Speak to a Protein: Language Interfaces

• Speak to a Protein is a research direction that enables bi-directional dialogue between natural language and protein data, bridging the modality gap between biological sequences and human language.
• These systems integrate techniques like supervised instruction tuning, cross-modal projection, and retrieval-based methods to convert protein sequences and structures into accessible text outputs.
• They support diverse applications such as protein captioning, question answering, and interactive analysis, advancing the integration of computational linguistics with molecular biology.

Searching arXiv for the cited papers and closely related work on protein–language interfaces. “Speak to a Protein” denotes a research direction in which proteins become objects of natural-language interaction rather than only inputs to task-specific predictors. In this literature, the interface may translate protein sequence into text, map text into protein sequence, answer free-form questions from sequence or structure, retrieve and synthesize protein-specific literature and ligand evidence, or reorganize protein embedding space so that biological relatedness is directly accessible to retrieval and nearest-neighbor reasoning (Wang et al., 2023, Liu et al., 2024, Wang et al., 2024, Xiao et al., 2024, Wang et al., 7 Feb 2025, Chen et al., 13 Oct 2025, Navarro et al., 1 Oct 2025, Ofer et al., 7 May 2026). The unifying problem is the modality gap between protein representations and human language; the main methodological families are supervised instruction alignment, cross-modal projection, multimodal early fusion, adaptive in-context exemplars, contrastive metric learning, and retrieval-grounded interactive analysis.

1. Conceptual scope and task landscape

The modern “speak to a protein” agenda emerges from a recurring asymmetry. LLMs are strong at human language but poor at raw biological sequence understanding, whereas protein LLMs are effective at sequence-level biology but cannot converse in natural language. Several papers identify additional bottlenecks: prior protein-text systems were often one-directional, classification-oriented rather than generative, weakly multimodal, or limited by annotation imbalance and the absence of instruction-style supervision (Wang et al., 2023, Liu et al., 2024, Wang et al., 7 Feb 2025).

Within that shared agenda, the operative task varies substantially. Some systems emphasize bidirectional generation between protein and text; some focus on protein-to-text generation for captioning or question answering; some ground answers jointly in sequence and 3D structure; some avoid training altogether and rely on adaptive contextual exemplars; and one system turns protein analysis into an interactive multimodal dialogue coupled to literature retrieval, structure inspection, ligand analysis, code execution, and live visualization (Navarro et al., 1 Oct 2025, Chen et al., 13 Oct 2025, Xiao et al., 2024).

Work	Primary substrate	Reported interaction mode
InstructProtein	protein sequence + natural language	bidirectional generation
ProtT3	protein sequence + text	captioning, protein question-answering, protein-text retrieval
ProtChatGPT	sequence + structure + user question	interactive conversations
ProteinGPT	sequence and/or structure + question	comprehensive analysis and responsive inquiries
Prot2Chat	sequence + structure + text	protein Q&A with early fusion
Protein-as-Second-Language	sequence + contextual exemplars	zero-shot protein understanding
Speak to a Protein	literature, structures, ligand data, code, 3D scene	interactive multimodal co-scientist
ProtSent	protein sequence embeddings	retrieval and k-nearest-neighbor probing

This landscape shows that “speaking” is not a single task. It can mean translation between modalities, open-ended explanation, question-conditioned summarization, design, or evidence-grounded scientific interaction.

2. Bidirectional translation between human language and protein language

“InstructProtein” frames the problem most explicitly as alignment between human and protein language, and presents a single LLM with bidirectional generation capabilities: taking a protein sequence as input to predict its textual function description, and using natural language to prompt protein sequence generation (Wang et al., 2023). The model is trained in two stages. First, multilingual pretraining is performed on both protein sequences and text using an autoregressive objective,

$$L(\mathcal{X}) = \sum_i \log P(x_i \mid x_{i-k}, \dots, x_{i-1};\theta),$$

starting from OPT-1.3B, with continued training on protein sequences from UniRef100 and text from PubMed abstracts. Proteins are tokenized at the amino-acid level, text uses GPT-2 BPE, pretraining runs for about 40k steps, and instruction tuning runs for another 20k steps.

