Prot2Chat

Prot2Chat is a multimodal protein question-answering framework that lets a user ask free-form natural-language questions about a protein and receive generated text answers, drawing on both the protein's amino acid sequence and its 3D structure. It was developed by Zhicong Wang and Ziqiang Cao at Soochow University with collaborators at Changping Laboratory and Peking University, posted as a preprint in February 2025 and published in Bioinformatics in July 2025.

Most computational protein annotation is framed as classification: a model assigns a label from a fixed ontology. That framing is rigid — it cannot answer questions the label set does not anticipate, and it gives no explanation. Language-model-based alternatives address the flexibility problem but typically consume only sequence, discarding structural context, and often require training a very large number of parameters. Prot2Chat targets both gaps at once.

Its distinguishing design choice is early fusion. Rather than generating a protein embedding independently and appending it to the prompt, Prot2Chat encodes the user's question first, then uses that question representation to compress protein information into a small set of virtual tokens. The protein representation the LLM sees is therefore already conditioned on what was asked. Note that despite the similar name, Prot2Chat is unrelated to Prot2Token, which is an ESM-2-based multi-task next-token prediction framework with a different architecture and purpose.

#Key Features

• Modified ProteinMPNN encoder: The protein encoder adapts ProteinMPNN so that sequence and backbone structure are encoded together in a unified representation, giving the language model access to geometry rather than sequence alone.
• Question-conditioned protein-text adapter: The adapter compresses the protein encoding into virtual tokens using vectors derived from the question, so the compression retains what is relevant to the specific query instead of a generic summary.
• Early fusion before decoding: Protein and text information merge before the LLM begins generating, rather than being concatenated as separate modalities at the input boundary.
• Frozen encoder with LoRA decoding: The protein encoder is frozen entirely and the LLM is adapted with low-rank adaptation, keeping training cost low relative to full fine-tuning of a comparable multimodal system.
• Open-ended answers: Output is generated text rather than a class label, so questions outside any predefined annotation vocabulary can still be answered.

#Technical Details

Prot2Chat pairs the modified ProteinMPNN encoder with LLaMA3-8B-Instruct as both the question encoder and the answer decoder. Trainable parameters total 109 million — 106,483,712 in the adapter and 3,407,872 in the LoRA modules on LLaMA3 — while the protein encoder contributes none. Training ran for 2 epochs using the Adam optimizer, requiring approximately 1,600 GPU-hours on an NVIDIA RTX 3090.

Evaluation used two datasets: the protein-oriented subset of Mol-Instructions, derived largely from UniProt annotations, and UniProtQA, introduced by BiomedGPT. Baselines included zero-shot LLaMA3-8B-Instruct, a LLaMA3 model fine-tuned on protein sequences, and larger protein-specialized chat systems. The authors report both automated text-generation metrics and expert human evaluation of answer quality, with Prot2Chat leading on both. The headline generalization result is on UniProtQA, where Prot2Chat evaluated zero-shot — with no training on that dataset — remained competitive, indicating that the learned protein-to-token compression transfers rather than memorizing dataset-specific phrasing.

#Applications

Prot2Chat suits researchers who want to interrogate a protein conversationally rather than through a fixed annotation pipeline: asking about likely function, family, catalytic role, or properties, and following up on the answer. Because structure is part of the input, it is applicable to proteins where fold carries information that sequence similarity alone would miss — including predicted structures. The authors ship a web interactive platform and a rapid installation path alongside model weights on Hugging Face and Google Drive, lowering the barrier for wet-lab groups without machine learning infrastructure. The frozen-encoder design also makes it a reasonable starting point for teams wanting to swap in a different LLM backbone.

#Impact

Prot2Chat's peer-reviewed publication in Bioinformatics gives it more standing than most protein-chat preprints, and its early-fusion design is a concrete argument that the point at which modalities merge matters as much as which modalities are included. Keeping trainable parameters at 109 million while using an 8-billion-parameter decoder is a useful demonstration that multimodal protein assistants need not be trained end to end. Limitations are real: evaluation rests on two datasets whose answers derive from curated annotation databases, so performance reflects agreement with existing annotations rather than discovery of new biology; generated text can be fluent without being correct, which is why expert evaluation was necessary; and the repository carries no license file, leaving reuse terms unclear.