For decades, the central dogma of structural biology has been that a protein's function is dictated by its three-dimensional shape. Consequently, predicting how a protein interacts with a drug, an RNA molecule, or another protein typically required first determining its atomic structure through expensive and time-consuming methods like X-ray crystallography or NMR. However, a growing body of research suggests that we can bypass the "3D bottleneck" entirely. By treating proteins as strings of text rather than complex geometric objects, new computational models are learning to predict biological function directly from the primary amino acid sequence.
The Bottleneck of Three-Dimensional Structure
The traditional reliance on 3D data creates a significant divide in biology. While we can sequence genomes at massive scales, our ability to solve protein structures lags far behind. As Edgar Liberis and colleagues point out in their work on Paratope Prediction using Convolutional and Recurrent Neural Networks, this reliance creates a bottleneck for research areas like antibody design, where sequence data is abundant but structural information is scarce.
The core intuition behind the sequence-only approach is that the "blueprints" for molecular recognition are already encoded in the linear arrangement of amino acids. If the sequence contains all the information necessary for a protein to fold into a specific shape (Anfinsen's dogma), it must also contain the information for the functional consequences of that shape—such as where a ligand will bind or how soluble the protein will be in a cell.
Extracting the Interaction Code
How can a model "see" a binding site without a 3D map? The answer lies in identifying local motifs and physicochemical patterns that characterize interaction interfaces.
Early successful models, such as RPISeq developed by Usha Muppirala, Vasant G Honavar, and Drena Dobbs, demonstrated that primary sequence data contains sufficient information to predict RNA-protein interactions with high accuracy. These models work by encoding the chemical propensities of amino acids—such as charge, hydrophobicity, and size—into numerical vectors.
Researchers like Sylvain Pitre have further refined this by identifying "interaction codes"—short, co-occurring polypeptide regions that serve as universal signals for protein-protein binding across different species. By focusing on these sequence-level "handshakes," models can map entire interactomes without needing a single crystal structure.
Scaling with Protein Language Models
The most significant leap in this field has come from applying Natural Language Processing (NLP) techniques to biology. Just as a large language model learns the grammar of English by reading billions of sentences, Protein Language Models (pLMs) learn the "structural grammar" of proteins by analyzing millions of unaligned sequences.
Models like OmegaFold, introduced by Ruidong Wu and others, have shown that large-scale pLMs can internalize structural and evolutionary information directly from sequence databases. This allows for high-resolution predictions even for "orphan" proteins that lack known evolutionary relatives.
This capability is being harnessed for therapeutic design. For instance, the SaLT&PepPr pipeline, developed by Garyk Brixi and colleagues, uses pLMs to identify interaction interfaces and design targeted protein degraders using only the amino acid sequence of a target’s binding partners. By bypassing 3D modeling, these tools can screen millions of sequences in the time it would take a traditional simulation to process one.
Beyond Proximity to Chemical Logic
Modern sequence-only models are moving beyond simply predicting if two molecules will touch to predicting how they will interact. Recent work on LINKER by Truong Son Hy demonstrates that sequence-based models can differentiate between specific biochemical interaction types, such as hydrogen bonds or salt bridges.
This suggests that deep learning models are not just memorizing patterns; they are capturing the underlying physics of molecular recognition. Whether it is predicting the druggability of a protein, the optimal pH for an enzyme by Xiangjun Li et al., or the solubility of a monoclonal antibody by Pietro Sormanni et al., the primary sequence is proving to be a surprisingly complete proxy for the complex reality of three-dimensional biology.