PPIFlow

Summary¶

PPIFlow generates de novo protein binder backbones by modeling residues as rigid-body frames in SE(3) and learning the flow-matching velocity field that transports random distributions into physically realistic structures. Coupled with a novel in silico affinity-maturation stage, the resulting pipeline produced picomolar-affinity binders for six of seven diverse therapeutic targets. It also produced picomolar-to-nanomolar VHHs for seven of eight targets, in both cases without any experimental affinity maturation.[1]

Method¶

Backbone generation: PPIFlow uses a pairformer module to model pairwise geometric/chemical interactions and generates backbones via flow matching (rather than diffusion), providing a geometrically consistent parameterization of rigid-body dynamics and stable conditioning on interface constraints.[1]
Sequence design: generated backbones are threaded with ProteinMPNN for monomers/binders/motifs, or AbMPNN for VHHs and scFvs.[1]
AF3Score filtering: a score-only adaptation of AlphaFold 3 (no MSA construction or recycling) that accelerates candidate prioritization roughly 100-fold relative to full AlphaFold 3 inference, used to filter designs before and after maturation.[1]
In silico affinity maturation: (1) interface rotamer enrichment identifies energetically favorable residues; (2) key interface residues are selected and merged; (3) critical residues are fixed while noise is added to neighboring backbone regions; (4) partial flow-based refinement of the unconstrained regions optimizes interface packing.[1]

Results¶

Minibinder benchmarking (75/120/160/200-residue binders against 15 targets): PPIFlow surpassed RFdiffusion, BindCraft, and BoltzGen on design success, novelty, and diversity, with a pronounced advantage for larger binders.[1]
Experimental validation, general targets (IL7RA, IFNAR2, IL17A, PD-L1, TRKA, PDGFR, VEGFA): of 210 designed candidates (30/target), all expressed successfully and 76 (36.2%) bound with sub-µM affinity by bio-layer interferometry; picomolar-affinity binders were obtained for 6 of the 7 targets without post-design experimental optimization (best K_D values from 1 pM to 74.1 nM across targets).[1]
Experimental validation, VHHs (CCL2, HNMT, PDGFR, 1433E, BHRF1, S100A4, EFNA1, IL13): sub-nanomolar-to-nanomolar binders were obtained for CCL2 (250 pM), PDGFR (4.09 nM), and 1433E (6.47 nM), with pM-to-nM affinity across 7 of the 8 targets.[1]
Ablations: interface rotamer enrichment combined with partial flow refinement systematically outperformed baselines and partial-flow-only ablations on AF3Score pTM/ipTM and Rosetta interface energetics (demonstrated in detail for the IFNAR2 and CCL2 targets).[1]

Availability¶

Code is intended for release at github.com/Mingchenchen/PPIFlow; AF3Score at github.com/Mingchenchen/AF3Score. This is a preprint, not yet peer-reviewed; confidence is set to medium pending independent replication of the experimental hit rates.

Citations¶

[1] PPIFlow authors (2026), "PPIFlow: flow-matching for de novo protein binder design with in silico affinity maturation", bioRxiv Supports: architecture, maturation pipeline, all benchmark and experimental figures above. Location: Results and Figures 1–4.