The traditional paradigm of structural biology states that a molecule's three-dimensional structure dictates its function. To determine whether a protein will bind to a ligand, an RNA molecule, or another protein, researchers historically relied on high-resolution 3D coordinates obtained via X-ray crystallography, cryo-EM, or intensive computational docking simulations. However, structural determination remains a massive bottleneck—expensive, slow, and unavailable for a vast portion of the proteome.
A rapidly growing body of research bypasses this bottleneck entirely. The core insight of sequence-based interaction prediction is that primary linear sequences contain a complete, self-contained blueprint for molecular recognition. Nature’s folding and binding rules are deterministic functions of sequence; therefore, if a machine learning model can effectively capture the implicit biophysical and evolutionary constraints embedded within a sequence, it can predict binding affinities and interaction interfaces directly, achieving massive computational speedups without ever calculating explicit 3D coordinates.
How 1D Strings Encode 3D Biophysics
How can a flat string of letters predict physical binding events occurring in three-dimensional space? The answer lies in the fact that primary sequences carry intrinsic physicochemical propensities and evolutionary histories that constrain their spatial configurations.
Physicochemical Propensities
Amino acid and nucleotide sequences are not just arbitrary text; they are chains of chemical entities with specific charges, hydrophobicity, and spatial volumes. As demonstrated by Nitin Bhardwaj et al. in Genome-wide sequence-based prediction of peripheral proteins using a novel semi-supervised learning technique, basic sequence features like electrostatic charge and hydrophobicity provide sufficient signal to accurately predict protein-membrane binding. Similarly, the authors of Computational prediction of associations between long non-coding RNAs and proteins showed that secondary structures and physicochemical propensities derived purely from sequence data can be mapped into numeric vectors to predict complex lncRNA-protein interactions.
Evolutionary Context via Language Models
Modern deep learning scales this concept by using pre-trained protein language models (pLMs). By training on millions of unannotated sequences, pLMs learn the "grammar" of biology. In a single-author study, Garyk Brixi demonstrated in SaLT&PepPr is an interface-predicting language model for designing peptide-guided protein degraders that pLMs can identify interaction interfaces using only primary amino acid sequences, performing comparably to methods requiring high-resolution structures. The model relies entirely on the evolutionary and physicochemical context captured during pre-training to locate binding hotspots.
Furthermore, Tianlai Chen et al. extended this to generative design in Target sequence-conditioned design of peptide binders using masked language modeling, proving that evolutionary sequence distributions serve as an effective surrogate for geometric modeling when designing high-affinity peptide binders.
[Primary Sequence] ──► [Pre-trained pLM] ──► [Latent Biophysical Features] ──► [Interaction Prediction]
(1D Amino Acids) (Learned Grammar) (Hydrophobicity, Charge) (No 3D Coordinates Needed)
Mapping the Interaction Space
Predicting an interaction requires modeling how two distinct sequences relate to one another. Instead of calculating spatial distances between atoms, sequence-only models look for statistical co-occurrence, alignment patterns, and cross-sequence attention.
Interaction Codes and Alignments
On a genomic scale, interacting molecules often co-evolve, leaving statistical signatures in their sequences. As shown by Sylvain Pitre et al. in Short Co-occurring Polypeptide Regions Can Predict Global Protein Interaction Maps, global protein interaction networks can be mapped by identifying "interaction codes"—short, co-occurring polypeptide regions across different sequences.
To handle flexible binding registers without relying on rigid structural templates, the authors of Predicting Class II MHC-Peptide binding: a kernel based approach using similarity scores utilized data-driven sequence similarity kernels. By evaluating all possible sub-sequence alignments, their method learns underlying binding rules directly from sequence data, bypassing the need for explicit three-dimensional structural modeling.
Cross-Attention and Paired Embeddings
To capture the intricate, non-linear dependencies between two sequences, modern architectures employ attention mechanisms. For example, Tingfang Wu et al. introduced a cross-attention mechanism in CAPLA: improved prediction of protein-ligand binding affinity by a deep learning approach based on a cross-attention mechanism to fuse protein and ligand sequence features, allowing the model to dynamically weigh which regions of the protein sequence interact with specific chemical constituents of the ligand.
Going a step further, Hungyu Chen et al. developed a joint processing framework in A paired sequence language model for protein-protein interaction modeling. Rather than embedding sequences independently and concatenating them, their paired sequence language model jointly learns from pairs of protein sequences, capturing mutual evolutionary constraints to significantly improve binding affinity and contact point predictions.
Generalization vs. Memorization
The primary vulnerability of sequence-only models is the "memorization wall." Because deep neural networks are exceptional pattern matchers, they can easily memorize sequence motifs present in the training set rather than learning the underlying biophysical principles of interaction.
This limitation was exposed by Filippo Grazioli et al. in On TCR binding predictors failing to generalize to unseen peptides. They demonstrated that while sequence-only T-cell receptor (TCR) predictors showed stellar performance on standard benchmarks, they failed to generalize to unseen peptides when evaluated using a more rigorous "hard split" protocol, performing barely better than random chance. This highlights the necessity of training models to extract true transferable features rather than superficial sequence correlations.
When done correctly, however, sequence-only models achieve remarkable generalization. In AI designed, mutation resistant broad neutralizing antibodies against multiple SARS-CoV-2 strains, Kevin Jin et al. demonstrated that a structure-free AI framework trained on large-scale sequence datasets could extract generalized binding patterns. Their model successfully predicted antibodies capable of neutralizing the Omicron variant even though Omicron data was completely absent from the training set—all while delivering a 10,000x computational speedup over traditional physics-based simulations.
By treating primary sequences as rich, high-dimensional representations of molecular identity, sequence-based interaction prediction transforms a complex, three-dimensional geometric problem into a scalable, data-driven inference task, opening up new horizons for high-throughput therapeutic design.
Key Foundational Work
IDENTIFYING INTERACTION SITES IN “RECALCITRANT” PROTEINS: PREDICTED PROTEIN AND RNA BINDING SITES IN REV PROTEINS OF HIV-1 AND EIAV AGREE WITH EXPERIMENTAL DATA (2005) — contributed 2 concept elements to this direction.
IDENTIFYING INTERACTION SITES IN “RECALCITRANT” PROTEINS: PREDICTED PROTEIN AND RNA BINDING SITES IN REV PROTEINS OF HIV-1 AND EIAV AGREE WITH EXPERIMENTAL DATA (2005) — contributed 2 concept elements to this direction.