Enregistré dans:
| Auteur principal: | |
|---|---|
| Format: | Recurso digital |
| Langue: | |
| Publié: |
Zenodo
2026
|
| Accès en ligne: | https://doi.org/10.5281/zenodo.20257600 |
| Tags: |
Ajouter un tag
Pas de tags, Soyez le premier à ajouter un tag!
|
Table des matières:
- <p>REAL FOLD ONE — Physics‑Based Refinement Engine</p> <p>MADE IN THAILAND.</p> <p>https://github.com/yoonalimsuwan/REAL-FOLD-ONE</p> <p>REAL FOLD ONE is a SOC‑driven universal refinement engine for proteins, DNA/RNA, ligands, and multimers. It uses full all‑atom physics (bond, angle, torsion, Ramachandran, LJ, Coulomb, H‑bonds, solvation) with a novel Self‑Organised Criticality (SOC) controller that adaptively tunes temperature, friction, and avalanche propagation during optimisation. No retraining is needed – it works directly on any input structure.</p> <p>REAL FOLD ONE HTS extends the engine to high‑throughput mutation scanning and epistasis analysis, enabling systematic ΔΔG evaluation for protein engineering and drug design.</p> <p>Why Refinement?</p> <p>For De Novo & Engineered Proteins (small to medium, N < 500)</p> <p>Predictors/detectors (AlphaFold3, ESMFold, RFdiffusion, ProteinMPNN) produce plausible backbones but often have:</p> <p>Sidechain clashes or wrong rotamers</p> <p>Distorted bond lengths/angles</p> <p>Unphysical Ramachandran outliers</p> <p>REAL FOLD ONE fixes all‑atom geometry, repacks sidechains, and resolves steric clashes using full physics (LJ, Coulomb, torsion, H‑bond, Ramachandran) – without retraining.</p> <p>Perfect as a physics‑based polish after any de novo design pipeline.</p> <p>For Large Natural Proteins / Complexes (N > 5,000 up to 100k+)</p> <p>Predictors have severe size limits – AlphaFold2 (~2,500 residues), ESMFold (~4k). They cannot handle viral capsids, ribosomes, or large multimeric assemblies.</p> <p>Cryo‑EM / tomography often produce medium‑resolution density maps → initial atomic models need physical relaxation. Homology models of large domains may contain steric clashes, distorted geometry, and incorrect sidechain packing.</p> <p>REAL FOLD ONE runs on any length (tested up to 100k+ residues) using sparse graphs and O(N) memory, without retraining. It refines structure while preserving global topology.</p> <p>Key Features</p> <p>SOC Controller – learnable kernel , avalanche‑driven stress relaxation, adaptive temperature & friction.</p> <p>Full‑atom protein sidechains – all 20 amino acids with correct topology (no duplicate CB).</p> <p>Full‑atom DNA/RNA – nucleotides with base pairing, stacking, and backbone.</p> <p>Ligand support – SDF, MOL2, PDB ligands with GAFF‑style force field.</p> <p>Multimer handling – chain boundaries, cross‑chain interactions, chain‑break penalties.</p> <p>Advanced electrostatics – Sparse PME, geometric multigrid Poisson solver, block‑wise multipole correction.</p> <p>Multiscale refinement – RG coarse‑graining for fast convergence.</p> <p>Itô SDE – Milstein‑scheme Langevin dynamics for stochastic exploration.</p> <p>Malliavin sensitivity – compute Greeks for temperature or other parameters.</p> <p>Antibody design – CDR H3 loop modelling with Rosetta‑style scoring.</p> <p>DNA origami – scaffold routing, staple design, oxDNA export, BV topology check.</p> <p>High‑throughput scanning (HTS module) – single‑mutation ΔΔG scan, epistasis analysis, publication‑ready plots.</p> <p>Scalability – O(N) graphs, chunked edge building, CPU‑only mode for low‑RAM environments (3 GB), multi‑GPU parallelism.</p>