STARLING

Roughly a third of the eukaryotic proteome is intrinsically disordered, and for those regions the question "what is the structure?" has no answer. An intrinsically disordered region (IDR) is properly described by a conformational ensemble — a population of interconverting shapes whose statistics, not whose single lowest-energy pose, determine function. Structure predictors built for folded proteins, including AlphaFold 2, return one confident-looking model and are therefore the wrong instrument entirely. The alternative, all-atom or coarse-grained molecular dynamics, gives real ensembles but costs hours to days of compute per sequence, which makes proteome-scale questions impractical.

STARLING closes that gap. Developed in the Holehouse lab at Washington University in St. Louis by Borna Novak, Jeffrey Lotthammer, Ryan Emenecker, and Alex Holehouse, it is a generative model that emits full conformational ensembles for a disordered sequence in seconds rather than days. Its name expands to conSTruction of intrinsicAlly disoRdered proteins ensembles efficientLy vIa multi-dimeNsional Generative models. It was posted as a preprint in February 2025 and published in Nature in 2026.

The design is a deliberate hybrid: rather than learning from sparse experimental measurements, STARLING learns from a large corpus of physics-based simulations, inheriting the calibration of a force field parameterized against experimental data. It sits alongside a small family of sequence-to-ensemble models — IDPForge, IDP Gyration-Radius MLP, and the conditioned pLM for inverse design — but generates complete ensembles fast enough to sweep a proteome.

#Key Features

• Ensembles, not single structures: The output is a population of conformers with the distributional properties that matter for disordered biology — radius of gyration, end-to-end distance, contact maps — rather than one static pose.
• Latent diffusion over distance maps: A variational autoencoder compresses inter-residue distance maps into a small latent space, and the diffusion model denoises inside that space, which is what makes generation cheap enough to run at scale.
• Runs on commodity hardware: Inference works on CUDA GPUs, Apple Silicon, and plain CPUs, so ensemble generation does not require a cluster allocation.
• Ensemble-aware sequence embeddings: Beyond conformers, the jointly trained encoder yields sequence representations that carry ensemble information, usable as features for downstream property models.
• Single pip install with bundled weights: The idptools-starling package ships trained checkpoints, so a sequence goes to an ensemble in one command with no separate weight download or configuration.

#Technical Details

STARLING is a latent-space denoising diffusion probabilistic model. A variational autoencoder is first trained to compress inter-residue distance maps into a 24 by 24 latent representation; the diffusion model then learns to denoise within that compressed space, and the VAE decoder reconstructs a distance map from each generated latent, from which three-dimensional coordinates are recovered. Working in latent space rather than on full distance matrices is the central efficiency decision, and it is why hundreds of conformers can be produced in seconds.

Training, validation, and test data came from coarse-grained molecular dynamics simulations run with the Mpipi-GG force field, covering sequences between 10 and 384 residues. The held-out test set comprised 10,437 sequences totalling 1,425,849 conformations. Accuracy was assessed against experimental data the model never saw: across 133 IDR sequences with published small-angle X-ray scattering measurements, STARLING-derived mean radii of gyration agree closely with the SAXS-derived values, and its ensembles agree with independently generated all-atom ensembles. Because the training simulations span 10 to 384 residues, the released model supports sequences up to roughly 380 residues; longer IDRs must be handled in fragments.

#Applications

The practical unlock is scale. A researcher studying one disordered tail can generate an ensemble and read off compaction, transient contacts, or accessibility in the time it takes to run a BLAST search, and a group studying a whole proteome can do the same for every annotated IDR. That supports hypothesis generation — which regions are unusually compact, which mutations reshape a chain, which ensembles are consistent with an observed FRET efficiency — and it supports interpretation, letting experimentalists compare a measured observable against a predicted ensemble instead of a guessed structure.

#Impact

STARLING's contribution is less a new architecture than a change in what is affordable. By distilling force-field simulations into a generative model, it converts an expensive physics calculation into a cheap inference call while keeping the physics calibration, and it does so with a distribution path — a pip package with bundled weights and hosted documentation — that puts it in reach of wet-lab groups with no simulation infrastructure. The peer-reviewed Nature publication and the LGPL-3.0 code release make it a credible default for sequence-to-ensemble work, though the article itself is CC BY-NC-ND 4.0, not open access. The honest limits are the training distribution: ensembles reflect the Mpipi-GG force field and will inherit its biases, sequences beyond roughly 380 residues are outside the trained range, and the model describes single chains rather than multi-chain assemblies or condensates. The Mpipi-GG distance-map corpus is not deposited, so the shipped weights can be used but not retrained.