GNMCADS

Sample 54 class A GPCRs a thousand times each with the diffusion module of AlphaFold3 and 81.1% of the 54,000 structures come back closed, against 3.8% open. That is a faithful answer to the question the model was trained on — the single most probable structure — and the wrong answer for receptors whose biology is the transition between states. The pairwise conditioning signal the trunk hands to the diffusion module determines the output so strongly that stochastic noise in the sampler cannot move it out of one basin.

GNMCADS (Gaussian Network Model Guided Condition Annealed Diffusion Sampler) intervenes on that conditioning signal. Condition annealing — progressively noising a diffusion model's conditioning during denoising — is an established diversity technique for image models, but applied wholesale to a protein diffusion model it degrades secondary structure and returns unusable geometry. GNMCADS makes the annealing selective, and picks where to apply it from physics rather than at random. It first generates a reference structure by default sampling, then builds a Gaussian Network Model (GNM) of it: an elastic network in which Cα atoms within a cutoff distance are joined by uniform harmonic springs, whose low-frequency normal modes describe the collective motions the fold is intrinsically built to make. Cross-correlations of residue fluctuations within one mode partition the protein into two dynamical domains — residues that move together, and residues that move against them. Only the inter-domain entries of the conditioning signal are annealed: the constraints holding each dynamical domain rigid survive, while the constraint fixing their relative arrangement is released.

GNMCADS was developed by Ahmed Selim Uzum and Turkan Haliloglu at Bogazici University and posted to bioRxiv in August 2026. Nothing is trained or fine-tuned: the GNM calculation is analytical and the diffusion backbone is frozen, so the method is a sampling-time strategy over an existing checkpoint — the same shape as AlphaFlex, which steers AlphaFold 2 by masking MSA columns. It is a preprint awaiting peer review, and the announced code repository is a placeholder at the time of writing.

#Key Features

• Physics-selected annealing targets: GNM cross-correlations, not a heuristic or a learned predictor, decide which residue pairs have their conditioning relaxed.
• Two composed masks: a diagonal mask spares the main diagonal and the nearest superdiagonals, where perturbation is known to wreck secondary structure; a GNM mask selects only anti-correlated inter-domain pairs. Their product is the annealing mask.
• Mode sampling per diffusion step: at each step the active GNM mode is drawn from a categorical distribution over the softmax of negative eigenvalues, so low-frequency functional modes dominate but the partition still varies across the trajectory.
• Channel-matched noise: injected noise is rescaled to the mean and standard deviation of the selected elements in each channel of the pairwise representation, keeping the perturbation in range — a departure from generic condition-annealed sampling.
• No retraining and no new weights: the method modifies inference on a frozen structure-prediction backbone and introduces no learned parameters of its own.

#Technical Details

The annealing schedule is linear in diffusion time, holding the conditioning intact until a first threshold, ramping to fully annealed at a second, and staying there. The authors implement this inside AlphaFold3's diffusion module and call the result AF3 GNMCADS, benchmarking it on 92 proteins: 54 class A GPCRs curated from GPCRdb with both G protein-bound and inactive states resolved, 15 transporters with experimental inward- and outward-facing structures, and 23 proteins with major domain motions. Every method generated 1,000 structures per target; AF3 GNMCADS used a single seed with 1,000 diffusion samples, against stock AlphaFold3's 200 seeds at 5 samples each.

On the GPCR cohort the closed/intermediate/open split moves from 81.1/15.1/3.8% for AlphaFold3 to 53.9/31.0/15.1% for AF3 GNMCADS. The other two datasets are scored by fill ratio, which measures how contiguously a method samples the path between two known conformations; AF3 GNMCADS improves it by more than 0.05 on 35 of 38 targets relative to AlphaFold3, and on a majority of targets against MSA-randomization samplers, latent-steering samplers, and the MD-emulation modes of Boltz-2 and BioEmu-1. Projected onto molecular dynamics-derived free energy landscapes, ensembles for calmodulin, adenylate kinase, the AT1a angiotensin II receptor and the AtSWEET13 sugar transporter populate extended, compact and intermediate states rather than clustering at one minimum.

#Applications

The target cases are those where a single predicted structure misleads: alternating-access cycles in membrane transporters, the activation pathway of class A GPCRs, and interdomain hinge motions in enzymes. Because the intervention is on internal representations rather than the multiple sequence alignment, it does not depend on MSA depth or heterogeneity, which makes it applicable to targets with sparse evolutionary signal — orphan sequences and de novo designed proteins — where MSA-perturbation methods have little to work with. Intermediate states matter to structure-based drug design, since transient pockets that open only between endpoints are invisible to single-state prediction.

#Impact

GNMCADS is an argument that the elastic network literature still has something to tell a diffusion model. The Gaussian Network Model dates to 1997 and predicts equilibrium fluctuations from geometry alone; used here it supplies the one thing a frozen structure predictor lacks, namely which parts of the protein are meant to move relative to which. The limits are the scope of the evaluation and the state of the release: results rest on 92 targets with experimentally resolved endpoint pairs, diversity is assessed against known conformations rather than by blind discovery, and no ensemble has been validated experimentally. The source code announced in the paper had not been published at the time of writing, so the results cannot yet be reproduced independently.