For decades, the "gold standard" for predicting a protein's structure or function has been the Multiple Sequence Alignment (MSA). By aligning a target protein with its evolutionary relatives, researchers like Martin Weigt and Debora Susan Marks demonstrated that we can spot conserved residues and co-evolving pairs that signal physical contacts or functional constraints.
However, this paradigm relies on a critical assumption: that the protein has a deep evolutionary history recorded in our databases. A significant portion of the "Dark Proteome"—including many eukaryotic proteins, viral variants, and de novo designed sequences—are "orphans" with few or no identifiable homologs. When evolutionary data is sparse, traditional models like DeepSequence or EVE struggle to distinguish biological signals from statistical noise.
Recent research has shifted toward modeling techniques that can extract functional insights even when the evolutionary well is dry.
The MSA Bottleneck and the Dark Proteome
The dependency on deep alignments creates a "twilight zone" where sequence identity is too low for traditional tools. As noted by Burkhard Rost, a large fraction of the proteome lacks the homologs necessary for MSA-based prediction. In these scenarios, the statistical power of models that rely on residue-residue correlations (like Potts models) evaporates.
To bridge this gap, researchers are moving away from family-specific models toward family-agnostic approaches that learn the "universal grammar" of all proteins.
Learning the Grammar of Life with Protein Language Models
Protein Language Models (pLMs) like ProtTrans and SeqVec treat amino acids as words and proteins as sentences. By pre-training on billions of unlabeled sequences, these models learn biophysical principles—such as hydrophobicity and secondary structure tendencies—that apply across the entire tree of life.
The intuition is that even an "orphan" protein must obey the same physical laws as a well-studied one. Models like VELM and VESPA use these learned embeddings to predict mutation effects "zero-shot," bypassing the need for MSAs entirely. This allows for the rapid characterization of proteins that were previously unannotatable.
Retrieval-Augmented Context
While pLMs are powerful, they often benefit from "reminding" the model of similar sequences at the moment of prediction. Pascal Notin and colleagues developed Tranception, which uses an autoregressive transformer combined with a retrieval mechanism.
Instead of building a static alignment beforehand, retrieval-augmented models like RSA search massive databases for similar patterns during inference. This "on-the-fly" context allows the model to find distant evolutionary signals that a rigid MSA pipeline might filter out, significantly improving performance on shallow alignments.
Integrating Structural and Physical Priors
When evolution is silent, physics can speak. If we don't have enough sequences to see how a protein changes, we can look at its 3D shape to understand what it can do.
Models such as SSEmb and S3F integrate sequence data with structural information, such as 3D backbones and surface topology. By grounding the model in the physical reality of the protein's fold, researchers like Lasse M. Blaabjerg have shown that models become much more robust to missing evolutionary data. If a mutation would clash sterically or destabilize a core fold, the model can flag it as deleterious even without seeing a single homologous sequence.
Test-Time Adaptation for Orphans
A final frontier in modeling low-depth proteins is Protein Test-Time Training (ProteinTTT). Developed by Anton Bushuiev, this method fine-tunes a general pLM on the specific target protein during the prediction process itself. By forcing the model to "memorize" the unique patterns of a single orphan sequence, the model reduces its internal uncertainty (perplexity), leading to more accurate predictions of fitness and structure without any external evolutionary context.
By combining these strategies—universal language modeling, intelligent retrieval, structural grounding, and per-sequence adaptation—the field is moving toward a future where "orphan" proteins are no longer left in the dark.