arrow_back Back to Blog

Beyond the Hype: Picking Rosetta & FoldX or AlphaFold/ESMFold/Boltz for Real Design Work

By Tensor Biodynamics • September 2025

Deep-learning structure predictors are astonishingly useful—but not universal solvents. This post outlines where Rosetta and FoldX still lead, where AlphaFold, ESMFold, and Boltz truly shine, and how to choose the right tool on evidence—not hype.

The past few years have redefined structural biology. Modern AI predictors made high-quality models accessible in minutes, and that success created a reflex: default to the newest neural model for every task. In practice, many engineering problems still benefit from constraint-aware, physics/empirical tools that let you control what changes and what stays fixed. This is where Rosetta and FoldX continue to deliver.

🧠 What DL Predictors Are Great At

Rapid, plausible 3D models from sequence. For new proteins or distant homologs, AlphaFold and ESMFold provide high-confidence monomer models quickly; newer diffusion/co-folding approaches (e.g., Boltz families) expand to assemblies and interfaces.

Throughput for exploration. Screening many sequences or variants for coarse structural plausibility is now tractable: generate models, check confidence metrics, and cluster shapes before deeper work.

Independent cross-checks. DL predictions give an orthogonal sanity check alongside physics-based scoring and experimental data—useful triangulation in early design.

🚧 Where DL Falls Short for Engineering

Limited control and constraints. Predictors typically rebuild the whole structure. They are not designed to “freeze” experimental coordinates while you alter a specific region under user-defined constraints.

Energetics are indirect. Confidence metrics and learned potentials don’t replace explicit ΔΔG for stability or binding, nor do they capture all environment- or condition-specific effects without careful interpretation.

Attractive, but sometimes misleading. Beautiful predictions can mask subtle packing errors, strained geometry, or lost contact networks that matter in function and manufacturability.

🛠️ What Rosetta & FoldX Still Do Uniquely Well

Local, constraint-aware remodeling (Rosetta). Define exactly which atoms can move, sample local backbone (Backrub/KIC/CCD), repack or design side chains, and relax the result—all while keeping the rest of the complex essentially fixed.

Fast empirical triage (FoldX). Rapidly estimate stability changes and complex binding energies, scan positions exhaustively, and prioritize variants before investing in heavy sampling or MD.

Transparent knobs. Task operations, resfiles, coordinate constraints, and score terms make assumptions explicit, enabling methodical iteration and clear rationale in reports and methods sections.

🧭 A Grounded Decision Guide

🔄 A Practical “Best-of-Both” Workflow

  1. Start from the most trustworthy structure. Clean the PDB, define the edit window, and document constraints.
  2. FoldX triage. Repair once; scan targeted mutations; compute ΔΔG for stability and complex interaction energy to down-select.
  3. Rosetta refinement. Freeze everything you want preserved; sample local backbone and repack/design allowed residues; relax under Cartesian coordinates.
  4. Lightweight dynamics. Run short restrained MD to confirm rigidity/consistency and catch hidden clashes or strain.
  5. DL cross-check (optional). Re-fold finalists with AlphaFold/ESMFold to confirm secondary-structure tendencies and gross compatibility; treat as corroboration, not ground truth.

⚠️ Pitfalls to Avoid

✅ Takeaways

Deep learning made protein models fast and accessible—an enormous win. But method fit matters: for controlled, local engineering with preserve-this-don’t-move-that requirements, Rosetta and FoldX remain the right starting points. Use DL predictors as powerful companions for exploration and cross-checks, not as one-size-fits-all solutions.