TMDiffusion (TMDF)

TMDiffusion (TMDF) is a diffusion model for the de novo design of transmembrane protein domains, developed in Patrick Barth's Laboratory of Protein and Cell Engineering at EPFL with experimental collaborators at Lehigh University and the Walter and Eliza Hall Institute. Membrane proteins are roughly 30% of the proteome and 70% of drug targets, yet under 1% of solved structures, and the deep learning tools that transformed soluble protein engineering do not transfer to them. Transmembrane helices associate through weak Cα–H···O hydrogen bonds, small-residue packing motifs, and buried polar networks whose energetics depend jointly on backbone geometry and sequence — interactions that pipelines generating a backbone first and threading a sequence onto it afterwards, such as RFdiffusion followed by ProteinMPNN, cannot model.

TMDF instead diffuses over sequence and all heavy atoms simultaneously. Every denoising step updates Cα coordinates, backbone frame orientations, sidechain dihedrals, and amino acid identity together, so sidechain chemistry shapes the backbone as it forms rather than being fitted to it. To cope with the scarcity of membrane protein structures, the model's scope is deliberately narrowed to transmembrane α-helical dimers — the motif underlying single- and multi-pass membrane proteins, including receptor tyrosine kinases and GPCRs.

The payoff is experimental rather than in-silico. Designs were tested in living cells, one was solved crystallographically at 1.7 Å, and the de novo helices were used to build synthetic receptors whose signaling output tracks their designed conformational equilibrium.

#Key Features

• Joint sequence-structure diffusion over all heavy atoms: Euclidean coordinates, SO(3) backbone orientations, SO(2) sidechain dihedrals, and categorical amino acid identity are noised and denoised in one pass, with no separate sequence design network in the loop.
• Contact maps as an inductive bias: Contact maps extracted from the noised structure are fed back into the network, reinforcing the transmembrane interactions — salt bridges, SmallxxxSmall packing motifs, Cα–H···O bonds, polar cores — that set association energetics.
• Multi-state switch design: Logits are shared across successive denoising steps so sequence sampling is conditioned on two target conformations at once, rather than intersecting sequence pools designed separately per state.
• Flexible conditioning: Designs can be generated unconditionally, guided by a native backbone at any noise level, or produced as an unbiased inhibitor chain against a fixed target monomer.

#Technical Details

Each training example is a 26-residue transmembrane dimer, 13 residues per chain, mapped to per-residue frames defined by the backbone N, Cα, and C atoms. The denoiser applies AlphaFold2-style invariant point attention with 8 heads of dimension 16 across four transformer instances. Training data was assembled from 190 homodimeric and 1,038 heterodimeric transmembrane structures plus dimeric interfaces extracted from 5,379 multipass membrane proteins, filtered to better than 3 Å resolution, fully membrane-embedded per the Orientations of Proteins in Membranes database, α-helical by DSSP, and free of tertiary contacts; the final set of 10,623 fragments was augmented to 42,492 by chain swapping and by generating unbound counterparts, then idealized with the Rosetta membrane score function. Training ran for 400 epochs at batch size 64 on two NVIDIA 3090 Ti GPUs. Designs were validated with the in vivo TOXGREEN association assay in E. coli: of 16 glycophorin A designs, 14 associated and 11 more tightly than wild type; every EphA2 left-handed design outperformed wild type, one by tenfold; 30 growth hormone receptor designs all oligomerized, 19 above wild type; and an inhibitor campaign against the FGFR3 A391E mutant reached a 47% success rate, with the authors reporting over 70% experimental success overall. The GpA3 design crystallized in lipid cubic phase and diffracted to 1.7 Å, matching the predicted model at 0.55 Å Cα RMSD.

#Applications

TMDF targets problems previously out of reach: transmembrane inhibitors that outcompete a disease-driving receptor dimer, chimeric single-pass receptors whose de novo transmembrane domain sets basal and ligand-induced signaling levels, and designed switches for mechanistic studies of membrane protein activation. Receptors built with TMDF-designed transmembrane domains responded to growth hormone at 86-120% of wild-type activity while spanning a tenfold range of basal signaling — the kind of tunable control that therapeutic cell engineering and biosensor design require.

#Impact

TMDF is a generative design model built specifically for the transmembrane protein universe, and its experimental record — cell-based association assays across four target systems plus an atomic-resolution crystal structure — is unusually strong for a de novo design method. Its scope is correspondingly narrow: the model generates α-helical dimers of fixed length, and the authors note recurrent steric clashes traced to modeling contacts via residue centers of mass. Extending the approach to multi-pass architectures would require assembling overlapping dimer interfaces. The work is a preprint and has not completed peer review. No code or model weights have been released, and the preprint is posted under an all-rights-reserved license, so the model cannot currently be run by others.