TrMRF

Inverse folding models answer the question "what sequence folds into this backbone?" — but most of them answer a weaker question than the name implies. A model trained purely on P(sequence | structure) sees only the target conformation, with no way to know whether the sequence it proposes might fold equally well into something else entirely. The consequence shows up in the lab as designs that express and fold but sit at low stability, or that adopt an alternative state the designer never modeled.

TrMRF — transform-restrained Markov Random Field — is the structure-to-sequence half of a system built to close that gap, from Sergey Ovchinnikov's group at MIT with collaborators at the University of Washington, the University of Tokyo, and Northwestern. It is deliberately constructed as the inverse of trRosetta: where trRosetta takes conservation and coevolution statistics from a multiple sequence alignment and predicts inter-residue geometry, TrMRF takes the geometry and predicts those features, maximizing the pseudo-likelihood of the alignment that would produce them. Because both directions share a representation, the two networks compose.

That composition is the paper's central result. Running TrMRF and trRosetta jointly — optimizing a sequence against both losses, using both pretrained models without retraining either — generates designs measurably more stable than either model alone or than ProteinMPNN. The joint procedure lets the backbone move to accommodate the best sequence, which is precisely what a fixed-backbone designer cannot do.

#Key Features

• Inverse of a structure predictor: TrMRF consumes one-hot structural bins from a backbone plus a sequence-identity channel, and emits pairwise potentials and conservation features — the same quantities trRosetta takes as input.
• Sequence, not weights, is optimized at sampling time: Generation starts from a random sequence and updates it by gradient descent through the frozen network via a straight-through estimator with Gumbel noise on the logits.
• Composable with a structure model: The joint TrROS + TrMRF protocol backpropagates two losses simultaneously, refining the backbone and the sequence together from pretrained checkpoints.
• Tunable homolog similarity: Fixing the sequence-identity channel at 1.0 during sampling biases generation toward sequences closely matching the reference.
• Experimentally validated designs: Every design arm was assayed by cDNA display proteolysis rather than scored only by computational proxies.

#Technical Details

The architecture uses ResNet blocks with varying dilation rates to capture features at several spatial resolutions, then branches into a path that reshapes the residual output into a higher-dimensional tensor of coevolution and conservation features. Training minimized cross-entropy between actual and predicted sequences over 15,051 nonredundant Protein Data Bank proteins at ≤30% pairwise sequence identity, released before May 2018.

The evaluation generated one design per structure from each of four methods — trRosetta hallucination, TrMRF, ProteinMPNN, and the joint model — totaling 20,668 sequences, of which 13,442 survived an AlphaFold2 inter-chain predicted alignment error filter excluding probable homo-oligomers. All designs were mini-proteins under 80 residues, cysteine excluded. Among designs sharing a structure, the joint model produced the more stable sequence in 80.5% of comparisons against trRosetta, 74.4% against TrMRF alone, and 84.7% against ProteinMPNN.

The same paper establishes a second, negative result that reshapes how designs are triaged: among sequences with measured ΔG above 5 kcal/mol, only 21.7% passed the standard filters of AlphaFold2 pLDDT > 85 and ProteinMPNN cross-entropy < 1.5 — highly stable designs are being discarded by the field's default screens. As a remedy, hybrid zero-shot scores outperform any single model: combining ESM-2 pseudo-likelihood with half-masked ESMFold pLDDT gives the best correlation with experimental stability, and fully masking the AlphaFold 2 MSA input while averaging eight seeded predictions raises the Spearman correlation from 0.50 to 0.54.

#Applications

TrMRF is used for fixed-backbone sequence design and, more distinctively, as the sequence half of a joint design loop where a target fold is refined alongside its sequence. Its audience is de novo protein designers who care about absolute stability rather than only about whether a structure predictor is confident — a group for whom the filtering analysis is as actionable as the model. The released toolkit packages TrMRF and trRosetta weights with notebooks for individual and joint design, plus the hybrid stability scorers.

#Impact

The joint-optimization argument is the durable contribution: modeling both directions of the sequence–structure mapping produces more stable proteins than either direction alone, and the effect held under experimental measurement across thousands of designs. Sequences from the joint model also showed enriched hydrophilic contacts, particularly lysine–glutamate pairs, suggesting a structural mechanism behind the stability gain. The limits are stated plainly by the authors: validation covers mini-proteins under 80 residues, and larger multi-domain designs are untested. TrMRF's own weights are small — three checkpoint files totaling around 19 MB — and arrive as a zip the repository's setup script fetches. That deposit sits on Zenodo's sandbox instance, which issues test DOIs under the 10.5072 prefix and carries no preservation guarantee, so neither the weights nor the accompanying stability measurements — the full 39,210-design set and its filtered subsets — have a permanently citable archive.