In recent years, large language models (LLMs) have demonstrated an extraordinary capacity for reasoning, summarizing scientific literature, and generating code. Concurrently, specialized biological foundation models—such as protein language models (PLMs) and genomic DNA models—have mastered the complex, evolutionary grammars hidden within raw molecular sequences.
Yet, a profound disconnect remains between these two paradigms. General-purpose LLMs excel at human language but are fundamentally blind to raw biological sequences, often performing at near-random levels when asked to interpret them directly. Conversely, specialized biological models can predict structural properties or mutation effects with high accuracy, but they are "single-task experts" that lack the generative flexibility to explain why a mutation matters or to design a new molecule based on a nuanced textual prompt.
To bridge this cross-modal representation gap, a rapidly growing body of research focuses on biotext-sequence alignment. By mapping the high-dimensional, physics- and evolution-based representations of biological data into the semantic space of human language, researchers are creating unified models capable of bidirectional biological reasoning.
The Root of the Misalignment: Tokenization and Modality Gaps
Why can't we simply feed a protein sequence (like MNGTEGPNFY...) or a DNA sequence directly into a standard LLM? The barrier lies in how these models perceive data.
First, standard LLM tokenizers are optimized for human text. When forced to process single-letter amino acid or nucleotide codes, they fragment the sequences into arbitrary sub-words that destroy the underlying biological grammar. As Kamyar Zeinalipour et al. observe in Design Proteins Using Large Language Models: Enhancements and Comparative Analyses, standard tokenizers cannot naturally capture the dense information embedded in biological chains without explicit retraining or vocabulary expansion.
Second, even if tokenized character-by-character, raw sequences lack the explicit logical structure of natural language. As Haonan He et al. demonstrate in Biology Instructions: A Dataset and Benchmark for Multi-Omics Sequence Understanding Capability of Large Language Models, state-of-the-art models like GPT-4o struggle significantly on raw sequence tasks despite possessing vast textual knowledge of biology. Knowing the literature about a gene does not inherently grant the model the ability to parse its raw code.
Conversely, trying to force an LLM to act as a raw sequence decoder can backfire. Kai Zhuang et al. point out in Lost in Tokenization: Context as the Key to Unlocking Biomolecular Understanding in Scientific LLMs that integrating raw molecular sequences alongside text can sometimes degrade reasoning performance due to the fundamental misalignment between physics-based structural representations and logic-based human text.
Mechanism 1: Contrastive Alignment (The CLIP Approach)
To make text and sequences mutually intelligible, models must learn a shared semantic space. One of the most effective ways to achieve this is through contrastive learning, mirroring the architecture of CLIP (which aligns images with captions).
[ Protein Sequence ] ---> [ Sequence Encoder ] ---\
===> [ Shared Semantic Space ] (Contrastive Loss)
[ Biomedical Text ] ---> [ Text Encoder ] -------/
In this framework, a sequence encoder (like a pre-trained PLM) and a text encoder are trained simultaneously on pairs of biological sequences and their corresponding textual descriptions (e.g., from scientific papers or databases like UniProt). The model is incentivized to maximize the dot product of correct sequence-text pairs while minimizing it for mismatched pairs.
Minghao Xu pioneered this approach with ProtST: Multi-Modality Learning of Protein Sequences and Biomedical Texts, showing that joint pre-training allows for powerful zero-shot capabilities, such as retrieving specific proteins using open-ended natural language prompts. Similarly, Mingze Yin et al. leveraged contrastive alignment in Multi-Modal CLIP-Informed Protein Editing to map text instructions (e.g., "increase catalytic activity") directly to sequence modifications, bypassing the need for rigid, manual engineering rules.
Mechanism 2: Cross-Modal Projectors for Generative Translation
While contrastive learning is excellent for classification and retrieval, it does not allow a model to generate original text or synthesize new sequences from scratch. To achieve generative fluidity, researchers use cross-modal projectors to feed biological embeddings directly into an LLM's autoregressive decoder.
In this setup, a specialized biological model acts as a "vision encoder" would in a multimodal LLM. When a protein sequence is passed through a PLM, the resulting high-dimensional hidden states are translated via a projection layer (such as a linear layer or a Perceiver sampler) into tokens that match the dimensionality and semantic expectations of the LLM's embedding space.
