Biasing Conformational Sampling in AlphaFold 3 and Boltz-2 via Pair Representation Scaling
Abstract
Deep learning has transformed protein structure prediction, yet most systems return a single dominant conformation with little control over alternative functional states. We introduce pair representation scaling, an inference-time method that biases conformational sampling in diffusion-based structure predictors by multiplying the latent pair representation by a single scalar before the Pairformer trunk, without retraining, an auxiliary model, or a second forward pass. On 86 two-state targets spanning domain motions and membrane transporters, scaling broadens the conformational ensembles of both AlphaFold 3 and Boltz-2 and recovers alternative states that default inference misses, most strongly in AlphaFold 3, where the gains extend even to targets deposited after the training cutoff. It approaches the alternative-state recovery of alignment-based sampling methods, and the benefit persists even without a multiple sequence alignment. The predicted distance distributions show that scaling shifts the encoded two-state distribution toward the experimentally observed alternative state, a directed modulation rather than arbitrary perturbation. Pair representation scaling is an interpretable, low-cost handle on the conformational ensembles of deep learning structure predictors. O_FIG O_LINKSMALLFIG WIDTH=200 HEIGHT=62 SRC="FIGDIR/small/701250v3_ufig1.gif" ALT="Figure 1"> View larger version (16K): org.highwire.dtl.DTLVardef@5e52dorg.highwire.dtl.DTLVardef@1090910org.highwire.dtl.DTLVardef@3218beorg.highwire.dtl.DTLVardef@f67068_HPS_FORMAT_FIGEXP M_FIG C_FIG Pair representation scalingA query sequence and its multiple sequence alignment enter the Pairformer trunk, where the latent pair representation z is rescaled by a single global scalar to (1 + {beta}) z and the diffusion module then generates a structure. The alignment and the trained weights are unchanged, and the same operation applies in AlphaFold 3 and Boltz-2. The scalar {beta} is a single explicit handle: sweeping it broadens the sampled ensemble toward alternative states, and its effect can be read out directly inside the network.
02 /Specification & Classification
Click any tag to browse matching papers- Citations
- 3
- Assay
- Research Focus
- Approach
In-depth summary
Claude Opus 5.5 · xhigh effort · 8 Nov 2026 · from the full text of version 3 (9 Aug 2026) · CC BY 4.0
Multiplying the AlphaFold 3 or Boltz-2 pair representation by a single scalar (1+β) at the Pairformer input, without retraining, broadens sampled ensembles toward alternative states across 86 two-state targets.
Structure predictors such as AlphaFold usually return one dominant conformation, and existing routes to alternative states reshape the input alignment (e.g. AF-Cluster), inject generic dropout noise, or optimise an auxiliary objective, none giving a single interpretable control. The new handle is one scalar on an internal representation, so it transfers unchanged between AlphaFold 3 and Boltz-2, still works without an alignment, and its effect can be read out inside the network.
Key results
ShownIn AlphaFold 3, sweeping β improved per-state success, worst-case recovery and path coverage in every target group over default inference. It matched the strongest alignment perturbation and also helped on targets released after the training cutoff.
Per-state success rate (2 Å) 0.60 default vs 0.73 scaling across 86 targets; worst-case minimum RMSD reduced significantly in every group (paired Wilcoxon, Holm-corrected); fill ratio more than doubled on transporters; after-cutoff worst-case RMSD improved 1.18 Å (95% bootstrap CI 0.47 to 2.03 Å, 23 targets). BioEmu still had the broadest fill ratio (0.43 vs 0.29). · 2.2 Benchmark across conformational-change types…
ShownIn Boltz-2 the same operation had a smaller effect: only the fill-ratio gain was significant against default inference, and the harder state was recovered better only when scaling was combined with multiple sequence alignment (MSA) subsampling.
