Reproduction of the CLAIRE de novo protein-binder design pipeline

Tue, 01 Sep 2026 00:00:00 +0000

A computational project — independent work alongside my wet-lab dissertation. I reproduced the CLAIRE small-molecule binder-design pipeline end-to-end, then pushed it onto two targets it had never been run on.

Three targets, not one

Target	Why this one
Progesterone	the repository’s built-in example — a correctness baseline
Thalidomide	generalization test
4-Hydroxytamoxifen (4-OHT)	generalization test

The latter two were chosen deliberately: their polar-handle configurations are chemically distinct enough to probe generalization rather than repeat a known result.

All three produced AlphaFold2-self-consistent candidates — best case pLDDT 94.1 at 0.62 Å RMSD.

Why it matters to my own work

My dissertation designs bioPROTACs, which need a binding domain against their target — so the binder is the limiting reagent, and borrowing one confines you to targets somebody has already solved. Designing binders directly is how that constraint comes off.

I extracted the computational recipe of Guo et al. (Science, 2025) from the primary literature and built a three-layer framework to test whether it transfers to the Ras superfamily:

Calibration — T35S, a known state-switching mutant: does the diagnostic detect what it claims to?
Clinical noise — G12D, G12C: does it fire on any substitution?
Application — the hypervariable-region mutation from my own dissertation.

The first two layers are what make the third worth believing. This also meant rebuilding the simulation system in AMBER after diagnosing that a prior model was nucleotide-free — the replacement is GppNHp- and Mg²⁺-bound, with de novo parameterization for the ligand and the farnesylated cysteine anchor.

Toolchain

Rosetta/PyRosetta · ProteinMPNN · AlphaFold2/ColabFold · GROMACS · AMBER/AmberTools · OpenMM · CHARMM-GUI · antechamber, AM1-BCC, GAFF2 · Python · Linux and cloud GPU

AlphaFold2 |

Reproduction of the CLAIRE de novo protein-binder design pipeline

Three targets, not one

Why it matters to my own work

Related: testing the generalizability of a multistate design method on KRAS

Toolchain