Addressing the Tertiary Structure Availability Gap

17 concepts16 papers1 proposal

The success of deep learning in structural biology has been profoundly uneven. While architectures like AlphaFold revolutionized protein structure prediction by capitalizing on the hundreds of thousands of experimentally determined structures in the Protein Data Bank, other critical biomolecules remain stranded in a data desert.

As highlighted by Shen et al., RNA structures constitute less than one percent of public structural databases. Similarly, Taubert et al. point out that the entire transcriptome yields only a few dozen non-redundant RNA families, compared to tens of thousands available for proteins. This scarcity extends to short peptides as well; the authors of a recent peptide design study note that peptides represent less than 0.2% of solved structures.

Without massive datasets, standard end-to-end deep learning models fail to generalize, biasing toward a few common motifs or predicting physically impossible conformations. To overcome this chronic data bottleneck, researchers have developed a distinct playbook: leveraging abundant sequence data, transferring structural intuition across molecular domains, and embedding strict geometric and physical priors directly into model architectures.


Bootstrapping Structure from Evolutionary Sequences

While high-resolution 3D coordinates are difficult and expensive to obtain experimentally, raw sequence data is abundant. The core intuition behind sequence-based bootstrapping is that evolution acts as an implicit structural filter. For a biomolecule to function, it must fold stably; therefore, the statistical patterns and co-evolutionary mutations embedded within millions of unannotated sequences contain latent geometric constraints.

[Massive Unlabeled Sequence Database] ──> [Language Model Pretraining] ──> [Learned Evolutionary Priors]
                                                                                      │
[Scarce 3D Structural Templates]      ────────────────────────────────────────────────┴─> [Accurate 3D Structure/Design]

To predict RNA 3D structures, Shen et al. developed RhoFold+, which pairs a geometry-aware transformer with a massive RNA language model pretrained on 23.7 million sequences. The language model captures long-range dependencies and structural "blueprints" before the geometric module ever processes a coordinate.

This strategy is equally effective for the inverse problem—generating sequences that fold into a target structure. Jiao et al. introduced EVO-RDesign, which injects evolutionary priors from pretrained RNA language models into an inverse folding network. By anchoring the model's search space within evolutionarily viable sequence distributions, the framework compensates for the lack of explicit 3D training templates and significantly improves sequence recovery.


Cross-Domain Knowledge Transfer

Because the physical forces driving molecular folding—such as electrostatic interactions, hydrogen bonding, and hydrophobic packing—are universal, models trained on data-rich domains can transfer their structural intuition to data-scarce domains.

[Data-Rich Domain: Proteins] ───> Learn Universal Biophysical Principles ───> [Data-Scarce Domain: RNA/Peptides]
                                      (via Translation or Fine-Tuning)

To predict RNA contact maps, Gao et al. developed CoT-RNA, a framework that translates RNA sequences into a protein-like "language." This allows a high-capacity model trained on massive protein datasets to map RNA interactions, bypassing the lack of native RNA training samples.

A similar logic applies to protein-RNA interactions. When predicting RNA-binding proteins (RBPs) across under-studied species, Peng et al. utilized a two-stage transfer learning framework (RBP-TSTL). The model extracts generalized biochemical features from a self-supervised protein language model before fine-tuning on sparse, highly specific RBP annotations, preventing overfitting on small organismal datasets.


Hardcoding Geometric and Physical Priors

Standard neural networks require vast amounts of data to learn that molecules cannot self-intersect or that certain bond angles are rigid. When data is scarce, these rules must be hardcoded into the model's architecture or generative process to restrict it to physically plausible outputs.

Thermodynamic Ensembles

Rather than treating a biomolecule as a static set of coordinates, models can incorporate its dynamic flexibility. For protein-RNA binding prediction, Su et al. developed ThermoNet, which integrates sequence embeddings with a "thermodynamic ensemble"—a mathematical collection of many possible ways an RNA molecule can fold. Capturing this structural "wobble" allows the model to learn robust binding rules from short, relatively unstructured training sequences.

Geometric Algebra and Flow Matching

For de novo design, researchers use advanced spatial mathematics to enforce structural realism. Anand et al. adapted $\text{SE}(3)$ flow matching to RNA backbones in RNA-FrameFlow, ensuring that the generative model respects the complex, flexible coordinate frames of RNA without biasing toward a handful of known training motifs. To better capture 3D spatial relationships, Wang et al. combined Geometric Algebra with Bayesian Flow Networks, enforcing rigorous spatial constraints directly within the network's layers.

Pre-trained Denoisers

Instead of training a generative model from scratch, researchers can embed an entire pre-trained structure predictor as a physical guide. Li et al. developed RiboDiff, a joint diffusion model for RNA sequence-structure co-design that repurposes the RoseTTAFold2NA structure predictor as its denoiser. The diffusion process is effectively guided by the deep molecular priors already memorized by the predictor, ensuring physical consistency despite the scarcity of training data.


Data-Efficient Representations and Regularization

A final line of defense against data scarcity is optimizing how data is represented and how the model is penalized during training.

  • Discrete Tokenization: Rather than forcing a model to learn continuous, noisy 3D coordinate distributions from a few samples, Zhang et al. introduced RiboSphere, which tokenizes geometric data into discrete, interpretable "structural tokens." This quantization simplifies the learning objective, stabilizing downstream tasks like inverse folding when data is limited.
  • Latent Space Alignment: For short peptides, Wang et al. developed LSSAMP, which maps amino acid sequences and secondary structures (such as alpha-helices) into a shared, vector-quantized latent space. Aligning these representations allows the model to generate novel antimicrobial peptides that are structurally predisposed to biological activity without requiring high-resolution 3D training data.
  • Diversity Regularization: When fine-tuning models on scarce data, architectures often suffer from mode collapse, generating a narrow set of safe sequences. To counteract this in peptide design, the authors of a diversity-regularized framework adapted Direct Preference Optimization (DPO) with online diversity rewards. This forces the model to explore a wider structural and sequence space, preventing it from overfitting to the tiny fraction of peptide structures available in public databases.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_structural_data_scarcity_biomolecular_da_93541343,
  title = {structural data scarcity (biomolecular data imbalance)},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/structural-data-scarcity-biomolecular-data-imbalance/rdr_93541343b94a4c9da876201d92da89b8},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-21},
}

This exact version (version 1, frozen text):

@misc{lacuna_structural_data_scarcity_biomolecular_da_93541343_v1,
  title = {structural data scarcity (biomolecular data imbalance)},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/structural-data-scarcity-biomolecular-data-imbalance/rdr_93541343b94a4c9da876201d92da89b8/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-21},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake