RoseTTAFoldNA

RoseTTAFoldNA extends the original RoseTTAFold design with a dedicated nucleic acid track that operates in parallel with protein-focused tracks. The architecture simultaneously processes information along three dimensions:

• 1D Sequence Track: Encodes amino acid and nucleotide sequences using language model embeddings and positional features
• 2D Distance Track: Predicts pairwise distances and orientation distributions between all residue pairs
• 3D Coordinate Track: Iteratively refines atomic coordinates in real space using SE(3)-equivariant transformations

Information flows bidirectionally between tracks at each iteration, allowing sequence-level evolutionary signals to inform spatial reasoning and vice versa. This integrated design enables the model to jointly reason about protein–DNA and protein–RNA interfaces without treating nucleic acids as an afterthought.

Unlike earlier methods that required separate docking steps or pre-specified binding sites, RoseTettaFoldNA predicts the full 3D structure of protein–nucleic acid complexes directly from sequence. The model accepts:

• Single or multiple protein chains
• Single or multiple DNA/RNA chains
• Arbitrary stoichiometries and chain combinations

During inference, the model generates the complete complex structure in a single forward pass, including protein–DNA interfaces, protein–RNA binding pockets, and nucleic acid tertiary folds. This capability is critical for understanding transcription factor binding, CRISPR-Cas complexes, and ribonucleoprotein assemblies without requiring experimental structural data as input.

The 3D track employs SE(3)-equivariant neural network layers that guarantee physical consistency under rotation and translation. Key properties include:

• Rotational Equivariance: If the input coordinates are rotated, the predicted structure rotates identically—no coordinate frame dependence
• Translational Invariance: Shifting the entire complex in space does not alter predictions
• Iterative Denoising: The model starts from random initial coordinates and progressively refines them through multiple recycling iterations

This equivariant design ensures that predicted protein–nucleic acid interfaces are geometrically plausible and that the model does not learn spurious coordinate-frame artifacts. The approach builds on the SE(3)-Transformer framework adapted for heterogeneous biomolecular systems.

RoseTettaFoldNA is trained to handle both DNA and RNA within the same unified framework, recognizing the distinct structural features of each:

• B-form DNA: Predicts canonical double-helical geometry with major and minor groove features
• A-form RNA: Captures the deeper major groove and wider minor groove characteristic of RNA duplexes
• Non-canonical Interactions: Models Hoogsteen base pairs, base triples, and A-minor motifs that stabilize RNA tertiary structure
• Single-Stranded Regions: Predicts flexible loops and linkers that connect structured elements

The model's training set includes crystallographic structures from the PDB covering diverse protein–DNA and protein–RNA complexes, enabling it to generalize across nucleic acid types without requiring separate models.

RoseTettaFoldNA employs a recycling mechanism where the model's output from one pass becomes the input for the next, enabling progressive refinement over multiple iterations. This approach:

• Improves Accuracy: Each recycling iteration reduces coordinate error, with diminishing returns after 3–4 cycles
• Resolves Ambiguities: Early iterations establish global topology while later iterations refine local geometry and side-chain packing
• Enables Complex Formation: Iterative refinement is particularly important for correctly positioning nucleic acid chains relative to protein binding surfaces

The recycling strategy is computationally efficient compared to running independent models, as intermediate representations are reused and updated rather than recomputed from scratch.

RoseTettaFoldNA outputs per-residue and per-residue-pair confidence estimates that enable users to assess prediction reliability without experimental validation:

• pLDDT (Predicted Local Distance Difference Test): Estimates local coordinate accuracy on a 0–100 scale; residues above 70 are generally reliable, above 90 are high-confidence
• PAE (Predicted Aligned Error): Quantifies expected positional error between residue pairs, useful for assessing relative domain orientations and interface confidence
• Interface Confidence: Specialized metrics for evaluating the reliability of predicted protein–nucleic acid contacts

These metrics are essential for downstream applications such as interpreting binding mechanisms, designing mutagenesis experiments, and filtering predictions for molecular docking studies.