The distinctive component is supervised instruction tuning from a protein knowledge graph constructed from UniProtKB annotations. The graph includes entities and relations for protein families, domains, sites, functions, locations, and biological processes, and triples are converted into instruction-response examples of the form $(\text{instruction}, \text{input}, \text{output})$. To address annotation imbalance, the paper introduces a debiased sampling strategy with

$$P((p, r, t)) = \frac{1}{m}\times \frac{1}{\lVert C_i\rVert}\times \frac{1}{\lVert p\rVert},$$

where $p$ is a protein in cluster $C_i$, $m$ is the number of clusters, and $\lVert p\rVert$ is the number of annotations on that protein. It also introduces “Knowledge Causal Modeling” (KCM), arranging triples in a directed acyclic graph from lower-level sequence features to higher-level functions, explicitly compared to chain-of-thought reasoning.

Experimentally, the model is evaluated on zero-shot classification tasks including subcellular localization prediction, Gene Ontology branches BP, MF, and CC, and metal ion binding prediction, using QA-style prompts. It reportedly outperforms OPT, LLaMA, Alpaca, Galactica, and BioMedGPT by large margins, substantially improving both accuracy and AUPR on localization and GO tasks while avoiding the severe category collapse seen in other LLMs. For design, it supports zero-shot instruction-protein pairing and de novo generation from structure-based or function-based prompts such as “generate a protein that enables heme binding”; generated sequences are examined with AlphaFold/ColabFold, ESM2 embeddings, DiffDock, and Smina. The paper presents this as a first step toward truly “speaking to a protein,” because the system can both explain proteins and answer with protein sequences.

3. Cross-modal protein-to-text generation

A second line of work focuses on making proteins “speak” by converting sequence representations into natural language. “ProtT3” is a two-stage protein-to-text framework built from a frozen ESM-2 protein encoder, a Q-Former-style cross-modal projector, and Galactica-1.3B as the text generator, with LoRA adaptation in the generation stage (Liu et al., 2024). The Q-Former contains learnable query tokens that attend to the PLM representation and extract a compact latent representation

$Z \in \mathbb{R}^{N_q \times d},$

which is then projected into the LM input space and used as prefix-like conditioning for prompts such as “Describe this protein’s function.”

Stage 1 trains the Q-Former on protein-text pairs using Protein-Text Contrasting, Protein-Text Matching, and Protein Captioning. Stage 2 connects the pretrained Q-Former to the LM and fine-tunes the LM with LoRA under the same autoregressive generation loss. The paper establishes benchmarks for protein captioning, protein question-answering, and protein-text retrieval. On captioning, ProtT3 reports over 10 BLEU-2 points improvement over the strongest baseline on Swiss-Prot and ProteinKG25, with Swiss-Prot Exact Match 25.74 and BLEU-2 55.03, and ProteinKG25 Exact Match 5.48 and BLEU-2 76.53. On PDB-QA it reaches 65.0% overall exact match, compared with 62.5% for LoRA fine-tuned Galactica, 15.5% for ProteinChat, and 60.2% for a question-only baseline. In retrieval, it reports Swiss-Prot P2T Acc 68.3 and T2P Acc 68.1, ProteinKG25 P2T Acc 55.8 and T2P Acc 55.6, exceeding baselines by more than 14% accuracy in test-set retrieval. Ablations show that replacing the Q-Former with an MLP hurts performance, skipping stage 1 reduces generation quality, and larger PLMs improve retrieval.

“ProtChatGPT” addresses the same modality gap with a different alignment stack: frozen ESM-1b for 1D amino-acid sequence, frozen ESM-IF1 for 3D structure, a Protein-Language Pertaining Transformer (PLP-former), a multi-level projection adapter, and Vicuna-13b as the decoder (Wang et al., 2024). The PLP-former uses learnable query tokens and cross-attention to extract the most useful protein information before the projection adapter compresses sequence and structure embeddings into soft protein prompts. Training is also two-stage. Stage 1 uses ProtDescribe for Protein-Text Contrastive Learning, Protein-grounded Text Generation, and Protein-Text Matching; stage 2 freezes PLP-former and LLM and trains the adapter on sequence-structure-description pairs from an RCSB-PDB-derived dataset. Evaluation on 1,000 held-out protein sequence-structure-description pairs uses BLEU, ROUGE-L, METEOR, CIDEr, SPICE, and PubMed BERTScore, and the full model performs best relative to variants without structure or without PLP-former. Qualitative examples further show homologous protein comparison, context-dependent functional interpretation, and open-ended conversational question answering.

