Boltz-Perturb: Improving Diversity and Accuracy in Protein-Ligand Co-Folding through Training-Free Conditioning Perturbation 1. The paper argues that many protein–ligand co-folding failures are not purely “model can’t represent the right pose”, but “sampling doesn’t reach it”: a true-coordinate injection at an intermediate diffusion step lets Boltz-2 refine into the correct binding-mode basin, implying the correct pose is often present in the learned energy landscape but under-sampled. 2. Boltz-Perturb is a training-free inference framework that increases binding-pose diversity by injecting time-annealed noise into the denoising module’s conditioning signals (the static bottleneck shared across samples in AF3-style architectures). 3. Two complementary perturbation targets are introduced: Token Conditioning Perturbation (TCP): perturbs the single per-token conditioning tensor (st), implemented here mainly on ligand tokens. Token Bias Perturbation (TBP): perturbs the attention logit bias (B) derived from pairwise representations (zij), focusing on protein–ligand interaction regions (cross-attention) and optionally ligand–ligand. 4. The key design principle is “time-aware perturbation”: noise magnitude decays over diffusion steps (TADS), so perturbations encourage exploration early/mid trajectory without destroying late-stage refinement. This parallels condition-annealed sampling ideas from diffusion in vision/language, adapted to structured co-folding latents. 5. To reduce manual schedule tuning, the paper proposes TADS-Auto: it monitors the model’s own stochastic noise variance during sampling and activates/calibrates perturbations only when injected noise won’t be dominated by the diffusion noise, selecting λauto(t) from a dynamic and a time-decay candidate scale. 6. Perturbations are made context-aware by scaling injected noise with local standard deviation of the targeted conditioning region, plus optional regularizers: moment-matching rescaling (keep mean/variance aligned to the original signal), norm-preserving projection (preserve magnitude; change direction), and interpolation (ψ) to smoothly blend original and perturbed conditioning. 7. On five diagnostic post-cutoff PDB complexes, vanilla Boltz-2 produced very low diversity and sometimes very high RMSD (>7 Å). With Boltz-Perturb, at least one perturbation setting achieved <2 Å ligand RMSD for all five targets; large gains came from TBP on protein–ligand cross-attention for high-error cases and TCP for others. 8. On a curated 57-target subset of Runs N’ Poses (single-chain/single-ligand, post training cutoff, and where prior Boltz-2 best RMSD >2 Å), perturbations improved oracle success substantially with fewer samples: Vanilla Boltz-2: oracle success 19.30% (180 samples/target). High diffusion temperature baseline: 22.81% (270 samples/target). TCP-C11 (mid-trajectory ligand-token perturbation): 26.32% (60 samples/target). TBP-C5 (s=12): 22.81% (60 samples/target). This corresponds to similar or better oracle success than high-temperature sampling while using ~78% fewer samples (and thus much less compute). 9. Increased accuracy tracks increased diversity: best-performing perturbation settings show higher ligand RMSF across the generated ensemble (with and without protein-alignment context), supporting the hypothesis that broader exploration improves the chance of landing in the correct binding basin. 10. Confidence is not a reliable proxy for pose correctness under perturbation: mean confidence distributions remain similar across methods, yet TCP/TBP improve top-20 success rates markedly (e.g., TCP-C11 doubles SR20 vs best baselines). This suggests pose selection likely benefits from complementary rescoring (physics-based or other) even when perturbation improves the candidate pool. 💻Code: github.com/MSDLLCpapers/bolt… 📜Paper: biorxiv.org/content/10.64898… #ComputationalBiology #ProteinLigand #DiffusionModels #AlphaFold3 #Boltz2 #DrugDiscovery #StructurePrediction #GenerativeModels #Sampling #Bioinformatics