The explosion of genomic sequencing has created a massive disparity in molecular biology: we know the raw amino acid sequences of millions of proteins, but we understand the three-dimensional structures and biological functions of only a tiny fraction. Traditionally, predicting how a protein behaves, where it localizes, or how it interacts with other molecules has depended heavily on obtaining high-resolution 3D structures or extensive experimental metadata.
However, acquiring structural data is slow and expensive. As noted by Xingyi Cheng et al. in their work on ProtGO, traditional experimental methods simply cannot keep pace with the influx of sequence data. This dependency creates a severe bottleneck in drug discovery (Taehan Kim et al. in Site4Drug) and protein engineering (Frédéric Cadet et al. in iSAR).
To bypass this bottleneck, a major research direction focuses on sequence-only prediction. The core intuition behind this approach is that a protein's primary sequence implicitly contains all the physical, chemical, and evolutionary blueprints required to dictate its form and function. By developing computational frameworks that extract these hidden signals directly from raw amino acid strings, researchers can predict complex biological properties without needing explicit structural templates or auxiliary functional annotations.
Transforming Sequences into Numerical Signals
The primary challenge of sequence-only prediction is translation: how do we convert a linear string of categorical letters (amino acids) into a continuous numerical representation that captures spatial and chemical meaning?
Early approaches solved this by mapping amino acids to their known physicochemical properties—such as hydrophobicity, charge, and polarity—and analyzing them within localized windows or across the entire sequence. For instance, Changhui Yan et al. demonstrated in their study on DNA-binding sites that local sequence windows, when enriched with evolutionary profiles, contain sufficient signal to accurately map out nucleic acid interfaces. Similarly, Yi Xiong et al. integrated evolutionary position-specific scoring matrices (PSSMs) with physical properties in their heme-binding residue predictor, proving that structural coordinates are not a strict prerequisite for identifying active binding sites.
[Primary Sequence] -> [Physicochemical Mapping] -> [Digital Signal Processing (FFT)] -> [Activity Prediction]
M-A-G-K-L-F... [Hydrophobicity, Charge] [Extract Periodic Frequencies] (Fitness/Stability)
To capture long-range dependencies along the sequence without relying on a 3D template, researchers introduced digital signal processing techniques. In innov'SAR, Nicolas Fontaine et al. converted amino acid strings into "extended numerical sequences" and applied Fast Fourier Transforms (FFT). This allowed the model to analyze the sequence in the frequency domain, capturing periodic structural and chemical patterns along the primary chain. As shown by Frédéric Cadet et al. in iSAR, this mathematical abstraction enables highly accurate protein fitness and mutation stability predictions even when training data is extremely scarce.
Mapping Macromolecular Interactions
Predicting how a protein interacts with other entities—such as ions, RNA, or other proteins—is inherently a spatial problem. Yet, sequence-only methods succeed here by learning the statistical regularities of molecular recognition interfaces directly from raw strings.
For protein-protein interactions (PPIs), the challenge lies in capturing the correlated mutations and complementary chemical environments between two distinct sequences. Chang Zhou et al. addressed this in their multi-scale encoding framework by combining multi-scale physicochemical descriptors with gradient boosting decision trees. Rather than looking at individual residues, this approach captures the broader chemical context across different scales of the sequence.
Other frameworks focus on local sequence combinations. Long Zhang et al. introduced the Local Conjoint Triad Descriptor (LCTD) in DNN-LCTD, which classifies amino acids into groups based on their dipoles and side-chain volumes, analyzing continuous triads of residues to map out interaction surfaces.
This sequence-only recognition capability is not limited to protein pairs. Usha Muppirala developed RPISeq to predict RNA-protein interactions using nothing but raw sequence strings, demonstrating that the fundamental rules governing cross-molecular recognition are deeply embedded within primary sequence statistics.
The Deep Learning and Language Model Paradigm
The modern evolution of sequence-only prediction moves away from hand-engineered descriptors toward deep learning architectures that discover biological features autonomously.
Early deep learning applications proved that neural networks could recognize complex spatial sorting signals without human guidance. For example, the authors of the Convolutional LSTM architecture for subcellular localization combined convolutional layers (to capture local motifs like signal peptides) with Long Short-Term Memory networks (to capture long-range contextual dependencies). This allowed the network to learn the "grammar" of cellular translocation entirely from raw sequence strings, bypassing the need for manually curated metadata.
Raw Sequence ---> [ Pretrained Protein Language Model ] ---> Latent Representation ---> Downstream Tasks
(M-A-G-K...) (Learns evolutionary grammar/context) (Embedding Vector) (Function/Interaction)
Today, this concept is fully realized through pretrained protein language models (pLMs). By treating amino acids like words and proteins like sentences, pLMs trained on massive, unannotated sequence databases learn the underlying constraints of evolution and biophysics.
In SPROF-GO, Qianmu Yuan et al. leveraged these rich latent embeddings alongside homology-based label diffusion networks to achieve state-of-the-art function prediction with remarkable computational efficiency. To model interactions directly, Hungyu Chen et al. developed the Protein Pair Language Model (PPLM). Instead of embedding proteins in isolation, PPLM jointly processes pairs of sequences, allowing the attention mechanisms to capture inter-protein relationships, binding affinities, and contact points implicitly.
By shifting the analytical burden from physical structures to deep statistical representations, sequence-only methodologies provide a scalable, high-throughput alternative that unlocks the functional meaning hidden within raw genomic data.
Key Foundational Work
Predicting DNA-binding sites of proteins from amino acid sequence (2006) — contributed 1 concept elements to this direction.
The SeqFEATURE library of 3D functional site models: comparison to existing methods and applications to protein function annotation (2008) — contributed 1 concept elements to this direction.