Polymorph-Dependent Success of Generative Protein Design Against Amyloid Fibrils: A Cross-Method Benchmark Including AL Amyloidosis
Polymorph-Dependent Success of Generative Protein Design Against Amyloid Fibrils: A Cross-Method Benchmark Including AL Amyloidosis
Where did the research take place?
The study site has not been established. Author addresses may differ from where the research occurred.
Publication status: preprint
A plain-language reading has not been prepared for this paper yet.
Original abstract
Abstract Generative protein design tools such as diffusion-based backbone generators and AlphaFold2-hallucination pipelines have been validated almost exclusively on globular receptor and signaling-protein targets. Whether these methods generalize to amyloid fibrils - flat, pocket-less, cross-beta surfaces that are structurally unlike any benchmark target used to date - is unknown. We benchmarked three generative design tools (RFdiffusion3, BindCraft, and BindEnergyCraft) for their ability to design binders against eight cryo-EM fibril structures spanning three amyloid diseases: AL (immunoglobulin light-chain) amyloidosis, Parkinson's disease (alpha-synuclein), and Alzheimer's disease (amyloid-beta). AL amyloid has not previously been attempted with any of these tools. Using a common AlphaFold2-Multimer scoring oracle and a mandatory ProteinMPNN sequence-redesign step for diffusion-generated backbones, we found that design success is polymorph-dependent rather than disease-dependent: amyloid-beta fibrils resisted all three methods (interface predicted template modeling score, i_pTM, 0.09-0.15 across both polymorphs tested, and zero accepted designs across 83 combined AF2-hallucination trajectories), alpha-synuclein was largely tractable (three of four polymorphs reaching i_pTM >= 0.70), and AL amyloid split evenly between the two extremes (i_pTM 0.235 for one polymorph, 0.585 for another). Diffusion-based backbone generation (RFdiffusion3) produced confident designs on several targets where AF2-hallucination-based methods (BindCraft, BindEnergyCraft) produced none. Across all RFdiffusion3 designs, binder beta-sheet-forming amino acid propensity correlated strongly with predicted interface confidence (Pearson r = 0.80, p < 0.001; n = 24), suggesting a compositional basis for engaging the fibrils' cross-beta surface. We additionally document a hardware-access barrier: a fourth candidate tool, ESMFold2, could not be run at all on our compute infrastructure (three NVIDIA TITAN RTX GPUs) because its bfloat16 Triton kernels require Ampere-generation (compute capability >= 8.0) hardware, excluding Turing-generation GPUs entirely. This is, to our knowledge, the first computational feasibility assessment of generative binder design against AL amyloid, and the first systematic comparison of these tools on any amyloid fibril target.