A Text-guided Protein Design Framework

TL;DR

ProteinDT introduces a unified multi-modal framework that combines textual protein descriptions with sequence data to enable text-guided protein generation and editing. It couples ProteinCLAP (text–protein contrastive pretraining), ProteinFacilitator (text-to-protein latent mapping via a Gaussian), and flexible decoders (autoregressive and diffusion) to produce protein sequences conditioned on text. The authors curate SwissProtCLAP (441k pairs) to train ProteinDT and demonstrate strong zero-shot performance: over 90% retrieval-like accuracy for text-to-protein generation, best-hit ratio across 12 editing tasks, and competitive or superior results on six protein-property benchmarks. The work highlights the potential of textual knowledge to guide protein design, enabling zero-shot editing and providing a scalable path toward functionally guided design without task-specific labeled data.

Abstract

Current AI-assisted protein design mainly utilizes protein sequential and structural information. Meanwhile, there exists tremendous knowledge curated by humans in the text format describing proteins' high-level functionalities. Yet, whether the incorporation of such text data can help protein design tasks has not been explored. To bridge this gap, we propose ProteinDT, a multi-modal framework that leverages textual descriptions for protein design. ProteinDT consists of three subsequent steps: ProteinCLAP which aligns the representation of two modalities, a facilitator that generates the protein representation from the text modality, and a decoder that creates the protein sequences from the representation. To train ProteinDT, we construct a large dataset, SwissProtCLAP, with 441K text and protein pairs. We quantitatively verify the effectiveness of ProteinDT on three challenging tasks: (1) over 90% accuracy for text-guided protein generation; (2) best hit ratio on 12 zero-shot text-guided protein editing tasks; (3) superior performance on four out of six protein property prediction benchmarks.

Figures (10)

• Figure 1: Pipeline for ProteinDT pretraining framework (a-c) and downstream tasks (d-f). (a) ProteinCLAP, a contrastive learning paradigm, aligns the representation space of the text and protein sequence modalities. (b) ProteinFacilitator model augments the mapping from text sequence representation to protein sequence representation. (c) A protein sequence decoder, which generates protein sequences conditioned on the representations produced from previous steps. (d) Downstream text-to-protein generation task. (e) Downstream text-guided protein editing task. (f) Downstream protein property prediction task.
• Figure 2: Visualization of text-to-protein generation and text-guided protein editing. (a) Visualization for evaluation of text-to-protein generation. The pretrained ProteinCLAP is used to calculate the similarity between the sampled text and generated protein sequence pairs. (b-c) Two methods for text-guided protein editing: latent interpolation and latent optimization. (d-e) Visualization for evaluation of text-guided protein editing. For four types of editing tasks, different evaluation metrics (marked in red) are applied accordingly.
• Figure 3: Visual analysis on text-guided protein editing with latent optimization. (a-c) visualize structure editing with more/less $\alpha$-helices editing. (d-f) visualize structure editing with more/less $\beta$-sheets editing. (g-i) visualize peptide binding editing on PDB 3IQI.
• Figure 4: The inference illustration for two decoder models (step 3 in ProteinDT). \ref{['fig:pipeline_autoregressive']} is an autoregressive (AR) model based on the Transformer. It generates the protein sequence token-by-token. \ref{['fig:pipeline_diffusion_model']} is ProteinDiff, and a Transformer-based transition network is displayed. It first randomly samples a noised sequence, then conducts a denoising process using the transition network.
• Figure 5: Illustration of retrieval accuracy in text-to-protein generation.