SurfFlow

SurfFlow is a generative model for therapeutic peptide design that introduces the molecular surface as an explicit design modality alongside sequence and three-dimensional structure. Peptides are an attractive therapeutic class for disrupting protein-protein interactions (PPIs), where binding is governed less by deep, well-defined pockets than by complementarity across broad, shallow interfaces. SurfFlow targets this regime by co-designing the geometric and biochemical surface of a candidate peptide together with its backbone and amino-acid sequence, rather than treating surface as a downstream consequence of an atomic model.

Developed by researchers at Stanford University (with Jure Leskovec and Jinbo Xu among the authors) and posted to arXiv in January 2026, SurfFlow uses multi-modal conditional flow matching to jointly learn distributions over sequence, structure, and surface. The work was presented in the KDD venue track and frames peptide binder generation as an "omni-design" problem in which all three modalities are generated coherently for a given target interface.

The model sits alongside recent diffusion- and flow-based peptide binder generators (such as PepGLAD and moPPIt), but is distinguished by elevating the molecular surface to a first-class, learned modality. This emphasis is motivated by the observation that surface shape and electrostatics are the most direct determinants of binding complementarity at flat PPI interfaces.

#Key Features

• Surface as a design modality: SurfFlow explicitly models molecular surface geometry and biochemical properties as a generated output, giving the model a representation directly aligned with the physics of interface complementarity rather than inferring surface only after atoms are placed.
• Multi-modal co-design: Sequence, backbone structure, and surface are generated jointly within a single conditional flow-matching framework, so the three modalities remain mutually consistent for the target.
• PPI-focused generation: The model is built for the difficult case of disrupting protein-protein interactions, where shallow interfaces make conventional pocket-based design less effective.
• Benchmark-leading results: On the PepMerge benchmark, SurfFlow reports improvements over full-atom baselines across the evaluated metrics.

#Technical Details

SurfFlow is a conditional flow-matching generative model. Flow matching learns a continuous transport from a simple prior distribution to the data distribution by regressing a velocity field, offering a stable, simulation-free alternative to denoising diffusion. SurfFlow applies this paradigm across multiple coupled modalities: it learns to generate the amino-acid sequence, the peptide backbone/structure, and the molecular surface (its geometry and associated biochemical features) conditioned on the target protein interface. By learning surface manifolds and their chemical properties directly, the model produces candidates whose generated surface is consistent with their generated structure and sequence. Evaluation is performed on the PepMerge benchmark for peptide design, where SurfFlow outperforms full-atom baseline methods across the reported metrics. As of the January 2026 preprint, no public code or model weights are available.

#Applications

SurfFlow is intended for computational design of peptide therapeutics, particularly binders that disrupt or modulate protein-protein interactions implicated in disease. Such interfaces are often considered "undruggable" by small molecules because they lack deep binding pockets, making surface-complementary peptides a compelling alternative. By co-generating sequence, structure, and surface, the model gives medicinal chemists and protein engineers candidate peptides whose predicted shape and surface chemistry are matched to a chosen target, which can be triaged with structure prediction and folding tools before synthesis. The approach is most useful early in a discovery campaign, where the goal is to propose diverse, target-conditioned starting points for a flat or feature-poor interface.

#Impact

By treating the molecular surface as a learned, generated modality rather than a byproduct of atomic coordinates, SurfFlow contributes a distinct perspective to the rapidly growing field of generative peptide design and to surface-aware molecular modeling more broadly. Its reported gains over full-atom baselines on PepMerge suggest that explicit surface conditioning is informative for the PPI-disruption setting that motivates much of therapeutic peptide development. The most important caveat is that, as a recent preprint, SurfFlow's results are computational and benchmark-based: there is no reported experimental (wet-lab) validation of generated binders, and no public code or weights were available at the time of release, which currently limits independent reproduction and adoption.