Open-source protein design tools
Most of the computational stack for protein design is now open. This is a map of what exists, grouped by the job it does. Licenses vary more than people assume, and several widely used academic tools are not licensed for commercial use, so check the license file in the repository before a tool enters a paid workflow.
Structure prediction
AlphaFold2 predicts structure from sequence and remains the reference point for monomers and many complexes. The code is open and the original weights are available for use under their own terms. ColabFold is the practical way most people run it, pairing the same models with a fast sequence search so a prediction takes minutes rather than hours.
ESMFold predicts structure from a single sequence using a protein language model, trading some accuracy for speed and for independence from sequence search. That trade is worth it when you are folding many sequences.
AlphaFold3 and the open co-folding models that followed it extend prediction to complexes with ligands, nucleic acids and modifications. Boltz and Chai are the two open implementations in wide use. Their licenses differ, and AlphaFold3's own terms are more restrictive than AlphaFold2's.
Design and inverse folding
RFdiffusion generates backbones with diffusion, including binders against a specified target site. It produces shapes, not sequences.
ProteinMPNN does the opposite job: given a backbone, it designs a sequence likely to fold into it. The pairing of a backbone generator with an inverse folding model is the standard modern recipe, usually with a prediction step afterward to filter designs that do not fold back to the intended shape.
BindCraft and similar pipelines wire those pieces together for binder design specifically, with filtering built in.
Antibody-specific tools
Antibodies have their own stack, because their structure is conserved except where it matters.
ANARCI numbers antibody sequences and assigns CDRs under the standard schemes, and almost every antibody pipeline depends on it somewhere. ImmuneBuilder and IgFold predict Fv structures quickly, which matters when you have a panel rather than a molecule. AntiFold and related models do inverse folding with antibody-specific training. RFantibody applies the diffusion and inverse folding recipe to antibody frameworks.
Check licenses carefully in this group in particular: several are free for academic use and require a commercial license otherwise.
Simulation and analysis
OpenMM and GROMACS are the two open molecular dynamics engines in common use, with mature ecosystems. MDAnalysis and MDTraj handle trajectory analysis in Python. PyMOL has an open-source build, and ChimeraX is free for academic use with a commercial license required for other users. FreeSASA, DSSP and PDBFixer cover the small jobs that every pipeline needs.
What the open stack does not give you
It does not give you a measurement. Every one of these tools produces a prediction, and predictions of binding in particular remain unreliable enough that they belong in a triage step rather than a decision step. The open stack is very good at generating candidates and at removing obviously bad ones. It is not good at telling you which of two plausible binders is better, which is what the bench is for.
It also does not give you data. Model quality follows from training data, and the public structural data is heavily biased toward what was easy to crystallize and what was interesting to publish.
How we use it
We use open tools where they are the best available, keep the license question in front of us because client work is commercial work, and treat every computational result as a hypothesis that earns its place only when an assay agrees with it.