Multi-View Feature Representation for Biological Sequence Learning

This direction is merged into another direction: multimodal protein representation. Its last version stays readable and citable here.

Version 1 of 122 papers24 conceptsfrozen 2026-10-04

Biological sequences—DNA, RNA, and proteins—are often treated by machine learning models as simple strings of characters. However, a sequence is more than a string; it is a physical entity with biochemical properties, a structural scaffold with 3D constraints, and an evolutionary artifact with a history of selection.

Recent research has converged on a powerful paradigm to capture this complexity: multi-view feature representation. Instead of relying on a single encoding method, researchers are integrating heterogeneous data "tracks" to provide models with a comprehensive understanding of biological function.

Beyond the One-Hot Vector

The simplest way to represent a protein is one-hot encoding, where each amino acid is a sparse vector. While this captures identity, it ignores similarity; it doesn't "know" that Valine and Leucine share similar hydrophobic properties.

To bridge this gap, Edgar Liberis, Petar Veličković, and colleagues developed Parapred, which augments one-hot identity with physical and chemical descriptors like hydrophobicity and isoelectric points. This hybrid approach allows the model to generalize across residues that behave similarly in a binding pocket, even if their specific identities differ.

The Three Pillars of Biological Context

The most successful models in this field typically integrate three distinct "views" of a sequence:

  1. Evolutionary Context: By using Position-Specific Scoring Matrices (PSSMs), models can see which residues are conserved across millions of years. Paul D. Yoo and Bing Bing Zhou demonstrated in SiteSeek that combining these evolutionary signatures with hydrophobicity scales significantly improves the detection of post-translational modification (PTM) sites.
  2. Physicochemical Properties: These represent the "physics" of the molecule. Chang Zhou and colleagues used multi-scale physicochemical encoding to capture properties like polarizability and charge, allowing a Gradient Boosting Decision Tree to predict protein-protein interactions with high accuracy without needing 3D structural data.
  3. Structural and Positional Motifs: Biological signals are often hidden in local patterns (k-mers) or global arrangements. For instance, Zhixun Zhao and Xiaocai Zhang used a high-dimensional feature space that combines local k-mer frequencies with global positional encoding to identify DNA methylation sites across different species.

Integrating Heterogeneous Views

The challenge of multi-view learning is how to fuse these disparate data types—categorical identities, continuous physical scales, and dense evolutionary matrices—into a single predictive framework.

One sophisticated approach is the Triple Coding Matrix. In TripHLApan, Chuqi Lei and Jianxin Wang combined physicochemical properties, evolutionary substitution matrices, and learnable latent embeddings. By stacking these into a multi-channel input (similar to RGB channels in an image), the model can use attention mechanisms to weigh which "view" is most relevant for a specific peptide-MHC binding event.

Similarly, the AMYPred-FRL model by Phasit Charoenkwan and Saeed Ahmed acts as a "meta-predictor." It extracts features using ten different encoding methods and feeds them into an ensemble of machine learning algorithms, ensuring that the final prediction is robust to the weaknesses of any single representation.

Multi-Track Foundation Models

The frontier of this research involves moving from task-specific feature engineering to "multi-track" foundation models. Orthrus, developed by Philip Fradkin, Quaid Morris, and others, represents RNA transcripts using a multi-track scheme. Alongside the raw nucleotide sequence, it incorporates auxiliary biological tracks like splice site locations and codon positions.

This explicit inclusion of biological domain knowledge allows the model to learn "functional" representations that outperform standard genomic models. It suggests that the future of biological AI lies not just in larger datasets, but in more nuanced, multi-faceted ways of describing the sequences themselves.

Papers in this version

History & citationVersion 1 · 2026-10-04

Merged into another direction · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_multi_view_feature_representation_for_bi_b489134a,
  title = {Multi-View Feature Representation for Biological Sequence Learning},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/multi-view-feature-representation-for-biological-sequence-learning/rdr_b489134af27647eb91dded6f3c215d72},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_multi_view_feature_representation_for_bi_b489134a_v1,
  title = {Multi-View Feature Representation for Biological Sequence Learning},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/multi-view-feature-representation-for-biological-sequence-learning/rdr_b489134af27647eb91dded6f3c215d72/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}