When two proteins or RNA molecules share a common evolutionary ancestor, their primary sequences mutate over time. Traditional sequence alignment algorithms—such as those utilizing substitution matrices or profile Hidden Markov Models (HMMs)—rely on these sequence patterns to find optimal alignments. However, when sequence identity drops into the "twilight zone" (often below 20%), the evolutionary signal in the primary sequence becomes too faint to distinguish from random noise.
The core insight of this research direction is that three-dimensional structure diverges much more slowly than primary sequence. Even when the letters of the genetic code have mutated beyond recognition, the structural scaffolding and local biophysical environments often remain remarkably intact. By predicting these structural features from sequence data and integrating them directly into the alignment scoring mechanics, we can align highly divergent sequences with high precision.
The Core Intuition: Complementary Information Channels
Why does adding predicted structural features improve alignment quality so dramatically? The answer lies in information orthogonality.
Traditional alignment methods look at a single channel: evolutionary sequence profiles (e.g., how likely alanine is to mutate into valine). This research direction introduces a second, complementary channel: local biophysical constraints.
Consider two residues in remote homologs. They might have entirely different amino acid identities, causing a sequence-only aligner to misalign them. However, if machine learning models predict that both residues sit in the middle of an alpha-helix, are completely buried away from solvent, and participate in similar local geometric configurations, an algorithm that weighs these features will correctly pair them.
As demonstrated by Tomas Ohlson et al. in Improved alignment quality by combining evolutionary information, predicted secondary structure and self-organizing maps, combining human-defined structural categories with unsupervised machine-learned clusters captures distinct, non-overlapping aspects of protein geometry. This synergy ensures that structural predictions and sequence-based profiles compensate for each other's weaknesses.
Mechanism 1: Expanding the Protein Scoring Function
To operationalize this intuition for proteins, researchers modify the fundamental mathematical framework of alignment algorithms—such as dynamic programming matrices or profile HMMs—to evaluate structural compatibility alongside sequence substitution scores.
In fRMSDAlign: Protein Sequence Alignment Using Predicted Local Structure Information for Pairs with Low Sequence Identity, Huzefa Rangwala and George Karypis developed a method to predict the geometric distance (fractional Root Mean Square Deviation, or fRMSD) between protein fragments from sequence data alone. By incorporating this predicted local structural similarity into the alignment score, they achieved significant accuracy gains for protein pairs with less than 12% sequence identity.
Taking this further, Xin Deng and Jianlin Cheng expanded the dimensionality of alignment calculations in two key papers:
- In MSACompro: protein multiple sequence alignment using predicted secondary structure, solvent accessibility, and residue-residue contacts, the authors integrated predicted secondary structure, solvent accessibility, and residue-residue contact maps directly into the fundamental probability calculations of multiple sequence alignments (MSAs).
- In Enhancing HMM-based protein profile-profile alignment with structural features and evolutionary coupling information, the authors extended profile-profile HMM scoring functions to include predicted backbone torsion angles and solvent accessibility.
By evaluating whether a residue's predicted local environment (e.g., its burial status or backbone twist) matches the template's environment, these methods prevent alignments from drifting out of register in low-identity regions.
Sequence-Only Alignment Score:
Score = Substitution_Matrix(Residue_A, Residue_B)
Structure-Enhanced Alignment Score:
Score = Substitution_Matrix(Residue_A, Residue_B)
+ w1 * Similarity(Predicted_Secondary_Structure_A, Predicted_Secondary_Structure_B)
+ w2 * Similarity(Predicted_Solvent_Accessibility_A, Predicted_Solvent_Accessibility_B)
+ w3 * Similarity(Predicted_Torsion_Angles_A, Predicted_Torsion_Angles_B)
Mechanism 2: Structural Constraints in RNA Alignment
The necessity of structural integration is perhaps even more pronounced in RNA alignment. Non-coding RNAs (ncRNAs) frequently exhibit extreme primary sequence variation while strictly conserving their base-paired secondary structures to maintain function.
Ignoring this structural context makes accurate alignment of long ncRNAs nearly impossible. To solve this, algorithms use predicted or known base-pairing configurations as rigid constraints or heavy priors during the alignment process.
In Two accurate sequence, structure, and phylogenetic template-based RNA alignment systems, David P. Gardner et al. demonstrated that using secondary structure information as an explicit constraint guides the alignment through highly variable sequence regions, compensating for primary sequence drift.
Similarly, Haruka Yonemoto et al. introduced CentroidAlign-Web in CentroidAlign-Web: A Fast and Accurate Multiple Aligner for Long Non-Coding RNAs, which leverages pre-existing or predicted secondary structure inputs to bypass the massive computational bottlenecks traditionally associated with simultaneous sequence-structure alignment, making the accurate alignment of massive molecules like ribosomal RNAs practical.
The Downstream Payoff: Better Structure Prediction
The benefits of structure-integrated alignment form a virtuous cycle in structural biology. Modern deep-learning structure prediction pipelines rely heavily on finding good homologous templates from databases of known structures. If the initial sequence-structure alignment is poor, the downstream 3D model suffers.
This utility is highlighted in DeepFold: enhancing protein structure prediction through optimized loss functions, improved template features, and re-optimized energy function, where Jae-Won Lee et al. improved upon AlphaFold2 architecture variants. By employing a refined sequence-structure alignment method based on conditional random fields and gradient boosted regression trees, the authors enabled significantly better identification and utilization of homologous protein templates, ultimately driving up the accuracy of final predicted side-chain positions and global topologies.
Key Foundational Work
Improved alignment quality by combining evolutionary information, predicted secondary structure and self-organizing maps (2006) — contributed 2 concept elements to this direction.
fRMSDAlign: Protein Sequence Alignment Using Predicted Local Structure Information for Pairs with Low Sequence Identity (2008) — contributed 1 concept elements to this direction.