Structure-conditioned masked language models for protein sequence design generalize beyond the native sequence space

Akpinaroglu, D.; Seki, K.; Guo, A.; Zhu, E.; Kelly, M. J.; Kortemme, T.

doi:10.1101/2023.12.15.571823

Home/Archive/bioRxiv

bioRxiv · 10.1101/2023.12.15.571823

Structure-conditioned masked language models for protein sequence design generalize beyond the native sequence space

Akpinaroglu, D.·Seki, K.·Guo, A.·Zhu, E.·Kelly, M. J.·Kortemme, T.

View original source ↗Export metadata ↓

Abstract

Machine learning has revolutionized computational protein design, enabling significant progress in protein backbone generation and sequence design. Here, we introduce Frame2seq, a structure-conditioned masked language model for protein sequence design. Frame2seq generates sequences in a single pass, achieves 49.1% sequence recovery on the CATH 4.2 test dataset, and accurately estimates the error in its own predictions, outperforming the autoregressive ProteinMPNN model with over six times faster inference. To probe the ability of Frame2seq to generate novel designs beyond the native-like sequence space it was trained on, we experimentally test 26 Frame2seq designs for de novo backbones with low identity to the starting sequences. We show that Frame2seq successfully designs soluble (22/26), monomeric, folded, and stable proteins (17/26), including a design with 0% sequence identity to native. The speed and accuracy of Frame2seq will accelerate exploration of novel sequence space across diverse design tasks, including challenging applications such as multi-objective optimization.

Explore related subjects

bioengineering Preprint

Keep this discovery

Explore connections, maps & timelines

BibTeX RIS

Akpinaroglu, D., Seki, K., Guo, A., Zhu, E., Kelly, M. J., Kortemme, T.. 2023-12-15. Structure-conditioned masked language models for protein sequence design generalize beyond the native sequence space. https://doi.org/10.1101/2023.12.15.571823

Cite the original work for its findings. Save a collection to share your selection of sources.

KEEP EXPLORING

Related preprints

bioRxiv2026

Synergy-informed bilevel optimization reveals joint-specific benefits for realistic exoskeleton torque prediction

Simulation-driven optimization can guide exoskeleton control design, but its utility is limited by the accuracy with which muscle coordination is predicted. Strict effort-minimization formulations predict non-physiological muscle shutdown under high exoskeleton assistance, leading to peak assistive torques that exceed values reported in human-in-the-loop experiments. We evaluated the accuracy of synergy-informed simulations at predicting assistance across four single-joint exoskeleton configurations: ankle plantarflexion, knee extension, hip flexion, and hip abduction. We implemented a bilevel optimization to predict a parameterized assistive torque that minimizes metabolic cost, with assisted muscle dynamics estimated via either effort-minimization simulations (minimizing muscle effort only) or synergy-informed simulations (minimizing muscle effort constrained to a synergy structure). We used experimental data from five unimpaired adults walking at their preferred speed and compared results descriptively between the two formulations within each subject. Assistance based on effort-minimization simulations shut down agonist muscle activations and overestimated peak assistive torque compared with experimental values. In contrast, constraining muscle controls to a synergy structure prevented non-physiological muscle shutdown and yielded assistive torques closer to those measured in human-in-the-loop experiments. This benefit was not observed equally across all assisted degrees of freedom or subjects, emerging only where no single synergy could be suppressed to unload the assisted muscles. Synergy weights can therefore be interpreted as joint-specific predictors of when simulations can reliably inform wearable robot design.

bioengineering↗

bioRxiv2026

Engineered microvessel networks -- Longitudinal transcriptomic profiling validates chip-grown tissues as a robust model for studying vasculature

Vascularised Organ-on-Chip (OoC) models are promising New Approach Methodologies (NAMs), but their translation requires validation for defined Contexts of Use (CoU). Here, we combine microfluidic microvessel networks (MVNs) with longitudinal bulk and single-cell RNA sequencing to determine how cell provenance influences engineered vascular phenotype. Stromal cell identity produced distinct network architectures and transcriptional trajectories, demonstrating that support cells, commonly treated as interchangeable, represent an important CoU variable. Primary human pancreatic microvascular endothelial cells failed to form stable perfusable networks, but supplementation with 10% Human Umbilical Vein Endothelial Cells (HUVECs) or Blood Outgrowth Endothelial Cells (BOECs) rescued self-assembly. Single-cell transcriptomics identified specialised endothelial populations associated with vascular morphogenesis, including tip, stalk, phalanx, endothelial to mesenchymal transition (EndMT) and pericyte-like endothelial cells. Failed networks lacked coherent endothelial specialisation, whereas rescued networks restored stabilising cell states and revealed FLRT2--UNC5B signalling within pericyte-like HUVECs, recapitulating a pathway implicated in angiogenic development in vivo. These findings establish cell provenance as a critical determinant of microvascular NAM performance and demonstrate how transcriptomic validation can define and strengthen the biological relevance of microvessel-on-chip models for specific CoUs.

bioengineering↗

bioRxiv2026

Benchmarking foundation models for resting-state functional magnetic resonance imaging time series in cognitive aging research

Foundation models trained on large neuroimaging datasets could help smaller studies investigate cognition from brain recordings. However, whether the foundation model-derived representations improve assessment of current cognition, prediction of follow-up cognition, or characterization of cognitive change remains unclear. We evaluated nine configurations from five model families that process regional resting-state functional magnetic resonance imaging time series, using 3,103 scans from 1,653 participants across two observational and five intervention aging cohorts. Primary frozen representations showed no consistent superiority over conventional functional connectivity for concurrent cognitive associations or diagnostic classification. Changing the readout improved selected concurrent cognitive associations without updating model weights, indicating that performance depended partly on how representations were extracted. Additional post-training and embedding alignment yielded gains in selected settings but did not consistently improve performance across outcomes. In the primary longitudinal analyses, imaging representations did not establish added value beyond baseline clinical measures. These findings show that readout choice influences concurrent cognitive associations, while the added value of these representations for longitudinal outcomes remains unestablished.

bioengineering↗