[ Raw Sequence ] -> [ PLM Encoder ] -> [ Projector ] -> [ Sequence Tokens ] \
==> [ LLM Decoder ] -> "This protein binds to..."
[ Text Prompt ] -> [ Text Tokens ] /
Using this architecture, Zhiyuan Liu et al. developed ProtT3: Protein-to-Text Generation for Text-based Protein Understanding, enabling an LLM to "read" amino acid sequences and output coherent, novel functional descriptions. To ensure that the LLM doesn't hallucinate or rely on generic descriptions when encountering entirely novel proteins, Xiao Fei et al. introduced Prot2Text-V2: Protein Function Prediction with Multimodal Contrastive Alignment, which combines generative decoding with a contrastive alignment loss to anchor the LLM's text generation strictly to the protein's actual biophysical properties.
Instruction Tuning: Teaching the Model to Reason
Once the structural alignment is in place, the model must be trained to perform specific tasks through instruction tuning. This involves curating massive datasets of sequence-text pairs formatted as explicit prompts and answers.
A foundational contribution to this effort is Mol-Instructions, a large-scale dataset introduced by Yin Fang et al. containing millions of biomolecular instructions designed to teach LLMs how to convert structural data into textual descriptions and vice versa.
Instruction tuning unlocks bidirectional capabilities:
- Sequence-to-Text (Understanding): The user provides a sequence, and the model explains its function, cellular location, or mutation effects. For example, the framework proposed by Massimo Hong et al. in MutaPLM bridges raw sequences with biomedical literature to explicitly explain and predict the functional impacts of protein mutations.
- Text-to-Sequence (Design): The user provides a functional requirement, and the model generates a valid biological sequence. In Annotation-guided Protein Design with Multi-Level Domain Alignment, the authors demonstrate that aligning text annotations at both global and local domain levels dramatically improves the success rate of de novo functional protein generation.
This bidirectional alignment is highly visible in specialized domains like immunology. Ling Luo et al. created AFD-INSTRUCTION, a dataset of over 430,000 antibody sequence-function pairs. Fine-tuning on this data allows an LLM to act as an interactive assistant that can both analyze an antibody's neutralization mechanism and write new antibody sequences tailored to specific viral targets.
Beyond Proteins: Genomics and Multi-Omics Integration
The principles of cross-modal alignment are not restricted to amino acid sequences; they are expanding to encompass the entire central dogma of molecular biology, including DNA and cellular expression profiles.
In genomics, Adibvafa Fallahpour et al. introduced BioReason, a framework that fuses DNA foundation models with LLMs. By optimizing the system with reinforcement learning, the model learns to analyze raw genomic sequences and output transparent, step-by-step mechanistic explanations detailing how specific genetic mutations lead to disease phenotypes.
Similarly, the approach applies to numerical, non-sequence biological modalities. In scReader, Cong Li et al. demonstrated that LLMs can be prompted to interpret numerical single-cell RNA sequencing (scRNA-seq) profiles by mapping ranked gene expression data directly to functional text descriptions, significantly improving automated cell-type annotation.
Why It Works: The Synergy of Two Worlds
Biotext-sequence alignment works because it respects the strengths of both underlying architectures. We do not force an LLM to learn the physics of protein folding from raw text strings, nor do we expect a protein language model to understand human grammar.
Instead, the biological model acts as an expert translator that compresses raw, evolutionary complexity into structured embeddings. The cross-modal alignment layer maps these embeddings into the LLM's vocabulary, allowing the LLM to deploy its pre-trained, multi-step reasoning capabilities over the biological data. The result is a unified interface where human scientists can converse directly with the code of life.
Key Foundational Work
ProtT3: Protein-to-Text Generation for Text-based Protein Understanding (2024) — contributed 2 concept elements to this direction.
Mol-Instructions: A Large-Scale Biomolecular Instruction Dataset for Large Language Models (2024) — contributed 2 concept elements to this direction.
TourSynbio: A Multi-Modal Large Model and Agent Framework to Bridge Text and Protein Sequences for Protein Engineering (2024) — contributed 1 concept elements to this direction.