Skip to main content
Back

AlloGen Teaches AI to Target One Protein Shape While Rejecting Another

AlloGen introduces a learned scoring network that steers any protein binder generator toward molecules selective for one conformational state.

Pro

  • AlloGen introduces a learnable scoring network (Q_theta) that ranks binders by conformational selectivity, not just affinity.
  • The scorer is an SE(3)-invariant graph transformer trained on paired apo/holo structures with a two-phase curriculum.
  • It plugs into any backbone generator (RFdiffusion, PXDesign, Proteina-ComplexA) as a reranker or gradient guide without retraining.
  • Benchmark spans 65 targets across 15 protein families; standard energy proxies fail where Q_theta succeeds.
  • Wet-lab validation: 5/10 designed peptides bound holo calmodulin (46.6 nM to 1.06 microM), none bound apo.
  • Model, code, and Colab notebook available at huggingface.co/ChatterjeeLab/AlloGen.

Protein binder design has spent years chasing tighter affinity, but for many of the most therapeutically interesting targets, affinity alone is a dead end. A new framework called AlloGen tackles a problem that current generative models systematically ignore: how do you design a binder that latches onto one shape of a protein while rejecting another? The team behind it, from the Chinese University of Hong Kong and the University of Pennsylvania, paired the method with wet-lab experiments showing that the computational selectivity signal actually predicts real binding behavior.

The blind spot in every binder generator

Proteins are molecular switches. Kinases, nuclear receptors, and GPCRs shuttle between active and inactive conformations, and which state a drug stabilizes determines whether it activates or inhibits the target. Despite a wave of generative tools like RFdiffusion, PXDesign, Proteina-ComplexA, BoltzGen, and BindCraft, all existing methods share a fundamental limitation: they condition on a single receptor conformation and optimize for fit to that structure alone. A binder that hugs the active state just as happily as the inactive one provides no functional specificity, no matter how strong the affinity reading.

Conventional scoring functions measure binding affinity but not differential affinity across states, and thus provide no signal for conformational selectivity. AlloGen's premise is that this gap is not a fundamental limitation of structure-based design, but a missing piece of supervision.

A modular scorer that plugs into any generator

Rather than building yet another generator, the authors built a scoring network and bolted it onto existing ones. The core component, called Q_theta, is a small SE(3)-invariant interface graph transformer that reads a receptor-binder complex and outputs a single number between 0 and 1 reflecting how well the binder fits that particular receptor conformation. SE(3)-invariance just means the score does not change if you rotate or translate the whole complex in space, which is a sanity-check property any physical scoring function should have.