RedNet

RedNet is a multiscale graph neural network for fixed-backbone protein binder sequence design, developed by Zihao Xie and Jinbo Xu at the Toyota Technological Institute at Chicago and released as a bioRxiv preprint in May 2026. Given the backbone of a protein-protein complex, RedNet predicts amino acid sequences for a binder chain that are compatible with the target interface, addressing the same inverse-folding problem as tools like ProteinMPNN and LigandMPNN but with explicit attention to the target's side-chain chemistry and to binding selectivity.

The model is motivated by a practical gap in existing inverse-folding methods: most encode only backbone geometry, discarding the target side-chain information that drives molecular recognition at a binding interface, and they optimize sequences for fold compatibility without regard to whether the designed binder might also recognize structurally similar off-target proteins. RedNet tackles both issues — it encodes target side-chain atoms at the interface, and it pairs the network with a novel contrastive decoding algorithm motivated by a thermodynamic decomposition of binding free energy, allowing designs to balance affinity and stability while improving discrimination against off-targets.

RedNet sits alongside the MPNN family of fixed-backbone sequence design models in the protein design landscape, but it is specialized for binder design and selectivity rather than general monomer or assembly design. The work is currently available as a CC-BY preprint with open code and pretrained weights; it has not yet undergone peer review.

#Key Features

• Target side-chain encoding: Unlike backbone-only inverse-folding models, RedNet incorporates the side-chain atoms of the target protein at the binding interface, supplying the chemical context that governs molecular recognition.
• Contrastive decoding for selectivity: A novel contrastive decoding algorithm, motivated by a thermodynamic decomposition of binding free energy, biases the designed sequence toward the intended target and away from a structurally similar off-target.
• Affinity/stability balancing: The decoding formulation is designed to trade off binding affinity against fold stability, rather than optimizing only for sequence recovery or fold compatibility.
• Multiscale graph representation: The network processes interface and structural context at multiple scales to capture both local interactions and broader binder geometry.
• Open weights and CLI inference: A pretrained checkpoint is distributed on Zenodo with a command-line inference tool, enabling reuse without retraining.

#Technical Details

RedNet is a multiscale graph neural network trained on heterodimer structures from the Protein Data Bank for fixed-backbone binder sequence design. On native sequence recovery for heterodimers, the authors report 43% recovery, compared with 37% for ProteinMPNN on the same task — a meaningful improvement attributed to the added target side-chain information. The contrastive decoding procedure operates at inference time: it requires a structure of the off-target complex as an additional input, and uses the contrast between target and off-target to steer sequence sampling toward selective binders. A pretrained checkpoint is released on Zenodo, the code is available on GitHub under the Apache 2.0 license, and the preprint is distributed under CC-BY.

#Applications

RedNet is aimed at researchers designing protein binders — for example, engineered binding proteins, biosensor components, or therapeutic candidates — where both high affinity for a target and selectivity against closely related off-target proteins matter. The contrastive decoding mode is particularly relevant when a designed binder must avoid cross-reacting with a structurally similar protein, a common challenge in designing binders against members of a protein family. Because it accepts a fixed backbone and outputs sequences via a CLI tool with released weights, it can be slotted into existing de novo design pipelines in place of or alongside MPNN-family sequence designers.

#Impact

By encoding target side-chain chemistry and introducing a thermodynamically motivated contrastive decoding scheme, RedNet contributes a selectivity-focused approach to the rapidly growing toolkit for protein binder design. Its reported gain in native sequence recovery on heterodimers over ProteinMPNN suggests that interface chemistry carries design-relevant signal that backbone-only models discard. Important caveats remain: the work is a preprint and has not been peer-reviewed, the headline benchmark is sequence recovery rather than experimentally validated binding or selectivity, and the contrastive decoding mode requires a known off-target structure at inference — a usage constraint that limits applicability when no such structure is available. As an open-weights, openly licensed release, RedNet nonetheless provides a concrete, reproducible starting point for groups working on selective binder design.