Protein structure predictors such as AlphaFold 3 now score near the ceiling on static benchmarks. Yet when a protein has two real shapes, they mostly predict the one they saw most often during training. A recent multi-state benchmark found that when a protein changes shape on binding a small molecule, over 80% of failures are "apo → holo" collapses: given the unbound (apo) input, the model returns the ligand-bound (holo) shape. A cheap, crystal-free experiment exists that could correct this. In cross-linking mass spectrometry (XL-MS), a chemical "molecular ruler" is added to the protein in solution, and pairs of residues that end up chemically bonded must lie within roughly 20–25 Å of each other in whatever shape the protein actually adopts. Ten years ago, a blind test showed that feeding such data to the prediction methods of the time did not help and sometimes hurt, because the data are sparse, unevenly spread and 10–20% wrong.
This proposal asks whether a modern diffusion-based structure model, meaning a generator that builds coordinates by gradually denoising random noise, can be steered at inference time by noisy cross-links into the conformation that is actually present in solution. It borrows a lesson from the code-benchmark literature: claimed gains must be measured under realistic data conditions, not idealized ones. The core contribution is twofold. The first part is a restraint likelihood that explicitly models false cross-links, applied as guidance during denoising. The second is an evaluation protocol that reports how much of any gain survives when clean simulated cross-links are replaced by realistically corrupted ones.
Core Claim
We claim that applying cross-link guidance inside the diffusion trajectory, using a likelihood that treats a known fraction of links as outliers, recovers alternative conformations that unguided sampling and post-hoc reranking cannot reach. The claim is expected to hold specifically for ligand-induced and protein-induced transitions. In those cases the correct state is almost never sampled, so filtering the samples afterward has nothing to select. We further claim that this robust likelihood reverses the failure seen in the 2016 blind test: denser but noisier cross-link sets should help rather than hurt. This mirrors the purified-albumin finding that thousands of links at high false-discovery rate (FDR) gave better models than a few hundred clean ones. We expect a sharp split. Reranking should roughly match guidance on intrinsic transitions, where the alternative state appears in some samples. Guidance should clearly win on ligand-induced transitions. If that split does not appear, the claim is wrong.
Where This Sits
The motivating evidence comes from three recent benchmarks. ProMiSE evaluates AlphaFold3, Boltz, Chai-1 and BioEmu on 577 conformational pairs and makes three findings:
- the ligand-induced "holo magnet";
- a "structure module bottleneck", in which internal predicted distance distributions keep both states but the final coordinates flatten to one;
- unphysical intermediate structures for protein-induced changes.
It only diagnoses these failures and does not try to correct them; our guidance is designed to act exactly where the bottleneck occurs. DisProtBench shows that the model's own confidence score (pLDDT) correlates at ρ = −0.48 with true disorder. It therefore constrains our design: we cannot use model confidence to decide which predictions to trust and must use external physical evidence instead. Density-guided AlphaFold supplies the closest method. It conditions AlphaFold 3 on X-ray electron density and recovers unmodeled β-turns. Its scope is limited to four-residue motifs and high-resolution crystals, whereas we target whole-domain rearrangements using measurements taken in solution without crystals.
The XL-MS side contributes both a cautionary result and a supporting one. The CASP11 blind tests (Belsom et al., CASP11; Blind Evaluation of Hybrid Methods) supplied 201–381 cross-links per target at 5% FDR. They found no significant improvement in GDT_TS, a standard 0–100 measure of structural similarity, and sometimes a decrease, which they attributed to uneven coverage and false links. This is the negative result our likelihood must overcome. The serum albumin study supplies the positive counterpoint. With sulfo-SDA cross-links fed to Rosetta, it modeled albumin domains even inside crude blood serum, and denser noisy link sets helped. Our approach differs from both because the conformational prior is a learned diffusion model rather than fragment assembly.
The benchmark-saturation literature serves as a methodological lens rather than a mechanism. FeatureBench shows coding agents dropping from over 70% on localized bug fixes to about 11% on integrated feature work. That drop arose because the earlier benchmark had removed the dependencies that make real tasks hard. Simulated cross-links taken from a known structure have the same flaw: they come with uniform coverage and no false positives. We therefore treat the gap between clean and realistic restraints as a first-class result rather than hiding it.
Research Program
The method adds a guidance term to the reverse diffusion of an open AlphaFold-3-class model (Boltz-1/Boltz-2 or Chai-1, chosen because their weights can be modified). Each cross-link $k$ joins residues $i$ and $j$ with a maximum span $d_{\max}$ of about 25 Å for sulfo-SDA. Let $d_{ij}(x)$ be the distance between the two residues in the current denoised coordinates $x$. We assign each link the likelihood
$$ p_k(x) = (1-\epsilon)\,\sigma\!\left(\frac{d_{\max}-d_{ij}(x)}{\tau}\right) + \epsilon , $$
where $\sigma$ is the logistic function, $\tau$ sets the softness of the distance bound, and $\epsilon$ is the expected false-link fraction (initialized from the reported FDR).
At each denoising step $t$, the model's predicted clean structure is nudged by
$$ \lambda_t \nabla_x \sum_k \log p_k(x). $$
Here $\lambda_t$ is a guidance strength that increases as noise falls. Because of the $\epsilon$ floor, a link that cannot be satisfied contributes a bounded, nearly flat penalty rather than an unbounded pull. This is how the method is meant to tolerate the 10–20% FDR that broke the CASP11 methods.
We also test a variant that applies the same likelihood to reweight the model's internal distance distribution (its distogram) before the structure module. ProMiSE located the collapse at that module, so injecting evidence before it may matter. The variants are compared directly. Every guided sample is relaxed and then scored by how well it satisfies a held-out half of the cross-links. These held-out links are never used for guidance, so they give an honest selection criterion.
