Td hits varied across target pairs and were lower at the highest Td values; CO2/N2 upper-bound exceedances were rare and based on surrogate predictions.The authors relate higher-Td difficulty to limited labeled-data support and describe above-bound candidates as requiring higher-fidelity evaluation.Source: §2.4 · p. 7
- LlamaUni achieved the largest per-attempt yield of valid, nonreplayed CO2/N2 target hits, while no representation dominated every conditioning task.Across tasks, it consistently combined low replay, high validity, and competitive target attainment.Source: §2.4 · pp. 7–9
2.5 Structural exploration and the latent origins of replay
- In CO2/N2, 84.3% of LlamaUni nonreplayed valid outputs exceeded the analogue-distance threshold, and LlamaUni yielded the most distinct analogue-novel clusters per attempt.Perturbation tests found rare exact labeled-polymer returns from LlamaUni decoding, providing a plausible decoder contribution to its lower replay.Source: §2.5 · pp. 9–10
3 Conclusion
PolyLatentFlow with LlamaUni combines high validity and low replay with broad structural exploration, while molecular representation shapes the balance between target control and exploration beyond labeled chemistry. The authors conclude that generated candidates still require prospective validation.
3 Conclusion
- LlamaUni produced the largest unconditional yield of valid structures novel relative to PolyInfo and, in CO2/N2 conditioning, the largest per-attempt yield of valid, nonreplayed target hits.Its low exact replay was accompanied by broader structural exploration, rather than only small modifications of labeled polymers.Source: §3 · p. 10
- Validity, replay, structural distance, and target attainment provide complementary evidence, and their balance should reflect whether the goal is analogue optimization or broader exploration.The data-scaling results show that more labeled data do not necessarily increase nonreplayed target-hit yield when conditioning improves alongside greater training-set reuse.Source: §3 · pp. 10–11
- Property attainment was evaluated with surrogate models, so prospective simulation, synthesis, and experimental characterization remain necessary to validate generated candidates.Reaction-aware assessment is also needed to establish practical synthesis routes.Source: §3 · p. 11
4 Methods
The methods combine polymer datasets, molecular representations, latent flow models, and surrogate-based evaluation to generate and assess polymers under unconditional and property-conditioned settings.
4.1 Data
- PI1M supplied unlabeled pSMILES for representation compression, unconditional flow training, and decoder training; PolyInfo and three labeled datasets supplied reference and property data.The Tg, paired Tg–Td, and paired CO2/N2 datasets contained 6,668, 3,690, and 623 polymers, respectively.Source: §4.1 · p. 11
4.2 Molecular representations and compact latents
- LlamaUni concatenates frozen Llama 3 pSMILES embeddings and UniMol embeddings of capped repeat-unit dimers, while Morgan, PolyBERT, and pSMILES-VAE provide comparators.Separate VAEs compressed molecular representations to 512-dimensional latents; the pSMILES-VAE instead encoded tokenized sequences with a bidirectional GRU.Source: §4.2 · p. 11
4.3 Flow matching and decoding
- Flow matching trained latent models along straight-line paths from Gaussian noise to data latents, and latent-conditioned GRU decoders reconstructed pSMILES.Polymer validity required RDKit parsing and exactly two repeat-unit attachment atoms.Source: §4.3 · p. 12
4.4 Conditional training and generation
- Conditional flows were initialized from unconditional checkpoints, or from Tg-conditioned checkpoints for joint Tg–Td tasks, and sampling balanced property quantiles.The study sampled 2,000 attempts per representation and target for Tg, Tg–Td, and CO2/N2 tasks, with classifier-free guidance scales of 1.5 for thermal and 3.0 for gas generation.Source: §4.4 · p. 12
4.5 Surrogate property prediction
- Property attainment relied on PolyLLMem surrogate predictors, and overlapping property compilations made target hit rates internal evaluations rather than independent prospective validation.For Tg, PolyLLMem and an independently trained Morgan random forest had Pearson r = 0.897; for Td, they had r = 0.725, with joint hit-rate differences no greater than 0.91 percentage points.Source: §4.5 · p. 13
4.6 Evaluation metrics
- Evaluation measured unconditional validity, uniqueness, novelty, diversity, distributional agreement, and similarity, alongside conditional replay, structural distance, target hits, and per-attempt yields.Gas targets required both predicted permeabilities to fall within 0.5 log10(P/Barrer) of their requested values; synthetic accessibility was treated as a heuristic, not proof of synthesizability.Source: §4.6 · pp. 13–14
4.7 Latent analyses and statistics
- Latent analyses tested decoder capture by perturbing labeled gas-polymer latents and assessed data availability using stratified subsets of 100, 250, and 500 polymers against the full 623-polymer dataset.The capture experiment used 96 labeled latents, with four stochastic decodes per perturbation radius.Source: §4.7 · p. 14