Together, these systems define a canonical architectural pattern: a frozen or partially frozen protein encoder, an intermediate alignment module that compresses protein information into text-relevant latents, and an LLM that converts those latents into fluent descriptions or answers.

4. Multimodal conversational question answering with sequence and structure

Other systems make “speaking to a protein” explicitly conversational and multimodal. “ProteinGPT” is trained so that a user can upload a protein sequence, a protein structure, or both, and then ask natural-language questions about that protein (Xiao et al., 2024). Its architecture combines the sequence encoder ESM-2 esm2_t36_3B_UR50D, the structure encoder esm_if1_gvp4_t16_142M_UR50, linear projection layers that produce soft prompts, and an LLM backbone chosen from Vicuna, LLaMA-2, LLaMA-3, or Mistral. Stage 1 aligns sequence and structure to text descriptions with the question prompt empty; stage 2 performs instruction tuning with GPT-4o-guided decomposition of descriptions into question-answer pairs. The dataset, ProteinQA, contains 132,092 proteins filtered from 204,826 raw proteins, with about 20–30 property tags per protein and 5–10 QA pairs per protein in the dataset description, while later training statistics report around 35 questions per protein and about 40 total QA pairs after augmentation. Stage 2 uses about 3.7 million QA samples. On a held-out evaluation over 160 proteins and 3,508 QA-protein pairs, ProteinGPT-Mistral reports ROUGE-1 0.461, ROUGE-2 0.048, ROUGE-L 0.460, ROUGE-LSum 0.457, BERTScore F1 0.829, PubMedBERT F1 0.784, and GPT-score F1 0.733; for closed-ended factual questions, LLaMA-3 and Mistral backbones reach around 80% accuracy, while Vicuna and LLaMA-2 remain above 70%.

“Prot2Chat” similarly aims at protein Q&A, but makes early fusion the central design principle (Wang et al., 7 Feb 2025). It modifies ProteinMPNN so that sequence and structure are encoded in a unified way, initializing the node representation directly from the sequence embedding,

$h_V = \text{Embedding}(S),$

and concatenating the outputs of nine released ProteinMPNN models to obtain an encoder dimension $D_c = 128 \times 9 = 1152$. A text-aware protein-text adapter then compresses the protein representation into question-aware soft prompt tokens, using the hidden state of the last question token from the same LLM that later generates the answer. The resulting prompt is read together with the question by the LLM. Training efficiency is achieved by freezing the encoder and applying LoRA only to the LLM’s $(\text{instruction}, \text{input}, \text{output})$0 and $(\text{instruction}, \text{input}, \text{output})$1 modules with rank $(\text{instruction}, \text{input}, \text{output})$2, alpha $(\text{instruction}, \text{input}, \text{output})$3, and dropout $(\text{instruction}, \text{input}, \text{output})$4. The adapter has 106,483,712 trainable parameters and LLaMA3 has only 3,407,872 trainable LoRA parameters, for a total of about 109M trainable parameters. On Mol-Instructions, Prot2Chat reports BLEU-2 35.85, ROUGE-1 57.21, ROUGE-2 38.09, and ROUGE-L 50.51; on UniProtQA, fine-tuned Prot2Chat reaches BLEU-2 6.72, ROUGE-1 15.71, ROUGE-2 9.25, and ROUGE-L 15.57. In KIMI and expert manual evaluation it ranks first overall, with average ranks 1.45 under KIMI and 1.49 under expert review. Ablations show that removing the fine-tuned LLM, removing protein sequence, or removing early fusion of text all degrade results, and the case study argues that structure-aware early fusion yields more accurate, domain-relevant answers than sequence-only prompting.

These systems move beyond captioning toward protein-conditioned dialogue. Their common claim is that sequence alone is insufficient for many protein questions, and that effective natural-language interaction often requires explicit structural grounding and question-aware fusion.

