Bridge-IF

A backbone does not determine a sequence. Homologous proteins that share barely a third of their residues fold to the same shape, so the map from structure to sequence is one-to-many, and a model trained to emit a single best amino acid per position is being asked for something the biology does not define. Autoregressive designers compound the problem: each residue is written conditioned on the residues already chosen, and an early mistake propagates through everything that follows without ever being revisited.

Discrete diffusion answers the first half of that — sampling naturally covers a distribution of sequences rather than one point estimate — but it answers it awkwardly. A diffusion model starts from a prior that carries no information, a uniform distribution over the twenty amino acid types, and spends hundreds of denoising steps rebuilding knowledge the backbone was holding all along. Bridge-IF replaces that noise prior with a data prior. A structure encoder reads the backbone and writes a complete draft sequence in one shot; a Markov bridge is then pinned between that draft and the native sequence, learning the stochastic path that connects two real distributions instead of a path that connects noise to data. Inference walks the path forward, progressively correcting the draft into a design.

Bridge-IF was developed by Yiheng Zhu and colleagues at Zhejiang University, Alibaba Cloud and the University of Science and Technology of China, and presented at NeurIPS 2024.

#Key Features

• Data-to-data generative process: The bridge is pinned at a structure-derived draft rather than at noise, so every trajectory starts already close to a plausible answer and a design converges in 25 steps.
• Frozen language model backbone: The pretrained protein language model that approximates the bridge is not updated at all; only the modulation and adapter parameters wrapped around it are trained, keeping the evolutionary knowledge intact and the fine-tuning cheap.
• Structure-conditioned transformer blocks: An adaptive layer norm variant regresses shifts to the scale and bias terms from the timestep embedding and a pooled structure representation, initialised to predict zero so the pretrained behaviour is preserved when training begins.
• Cross-attention structural adapter: A bottleneck adapter with residual connection lets each block attend back to the encoder's fine-grained structural features, recovering detail the pooled representation discards.
• Simplified training objective: Reparameterising the bridge turns the variational bound into a reweighted cross-entropy computed over the tokens not yet converted to the target, which is both simpler to optimise and focused on the residues still needing refinement.

#Technical Details

The structure encoder is the one-shot discriminative model PiFold, supervised by native sequences to propose the prior; the bridge network is ESM-1b 650M or an ESM-2 checkpoint, modulated as described above. Training uses the standard CATH splits — 18,024 / 608 / 1,120 proteins for CATH v4.2 and 16,153 / 1,457 / 1,797 for v4.3 — with a cosine schedule, 25 timesteps, batches of roughly 6,000 residues, and up to 50 epochs on a single NVIDIA 3090. On the full CATH v4.2 test set the ESM-2 650M variant reaches 58.59% median recovery at perplexity 3.83, and the ESM-1b variant reaches 58.93% on CATH v4.3. Scaling the base model across the ESM-2 series from 8M to 3B parameters produces a clean log-linear recovery trend, topping out at 61.27%. Foldability was assessed on an external 82-structure set with 100 sequences sampled per backbone and refolded by ESMFold: TM-score 0.81 and 54.08% recovery, against 0.80 and 41.44% for ProteinMPNN. On de novo backbones generated by Chroma at lengths from 100 to 500, self-consistency TM-score is 0.73 and designability 0.85.

#Applications

The model serves the two standard inverse folding jobs. Given an experimentally determined backbone it proposes sequence variants for protein engineering, sampling an ensemble that can be filtered downstream rather than returning one candidate; given a backbone invented by a generative structure model it supplies the sequences that make the design synthesisable — the step every de novo pipeline needs between structure generation and the bench.

#Impact

Bridge-IF establishes the Markov bridge as a practical formulation for discrete biological sequence generation, showing that a well-chosen data prior buys both accuracy and a twenty-fold reduction in sampling steps over noise-based discrete diffusion. Its frozen language model with lightweight structural conditioning is a template other structure-to-sequence problems can borrow. The evaluation is entirely in silico, with no wet-lab characterisation of the designs. Distribution is limited: the code is MIT-licensed and built on the LM-Design/ByProt codebase, but no trained Bridge-IF weights are published anywhere. The Zenodo deposit holds two LM-Design checkpoints, one per CATH split — ESM-1b 650M with a PiFold encoder, supplying the pretrained structure encoder that proposes the prior, not the bridge itself — so a working designer has to be trained from the public CATH corpora. There is no hosted API or packaged release.