PROPEND

The standard computational route to a designed protein has two steps: settle on a backbone, then find a sequence that folds into it. Every design goal has to be smuggled into that geometry — catalysis becomes the spatial arrangement of an active site, stability becomes the shape of the core. But the residues that hold a fold together are not the residues that do the chemistry, so a single backbone is a lossy way to state two requirements at once.

PROPEND (PROmpt-guided ProtEiN Design) removes the geometric middleman by conditioning a protein language model directly on whatever the designer can state. Borrowing the "pre-train and prompt" paradigm from natural language processing, it leaves a pretrained sequence decoder frozen and steers it with learned prompts of three kinds — an atomic 3D backbone, a blueprint giving only the identity and order of secondary-structure elements, and a natural-language functional tag. Because all three are embedded into the same latent space, they can be summed and applied together.

PROPEND was developed at Zhejiang University with the ZJU-Hangzhou Global Scientific and Technological Innovation Center and posted to bioRxiv in November 2024. It builds on ProGen2 as the frozen decoder and on the ProteinMPNN architecture as its structural prompt encoder.

#Key Features

• One model, several design languages: a single trained framework handles inverse folding from a backbone, generation from a blueprint, generation from a function tag, and any combination of the three.
• Global and local prompts: the global prompt names the design scenario and is prepended once; the local prompt carries per-position detail, so each residue sees conditioning specific to it rather than a shared prefix.
• Dual-stream self-attention: decoder blocks alternate between attention over the growing sequence and attention over the per-residue prompt stream, letting domain-level and residue-level constraints coexist in one generation.
• Frozen decoder, trained prompt encoders: only the prompt encoders and adapters are optimized, roughly 10% of the parameters, leaving the pretrained sequence model untouched.
• Masked causal language modeling: masking a fraction of residues during training removes the model's ability to coast on sequence completion and forces it to rely on the prompt.

#Technical Details

The sequence decoder is a pretrained ProGen2 model, used at both the 764M-parameter base and 6.4B-parameter xlarge sizes. Backbone prompts are encoded by a message-passing network using the ProteinMPNN v_48_010 architecture; blueprints and functional tags are encoded by a T5 text encoder, with a [PROMPT] token whose representation serves as the functional local prompt. Encoders are trained from scratch with AdamW at a learning rate of 1e-4, a 2,000-step linear warmup, dynamic batching and early stopping on validation perplexity, in PyTorch Lightning; the base configuration trained on four RTX 4090 GPUs and the xlarge on four V100s. On the CATH v4.2 standard splits (n = 1,120 test structures), PROPEND reached 63.21% native sequence recovery at 3.13 perplexity, against 52.91% and 4.02 for the strongest of ProteinMPNN, ESM-IF and PiFold, with the largest gains in surface-exposed positions. Its substitution pattern resembles BLOSUM62, and AlphaFold 2 pLDDT distributions for its designs track native sequences, including their low-confidence, disorder-like regions. Blueprint-conditioned designs recovered the intended three-state DSSP assignment at low sequence identity to natural proteins, with the eight-state confusion matrix confirming the finer-grained alignment, and reproduced the non-natural α/β topology of Top7.

#Applications

The functional-tag route is the one that reaches the bench without a structure. Prompted with the tag "DNA nucleotidylexotransferase", PROPEND generated 1,000 candidate terminal deoxynucleotidyl transferases; filtering for length and for the three catalytic aspartates gave a refined set, from which ten sequences at 40–60% identity to natural TdT were selected, all of which expressed solubly in E. coli and showed template-free primer-extension activity. In a multi-objective test, the NlpC/p60 peptidoglycan hydrolase SagA was redesigned from its crystal structure (PDB 6B8C) plus a functional tag; of five designs assayed, the best reached 105.2% of wild-type activity, where five designs from a structure-only pipeline topped out at 50.8%, and two of its designs melted at higher temperatures than wild type. The natural users are enzyme engineers and de novo design groups whose specification is easier to write down than to draw.

#Impact

PROPEND demonstrates that a pretrained protein language model can be treated as a fixed substrate and steered by modality-specific prompt encoders, rather than fine-tuned into a new model for every objective — the shift prompt tuning brought to natural language processing. Its practical contribution is multi-objective conditioning: structure and function stated separately and satisfied together, which a backbone-first pipeline cannot express. The evidence base is small — the in vitro work rests on ten TdT and five SagA sequences — and the work remains a preprint that has not been through peer review. Neither inference code nor trained prompt-encoder weights have been released; the training datasets are on Zenodo and the paper points at the third-party components it reuses, so reproducing PROPEND means rebuilding it.