5. Prompt-only bilingual interpretation and embedding-space formulations

A different strategy dispenses with further gradient updates. “Protein as a Second Language for LLMs” reformulates amino-acid sequences as a symbolic language that a frozen LLM interprets in context through adaptive sequence-question-answer triples (Chen et al., 13 Oct 2025). The framework constructs a bilingual corpus of 79,926 protein-QA instances from Swiss-Prot, beginning from 573,661 Swiss-Prot entries with GO annotations, pruning the GO DAG with a depth-adjusted minimum support threshold

$(\text{instruction}, \text{input}, \text{output})$5

and a child-imbalance ratio

$(\text{instruction}, \text{input}, \text{output})$6

compared against

$(\text{instruction}, \text{input}, \text{output})$7

Redundancy is reduced by MMseqs2 sequence-level deduplication at 70% sequence similarity and by annotation-semantic deduplication based on GO information content. The resulting corpus spans 11,693 attribute-based QA, 12,108 knowledge-based QA, 23,681 descriptive text QA, and 32,444 true/false QA instances. At inference time, the prompt contains a query sequence, similar protein examples, natural-language descriptions or QA pairs, and the target question, with exemplar selection based on both sequence homology and text/QA similarity. Reported gains are an average ROUGE-L improvement of +7% and a maximum gain of +17.2%, with adaptive context often surpassing fine-tuned protein-specific LLMs. Using only sequence similarity hurts average performance by 5.2%, and using only text/QA similarity hurts by 2.8% relative to the dual criterion.

“ProtSent” addresses the same interface problem at the level of embedding geometry rather than prompting (Ofer et al., 7 May 2026). It fine-tunes ESM-2 35M and 150M backbones end-to-end within the Sentence Transformers framework, mean-pooling non-padding residue tokens into a single vector and optimizing MultipleNegativesRankingLoss with temperature 0.05 across five protein-pair datasets: Pfam families, Pfam hard negatives, AlphaFold DB structural pairs, STRINGDB protein–protein interactions, and Deep Mutational Scanning data. The main loss is

$(\text{instruction}, \text{input}, \text{output})$8

and DMS uses CoSENTLoss to preserve ranking of continuous fitness values. Evaluation is deliberately restricted to frozen embeddings and a k-nearest-neighbor probe with $(\text{instruction}, \text{input}, \text{output})$9 across 23 downstream tasks. On ESM-2 150M, ProtSent improves 15 of 23 tasks, including +105.0% on remote homology detection, +17.3% on variant effect prediction, and Recall@1 on SCOPe-40 structural retrieval from 0.423 to 0.507, a +19.9% gain. On ESM-2 35M it improves 16 of 23 tasks, including +40.5% on remote homology and Recall@1 on SCOPe-40 from 0.385 to 0.445, a +15.5% gain.

Taken together, these papers suggest two low-overhead but conceptually distinct routes to protein-language interaction. One keeps the model frozen and teaches protein interpretation through adaptive bilingual context. The other trains the embedding space itself so that nearest-neighbor structure becomes biologically meaningful.

6. “Speak to a Protein” as an interactive multimodal co-scientist

The title “Speak to a Protein” is used directly by an interactive system for protein analysis that combines natural-language dialogue with retrieval, visualization, and code execution (Navarro et al., 1 Oct 2025). Its frontend integrates a chat interface, a browser-based molecular viewer powered by Mol*, and an integrated Python sandbox implemented with Pyodide. The backend is an LLM agent that can run locally or via an external API; in the reported implementation it uses OpenAI’s GPT-4.1 and orchestrates tool calls through Model Context Protocol servers. Available tools include literature search, UniProt search and lookup, ChEMBL bioactivity retrieval, PDB/PDBe structural lookup, a MoleculeKit codebase search utility, and Python execution.

The retrieval layer is protein-conditioned. From a UniProt accession, PDB ID, or FASTA sequence, the system resolves protein identity, gathers associated structures and references, normalizes citation identifiers to PubMed Central IDs when possible, downloads accessible full-text PMC papers in XML, cleans them into passages, and builds a retrieval-augmented generation corpus with LlamaIndex. The initial corpus construction takes about one to two minutes for a new protein, while cached follow-up queries are answered in seconds. UniProt provides the canonical annotation layer, including sequence, functional comments, domains, active and binding sites, post-translational modifications, sequence variants, literature references, and cross-references. ChEMBL contributes assay-level SAR data with compound IDs, SMILES, assay descriptions, publication context, and standardized measurements. PDB/PDBe supplies structural metadata and co-crystallized ligands, filtering out common solvents and ions.