Fill ratio vs default significant after Holm correction (p < 0.001 in both pre-cutoff groups, p = 0.04 after cutoff); per-state success rose by up to 0.09 and worst-case minimum RMSD fell 0.28 to 0.60 Å, not significant (p ≈ 0.06 for worst-case, pre-cutoff). Adding MSA subsampling brought worst-case RMSD below default in both categories (p < 0.001) and raised fill ratio over scaling alone (p < 0.01). · 2.2 Benchmark across conformational-change types…
ShownThe gain depends on a structured, global rescaling at the Pairformer input. In AlphaFold 3, additive noise of matched variance did not reproduce it (Boltz-2 tolerated it, recovering close to scaling). In both predictors, neither other trunk positions nor oracle scaling of only moving contacts reproduced it.
Under variance-matched Gaussian noise, AlphaFold 3 gave steric clashes and recovered neither state; Boltz-2 tolerated it, recovering close to scaling. On 38 targets, position mattered (Cochran's Q p = 0.0001 AlphaFold 3, 0.03 Boltz-2): scaling before the outer-product mean gave no gain, after the Pairformer halved AlphaFold 3 fill ratio, and oracle scaling of only pairs moving >4 Å (Cα) did not beat default. · 2.5 Dissecting the scaling effect; 5.1.3 Gaussian-noise control; 5.2.4 Ablation subset
ShownScaling moves AlphaFold 3 distograms toward the alternative-state distances in three ways: reweighting pairs already bimodal at default, creating new modes, and shifting single peaks. Many new bimodal splits are spurious, and the effect is weak in Boltz-2.
Per-target correlation between distogram change and the experimental inter-state distance change: median Pearson r = 0.56 over 35 AlphaFold 3 targets with a recovery gain, positive in 34. Bimodal fraction of change-contact distograms rose (paired Wilcoxon p < 10^-11, both predictors, 3 and 10 recycles), yet about half (AlphaFold 3) and three quarters (Boltz-2) of bimodal pairs matched neither reference. · 2.5.1 Internal representations under scaling; 2.5.2 Distogram bimodality; S4 Distogram–experiment agreement across targets
ShownScaling shifts ensemble mass from the dominant to the harder state rather than adding coverage on top. Where the harder state is reached, it remains a small share of the 250 models.
Targets never reaching the harder state: 63% to 44% (AlphaFold 3), 66% to 55% (Boltz-2). Median AlphaFold 3 domain-motion harder-state occupancy rose from 0 to 0.052 (13 of 250 models). Median dominant-state occupancy fell 0.97 to 0.74 (AlphaFold 3) and 0.86 to 0.73 (Boltz-2); harder-state occupancy did not rise in the Boltz-2 transporter group (p = 0.73). · S2 Ensemble occupancy of the two states; 3 Discussion
Methods, data and reproducibility
Pair representation scaling multiplies the pair tensor by (1 + β) at the Pairformer input on every recycle, leaving sequence, MSA and weights unchanged, with β in {±0.15, ±0.30, ±0.45, ±0.60, ±0.75}. Ensembles of 250 models per target were scored against both reference states (TM-score, Cα RMSD, fill ratio) and compared with sampling baselines. Distogram, ablation and denoising-trajectory analyses were added.
- Domain-motion targets with hinge or lid rearrangements: 22 BioEmu domain-motion proteins plus 17 OC23 (AFsample2) proteins; UniProt sequences with paired PDB reference structures · n = 39 targets
- Membrane transporters interconverting between inward- and outward-facing states: 15 of 16 IOMemP transporters plus 32 of 37 entropy-guided folding benchmark membrane proteins · n = 47 targets
- Deposited intermediate PDB structures cross-referenced from UniProt, with TM-score 0.70 to 0.95 to both references · n = 75 structures from 13 targets
- Ablation subset: 23 OC23 domain-motion proteins and 15 IOMemP transporters, 150 models per target · n = 38 targets
- Input MSAs taken once from the AlphaFold 3 server and reused across methods (Boltz-2 at depth 1024 sequences); sequence-only runs without MSA
- Methods & tools
- AlphaFold 3 3.0.1, Boltz-2 2.2.1, BioEmu 1.3.1, US-align, ChimeraX, DSSP, DBSCAN, AF-Cluster (clustering protocol), AFsample (inference-time dropout), UniProt, RCSB Protein Data Bank, Li et al. distogram peak-finding protocol
- Software & environment
- AlphaFold 3 3.0.1, Boltz-2 2.2.1 and BioEmu 1.3.1 at released default inference settings, with recycles fixed at three for the two diffusion predictors (Table 2). No container or requirements file is described.
