EvoBind-multimer

A PROTAC links a known binder for an E3 ligase to a known binder for the target, so the target must already have a chemical handle — and of the 600-plus E3 ligases in the human genome, only a handful do. Molecular glues drop the linker: one monovalent molecule stabilises an interface between two proteins with no natural affinity. But nearly every approved glue, thalidomide included, was found by accident and rationalised afterwards.

EvoBind-multimer (EBM) attacks that problem generatively. From the sequences of two proteins alone — no binding site specified, no pre-existing ligand for either partner, no experimental ternary structure — it designs a macrocyclic peptide that binds both at once. The central move is architectural: a single-chain AlphaFold 2 network is made to process three entities — the two targets and the peptide bridge — via artificial chain breaks encoded as residue-index offsets, so a network never trained on protein complexes is made to fold one. A design loop then predicts the complex, scores it with a loss combining inverse peptide pLDDT with the peptide's mean atomic distance to each target, accepts improvements and mutates one random position, for 1,000 iterations.

EBM comes from Patrick Bryant's group at Stockholm University and SciLifeLab, with cell and tumoroid work at Karolinska Institutet and Lund University, and was posted to bioRxiv in August 2026. It extends the lab's EvoBind line, which designs binders against a single target, to the two-target ternary case; sibling model RareFold applies the same inverted-predictor strategy to noncanonical amino acids.

#Key Features

• Sequence-only ternary design: The two target sequences are the sole required inputs; interface residues, peptide length and complex geometry are discovered rather than supplied.
• Complex-naive backbone by design: Having seen only single chains, the network carries no prior toward interfaces already in the PDB — an advantage when the goal is a new interface, not a known one.
• Macrocyclic output: Head-to-tail cyclic peptides of 4–10 residues, kept small to improve the odds of crossing cell membranes.
• Simultaneous sequence and structure generation: Target interfaces adapt conformationally over the trajectory, capturing induced fit rather than docking onto a rigid holo structure.
• Cellular mechanistic validation: Designs were tested by live-cell NanoBRET, siRNA knockdown of VHL and neddylation inhibition with MLN4924, not by in-silico scores alone.

#Technical Details

MSAs are built with HHblits 3.1.0 against uniclust30_2018_08, then paired and block-diagonalised so the network reads coevolutionary signal within and between the two targets. One iteration takes 30–40 seconds, so a 1,000-step trajectory completes in 8–11 hours on a single 40 GB NVIDIA A100. Five initialisations across peptide lengths 4 to 10 gave 35,000 candidates per target pair; the top 10% by loss, filtered at peptide pLDDT ≥ 85, left 646 designs for VHL–KRAS and 353 for VHL–BRD4.

Four macrocycles were synthesised: KRAS decapeptides K587 (APGDPVCSWW) and K743 (RPGDPVCSWW) at pLDDT 90.0 and 90.5, and BRD4 designs B195 (YTCNNP) and B938 (NRYCLPHYFV) at 91.3 and 90.6. All four raised the live-cell NanoBRET ratio significantly under proteasome inhibition. By western blot, K743 cleared endogenous KRAS to 0.3 of the DMSO level — matching the PROTAC LC-2 — and B195 and B938 cleared BRD4 to 0.4 and 0.5 against 0.6 for MZ1, with MYC downregulated alongside; VHL knockdown and MLN4924 rescued target levels in both systems, confirming a Cullin-RING-dependent mechanism. Re-predicting the complexes with AlphaFold-Multimer, AlphaFold 3 and Boltz-2 placed the interfaces on VHL's canonical small-molecule glue pocket at low confidence — training priors overriding a genuinely novel interface.

#Applications

The immediate application is degrading proteins with no known ligand: EBM needs none for either partner, and can target any E3 ligase with a known substrate-binding domain sequence. More broadly, the same machinery designs induced proximity for its own sake — recruiting phosphatases, kinases or trafficking machinery to a chosen target — making it relevant to synthetic biology and chemical-probe work as much as drug discovery.

#Impact

EBM shows that ternary interface design is computationally addressable from sequence, and that a complex-naive model can be a better starting point for novel interfaces than one trained on them. The work is candid about its limits: peptides were delivered by lipid-mediated transfection rather than crossing membranes unaided, and validation covers one E3 ligase and two targets. Most instructive is a divergence the authors foreground: in patient-derived neuroblastoma tumoroids, B195 and B938 degraded BRD4 in the LU-NB-2 model but drove VHL-dependent accumulation of it in LU-NB-3, turning the same molecule from a degrader into a stabilising LOCKTAC. Proximity was established in both; only its downstream processing differed. The work remains a preprint awaiting peer review. Code inherits AlphaFold 2's Apache 2.0 license, the AlphaFold 2 parameters are used unmodified under CC BY 4.0, and the EBM design protocol carries CC BY-NC 4.0, restricting commercial use.