A central feature is grounding answers in a live 3D scene. The agent can highlight residues, change representations, center or zoom on regions of interest, measure distances and angles, remove waters, split chains, extract subsets, and superimpose structures using expressive VMD-like selection logic. The system can also generate and run Python for custom filtering, plotting, alignment, and table generation, with the executed code shown transparently in the interface.

The case studies are quantitative. For the D3 dopamine receptor, the system identifies the relevant UniProt entry, loads structure 3PBL, and retrieves inhibitors from ChEMBL, including CHEMBL5841759 with $$P((p, r, t)) = \frac{1}{m}\times \frac{1}{\lVert C_i\rVert}\times \frac{1}{\lVert p\rVert},$$0 nM and CHEMBL5802711 with $$P((p, r, t)) = \frac{1}{m}\times \frac{1}{\lVert C_i\rVert}\times \frac{1}{\lVert p\rVert},$$1 nM. For a D3R versus D2R comparison, the literature tool retrieves 12 passages for D2R and 10 for D3R, and the synthesized answer emphasizes a conserved orthosteric core but selectivity differences in the extended binding pocket and extracellular loops, mentioning Trp100, EL1/EL2, and positions 1.39 and 7.35. The CDK2 case study is more extensive: the system resolves human CDK2 as UniProt accession P24941, finds 462 associated PDB structures, extracts 479 unique ligand/structure pairs with relevant co-crystallized small molecules, maps 258 unique ligands to ChEMBL, finds 132 with activity data, produces an IC50-only filtered dataset of 353 unique records, and then deduplicates to about 100 entries. The top 20 most potent complexes span sub-nanomolar to 15 nM IC50 values; 14 of 20 structures are successfully rendered; aligned pocket views are restricted to residues within 6 Å of the ligand while excluding solvents and ions such as SO4, CH3, SGM, GOL, ACT, and ACE. The paper reports that the platform already had more than 18,000 registered users.

This system differs from protein-LLMs that only emit text. It treats protein analysis as a language–code–3D loop in which an answer is expected to remain synchronized with literature evidence, structure manipulation, ligand tables, and executable analyses.

7. Limitations, misconceptions, and open problems

The literature does not claim direct biochemical communication with proteins. This usage suggests a model-mediated interface in which sequence, structure, annotations, literature, ligand data, or embedding neighborhoods are aligned with natural-language queries and responses. The metaphor is therefore informational rather than molecular (Wang et al., 2023, Navarro et al., 1 Oct 2025).

Several limitations recur. InstructProtein notes that the model still struggles with numerical reasoning and quantitative protein properties, making it better suited to qualitative descriptions than to highly constrained protein engineering tasks (Wang et al., 2023). ProtT3 reports substantial unimodal bias in PDB-QA and identifies numerical answers as especially difficult, while supplementary-information questions are harder than structure/property questions (Liu et al., 2024). Prot2Chat implies that evaluation remains imperfect and therefore combines BLEU and ROUGE with KIMI ranking and expert review; it also suggests that differences between datasets reflect style and length differences in target texts (Wang et al., 7 Feb 2025). The interactive Speak to a Protein platform currently accesses only public information, can be stressed by large numbers of complex structures, can suffer from mismatches across databases with different residue or structure representations, and can overwhelm the model’s context window with long tool outputs (Navarro et al., 1 Oct 2025). ProtSent further shows that not all tasks benefit equally from contrastive restructuring: some stability-related regression tasks do not improve and sometimes degrade, AlphaFold-based structural pairs are unsupervised rather than curated SCOPe labels, and the reported results are from single training runs without multi-seed uncertainty estimates (Ofer et al., 7 May 2026).

A broader methodological tension remains unresolved. Some systems rely on explicit supervised alignment and large protein-text corpora; some rely on question-aware multimodal compression; some rely on prompt construction alone; some relocate the problem into retrieval or embedding geometry. A plausible implication is that “speaking to a protein” is best understood not as a single model class but as an interface layer spanning translation, explanation, design, retrieval, and interactive scientific reasoning. The shared trajectory is toward systems in which a protein can be queried in natural language while the answer remains grounded in biologically meaningful sequence, structure, or evidence.