- Assay conditions & compute
- Predictions run on NVIDIA RTX A6000 and H100 GPUs; Pegasus (University of Tsukuba) and SQUID (The University of Osaka) resources acknowledged. Run times and memory not reported.
- Reproducibility gaps
- No deposition of prediction ensembles or MSAs is linked; the target list and reference chains are in supplementary Table S1, and the sampling and inference settings are in Tables 1 and 2.
Limitations (6)
- Gains are partial and target-dependent: some targets never reach the alternative state at any β, and for Boltz-2 only fill ratio improved significantly over default inference. · stated by the authors
- Without a reference structure the sampled alternative state cannot be identified, because the helpful sign of β is target-specific. Picking the two highest-pLDDT (predicted confidence) models fell below random, though structural clustering beat chance. · stated by the authors
- The after-cutoff split uses reference release dates only, so homologs of post-cutoff targets may have been seen in training; the authors note a sequence- or structure-similarity split would be stronger. · stated by the authors
- Only two-state proteins were benchmarked; disordered regions, multi-state systems, complexes and ligand-induced changes remain untested. Sampled states carry no equilibrium populations, and sequence-only gains rescue isolated cases rather than offering a general route. · stated by the authors
- Ablations (contact-localized scaling, scaling position, MD-conditioned mode) used a 38-target subset at 150 models per target, and the denoising-trajectory evidence comes from two AlphaFold 3 targets (FlgA and protein Q). · noted in this summary
- AF-Cluster was adapted to a fixed k = 10 ranked clusters rather than exhaustive clustering, which may understate it. Column masking used the 40% rate AFsample3 tuned for AlphaFold 3, and dropout (0.25, following AFsample) was applied identically to Boltz-2, so these settings were not tuned per predictor. · noted in this summary
Who should read:Read this if you run AlphaFold 3 or Boltz-2 and want a one-parameter, retraining-free way to sample alternative conformations such as transporter inward/outward states or hinge motions, alone or with MSA subsampling.
Reproducibility ledger
Nobody has reported trying to reproduce this paper yet. Did you replicate an experiment or run the code? Your report saves the next lab a week.
Sign in to file a reproducibility report.
Related papers
- AFsample3: Generating and selecting multipleconformational states with Alphafold3
Yogesh Kalakoti, Björn Wallner·bioRxiv·Jan 17, 2026
- AlphaConformers: Structure-guided sampling enables prediction of multiple protein conformations
Julie Daniel, lucas vitoriano, Diego Javier Zea·bioRxiv·Aug 22, 2026
- SteerAF: Distogram-based Steering of AlphaFold2 toward Alternative Conformations
Jie Tang, Zefeng Zhu, … Chen Song·bioRxiv·Jun 19, 2026
- ConforFold Recovers Alternative Protein Conformations Beyond MSA Subsampling
Raulia R. Syrlybaeva, Eva‐Maria Strauch·Protein SciencePeer-reviewed·Apr 9, 2026
- ConforFlux: Particle-Guided Trunk Repulsion for Diverse Protein Conformations
Shosuke Suzuki, Toshiyuki Amagasa·bioRxiv·May 17, 2026
- Sampling in structure-token space enables accurate prediction of multiple protein conformations
Zhiyuan Wang, Yue Yu, … Dongbo Bu·bioRxiv·Sep 9, 2026