Deep learning has fundamentally transformed structural biology, most notably through end-to-end architectures like AlphaFold (Jumper et al.). However, these models harbor a critical vulnerability: their heavy reliance on deep Multiple Sequence Alignments (MSAs). When a target protein lacks a rich family of known evolutionary relatives—such as orphan proteins, synthetic de novo designs, or viral dark matter—the predictive power of traditional co-evolutionary models plummets.
To expand machine learning to the entire proteome, researchers are developing methods to overcome this "evolutionary data scarcity." Understanding why this bottleneck exists, and how modern architectures bypass it, requires looking closely at how statistical signals dissolve into noise when data is sparse.
The Mechanics of Co-evolutionary Inference
The core mechanism behind modern protein structure prediction relies on a simple biological premise: if two amino acids are physically touching in a 3D structure, a mutation in one residue will likely disrupt the protein's stability unless it is compensated by a matching mutation in the other.
In her pioneering work on de novo structure prediction, Debora Susan Marks demonstrated in Protein 3D Structure Computed from Evolutionary Sequence Variation that analyzing these correlated mutations across thousands of homologous sequences allows us to mathematically infer spatial proximity. By calculating a covariance matrix across an MSA, global statistical models can distinguish direct physical contacts from indirect correlations.
Deep MSA (Thousands of sequences)
Seq 1: ... A ... T ... (Residues mutate together over evolutionary time)
Seq 2: ... G ... C ...
Seq 3: ... A ... T ... ==> High statistical confidence of physical contact
Seq 4: ... G ... C ...
Shallow MSA (Fewer than 30 sequences)
Seq 1: ... A ... T ...
Seq 2: ... A ... C ... ==> Cannot distinguish structural constraints
Seq 3: ... G ... T ... from random phylogenetic noise
However, this statistical machinery demands significant data depth. Debora Susan Marks noted that traditional co-evolutionary analysis requires roughly 1,000 sequence homologs to achieve sufficient statistical power. Even highly optimized deep learning models like AlphaFold experience a sharp decline in accuracy when an MSA contains fewer than 30 sequences (Jumper et al.).
Why Shallow Alignments Fail
To understand why shallow MSAs fail, imagine trying to determine if two people are close friends based solely on how often they attend the same events. If you only observe three events, and they attend all of them, they might be inseparable—or they might simply be coworkers who share an office. You need to observe them across hundreds of diverse settings to prove a true social bond.
In a shallow MSA, global statistical models face the exact same problem. They cannot distinguish a meaningful structural correlation from random phylogenetic noise (Monasson et al.). If two residues appear to mutate together across only a dozen sequences, it is often because those sequences share a recent common ancestor, not because the residues physically interact. The background noise dominates, leaving downstream models with incomplete or noisy contact maps.
Substituting Local Context with Global Priors
The dominant strategy for overcoming evolutionary data scarcity is to replace family-specific MSAs with Protein Language Models (pLMs). Instead of looking deeply at a single protein family, pLMs look broadly at the entire protein universe.
The authors of the MSA Transformer (Rao et al.) demonstrated that pre-trained evolutionary models can extract clean structural signals from shallow alignments containing as few as 8 to 16 sequences. By training on millions of unaligned sequences via masked language modeling, pLMs learn a generalized "biophysical grammar"—the fundamental rules of hydrophobicity, charge, and secondary structure propensity. When faced with an orphan protein, the model uses this global prior knowledge to fill in the blanks, inferring structural constraints from the single sequence alone.
[Massive Unaligned Database] ──> Pre-trained pLM ──> Learns Global Biophysical Grammar
│
[Target Orphan Sequence] ─────────────────────────────┴──> Accurate Prediction (No MSA needed)
Taking this concept further, Marquet et al. developed VESPA, which completely eliminates alignments by using single-sequence pLM embeddings to predict variant effects. Similarly, Barrett et al. integrated pLMs directly into streamlined structure prediction pipelines, maintaining high accuracy for rare sequences without the heavy computational burden of generating deep MSAs.
Fusing Geometry with Sequence Evolution
When evolutionary history is sparse, physical and geometric constraints can act as a substitute. By fusing sequence embeddings with explicit 3D structural data, models can compensate for missing evolutionary context.
For instance, Blaabjerg et al. introduced SSEmb, a framework that merges sequence evolution with 3D structural graphs into a joint embedding. This allows the model to remain robust during variant effect predictions even when sequence data is scarce, because the physical laws governing the 3D backbone constrain the allowable mutations.
To completely decouple protein modeling from MSA dependencies, Zhu et al. developed FlexRibbon, a foundation model that jointly learns sequence and structure via diffusion-based denoising. Because it does not rely on evolutionary couplings to infer spatial relationships, it excels at modeling highly variable or engineered regions, such as antibody loops, where traditional alignments break down entirely.
Deepening Alignments via Dense Retrieval
A final, complementary angle focuses on the retrieval pipeline itself. Often, evolutionary relatives for an orphan protein do exist in genomic databases, but traditional character-matching search tools (like BLAST or HMMER) fail to find them because the sequences have diverged too far past the "twilight zone" of sequence similarity.
To solve this, Hu et al. developed the Dense Homolog Retriever (DHR). Instead of matching exact amino acid characters, DHR maps sequences into a high-dimensional vector space learned by a neural network.
Traditional Search (BLAST): Matches exact characters ──> Misses remote homologs
Dense Retrieval (DHR): Matches semantic vectors ──> Uncovers "dark matter" homologs
By performing "dense retrieval" based on semantic and structural meaning rather than surface-level sequence identity, the authors successfully uncovered remote homologs—the hidden "dark matter" of the proteome. This artificially deepens what would otherwise be a shallow MSA, restoring the vital co-evolutionary signals required by downstream structural models.