The evaluation protocol is the second contribution. For every ProMiSE pair we generate cross-link sets from the target state in two forms:
- Clean: all residue pairs within $d_{\max}$, subsampled at random.
- Realistic: coverage biased toward reactive, solvent-exposed residues; ambiguous residue assignments; 5%, 10% and 20% false links drawn from the chemically reachable but spatially distant pairs; and densities spanning the CASP11 range up to the 881–1495 links of the albumin dataset.
We report the drop from clean to realistic for every method. Real cross-link data provide the external check: the albumin dataset (PXD001692, ground truth PDB 1A06, both purified and in-serum) and the four CASP11 targets Tx781, Tx808, Tx767 and Tx812.
Baselines
The naive reference is unguided multi-seed sampling from the same model at the same total number of denoising steps. ProMiSE reports that standard models sample all known states of intrinsically multi-state proteins only 8%–24% of the time, and that over 80% of ligand-induced failures are apo→holo collapses. Beating it means a clearly higher state-recovery rate on the same clusters at matched compute.
The strongest learned competitor for conformational diversity is BioEmu, whose structure module was fine-tuned on molecular-dynamics trajectories. ProMiSE reports a 29% sampling success for it on intrinsic transitions. We run it with the same sample budget and the same held-out-link selection. To beat it, unguided Boltz plus guidance must exceed 29% on intrinsic clusters and must also improve ligand-induced recovery, where BioEmu has no external evidence to act on.
The strongest restraint-aware competitor is Rosetta model-based search with cross-link restraints, the pipeline used in the albumin and CASP11 work. On albumin it reached first-selected RMSDs of 2.8/5.6/2.9 Å (domains A/B/C) purified and 4.5/5.9/4.8 Å in serum (serum albumin study). Lower per-domain RMSD on the same link sets at the same FDR thresholds counts as beating it. For the CASP11 targets the file records only "no significant improvement" in GDT_TS and gives no numbers, so we rerun that pipeline ourselves under the same restraint files.
Two ablations isolate the key ingredients. One sets $\epsilon = 0$, giving a plain harmonic-style bound that should reproduce CASP11-style brittleness as FDR rises. The other replaces guidance with reranking, as described next. Every method receives the same restraint files, the same per-target GPU budget, and the same small tuning budget for $\lambda_t$ and $\tau$. Tuning uses a disjoint set of ProMiSE clusters.
Why not just…?
Why not just oversample seeds and keep the samples that best satisfy the cross-links? This is the most serious threat, and we run it at matched compute. Reranking can only select what has been sampled. ProMiSE's figure of over 80% apo→holo collapse on ligand-induced pairs suggests the correct state is rarely in the sample pool. On intrinsic transitions, however, states appear 8%–24% of the time, so reranking may come close to guidance there. We predict exactly this split. If reranking matches guidance on ligand-induced pairs as well, the guidance contribution is unnecessary, and the paper reduces to the realistic-noise evaluation protocol.
Why not just fine-tune the model to take cross-links as an input feature? Fine-tuning needs paired structure and cross-link training data. Real paired data are scarce, so training would rely on simulated, clean links, which is the very idealization our clean-versus-realistic comparison is designed to expose. Guidance requires no retraining and lets us change $\epsilon$ at test time to match each experiment's FDR. We treat fine-tuning as follow-up work, justified only if guidance shows the signal is usable at all.
Why not just run molecular dynamics, or use BioEmu? MD fine-tuning gives BioEmu its 29%, but this addresses the model's prior and not the experimental evidence. Force-field relaxation alone did not discover the backbone flips that density guidance found in Density-guided AlphaFold. We include BioEmu both as a baseline and as an alternative prior that can itself be guided. If guided BioEmu outperforms guided Boltz, that changes which prior we recommend but does not undermine the claim.
Why not just trust the model's confidence to flag collapsed predictions? DisProtBench shows pLDDT is anti-correlated with true disorder (ρ = −0.48), so it cannot identify cases that are confidently wrong. For this reason held-out cross-link satisfaction, not pLDDT, serves as our selection score.
Decisive Checks
The primary metric is state-recovery rate on ProMiSE, broken down by transition type and FDR level. A sample counts as a success if its TM-score (a 0–1 measure of fold similarity) to the target state is at least the benchmark's threshold. Secondary metrics are per-domain RMSD on albumin, GDT_TS on the CASP11 targets, and stereochemical clash counts.
Three results would settle the proposal:
- Guided sampling must beat reranking on ligand-induced pairs at matched compute.
- Under the robust likelihood, accuracy must not fall as link density rises from CASP11-like to albumin-like levels at 10–20% FDR, while the $\epsilon = 0$ ablation does fall.
- In-serum albumin RMSD must improve on the Rosetta numbers above.
Alongside these, we publish the clean-to-realistic drop for every method. That gap, analogous to FeatureBench's 70%→11% cliff, is the honest measure of how far restraint-guided structure prediction remains from real experiments.
References
- ProMiSE: Protein Multi-State Evaluation Benchmark in Biological Contexts
- DISPROTBENCH: A Disorder-Aware, Task-Rich Benchmark for Evaluating Protein Structure Prediction in Realistic Biological Contexts
- Density-guided AlphaFold reveals unmodeled alternative turn conformations in protein structures
- Blind testing of cross‐linking/mass spectrometry hybrid methods in CASP11
- Blind Evaluation of Hybrid Protein Structure Analysis Methods based on Cross-Linking
- Serum Albumin Domain Structures in Human Blood Serum by Mass Spectrometry and Computational Biology
- FeatureBench: Benchmarking Agentic Coding for Complex Feature Development