The defining innovation of RoseTTAFoldNA that simultaneously processes information along three parallel tracks:

• 1D Sequence Track: Processes amino acid and nucleotide sequence information using transformer layers to capture evolutionary couplings and residue-type dependencies
• 2D Distance Track: Operates on pairwise distance and orientation matrices, refining inter-residue geometry predictions through iterative attention mechanisms
• 3D Coordinate Track: Directly updates 3D atomic coordinates using SE(3)-equivariant transformers, ensuring physical consistency under rotation and translation

The three tracks exchange information at every iteration, allowing sequence-level constraints to inform spatial geometry and vice versa. This contrasts with AlphaFold 2's two-track design and enables simultaneous modeling of protein and nucleic acid chains.

A SE(3)-invariant loss function central to RoseTTAFoldNA's coordinate track training. FAPE measures the deviation between predicted and ground-truth atomic positions within local residue frames rather than a global coordinate system.

Key properties:

• Computes errors in local reference frames defined by each residue's backbone geometry (Cα, N, C atoms for proteins; C4', P, N1/N9 for nucleic acids)
• Invariant to global rotation and translation, eliminating the need for structural alignment before loss computation
• Penalizes both local backbone geometry errors and long-range tertiary contact inaccuracies
• Enables end-to-end training directly on 3D coordinates without intermediate supervision targets

FAPE loss provides a smoother optimization landscape compared to RMSD-based losses and is critical for achieving high-accuracy nucleic acid structure prediction.

The neural network layer type used in RoseTTAFoldNA's coordinate track that guarantees physical consistency of predicted structures. Unlike standard transformers that operate on scalar features, SE(3)-Transformers process 3D vectors and tensors while respecting rotational and translational symmetries.

Core mechanisms:

• Equivariant attention: Attention weights depend on both feature similarity and relative 3D positions encoded via spherical harmonics
• Tensor product layers: Combine features of different rotation orders (scalars, vectors, higher-order tensors) using Clebsch-Gordan coefficients
• Fiber bundle representations: Features are organized by their transformation properties under SO(3) rotations

This architecture ensures that if the input coordinates are rotated, the predicted structure rotates identically—a mathematical guarantee absent from standard coordinate regression approaches.

RoseTTAFoldNA's specialized capability to model the binding interface between protein chains and DNA/RNA molecules with atomic precision. The model jointly predicts:

• Interface residue pairs: Which amino acids contact which nucleotides across the binding surface
• Hydrogen bond geometry: Specific donor-acceptor distances and angles for protein sidechain-to-nucleobase interactions
• Electrostatic complementarity: Spatial alignment of positively charged protein patches with the negatively charged nucleic acid backbone
• Shape complementarity: Geometric fit between protein binding pockets and nucleic acid grooves (major/minor)

Training on protein-DNA and protein-RNA complexes from the PDB enables the model to learn recognition patterns including zinc finger motifs, helix-turn-helix domains, and RNA recognition motifs (RRMs).

RoseTTAFoldNA performs direct structure prediction without relying on separate secondary structure prediction, fragment assembly, or energy minimization stages. The model:

• Takes raw protein and nucleic acid sequences as sole input (optionally with multiple sequence alignments)
• Iteratively refines both pairwise geometry and 3D coordinates through recycling—feeding outputs back as inputs for multiple refinement cycles
• Produces all-atom coordinates for proteins and nucleic acids in a single forward pass
• Outputs per-residue confidence scores (pLDDT) and predicted aligned errors (PAE) for self-assessment

This end-to-end approach eliminates error propagation from intermediate steps and contrasts with modular pipelines like Rosetta FARFAR2 that require separate secondary structure prediction, fragment picking, and Monte Carlo assembly stages.

RoseTTAFoldNA leverages evolutionary coupling information extracted from Multiple Sequence Alignments (MSAs) of homologous protein and RNA sequences. The MSA processing pipeline:

• Searches genetic sequence databases (UniRef, Rfam, NCBI nt) to identify homologs for each input chain
• Constructs paired MSAs that capture co-evolutionary signals—correlated mutations indicating spatial proximity or functional interaction
• Processes MSA rows through axial attention mechanisms that attend across both sequence positions and alignment depth
• Handles the challenge of limited RNA sequence databases by incorporating metagenomic data and using RNA language model embeddings as supplementary features

For RNA targets with few known homologs, the model relies more heavily on the single-sequence track and learned structural priors from self-supervised pre-training.