Prescient Design / Genentech
SE(3)-equivariant protein structure prediction model using a novel coarse-grained representation for orders-of-magnitude faster inference.
Accurate protein structure prediction is a cornerstone capability for modern computational biology and drug discovery, but the dominant approaches — led by AlphaFold 2 and RoseTTAFold — impose computational costs that limit their use in high-throughput workflows. Both systems depend on generating and processing multiple sequence alignments (MSAs) and often leverage pre-computed protein language model embeddings, making each prediction a multi-minute to multi-hour operation even on modern GPU hardware. For applications such as protein design, where a researcher may need to evaluate tens of thousands of candidate sequences, this latency creates a practical bottleneck that prevents structure-informed design from being integrated into iterative optimization loops.
EquiFold, developed by researchers at Prescient Design (Genentech), addresses this bottleneck through two complementary innovations. First, it introduces a novel coarse-grained structure representation that dramatically reduces the computational cost of representing and processing protein geometry. Second, it applies SE(3)-equivariant neural network operations — computations that are guaranteed to produce physically consistent outputs regardless of how the input structure is rotated or translated in three-dimensional space — directly to this reduced representation, enabling end-to-end differentiable all-atom structure prediction without the MSA preprocessing step. The result is a model that achieves accuracy comparable to AlphaFold 2 on standard benchmarks while running orders of magnitude faster, making it practical to integrate structure prediction as an inline component in protein design pipelines.
Released as a bioRxiv preprint in October 2022, EquiFold was developed by Jae Hyeon Lee, Payman Yadollahpour, Andrew Watkins, Nathan C. Frey, Andrew Leaver-Fay, Stephen Ra, Kyunghyun Cho, Vladimir Gligorijević, Aviv Regev, and Richard Bonneau. The model's architecture is substantially smaller than AlphaFold 2, whose 93 million parameters are concentrated in the Evoformer's MSA processing stack. By eliminating the MSA requirement entirely and replacing it with a geometrically informed coarse-grained representation, EquiFold achieves a compelling speed-accuracy trade-off that positions it as the preferred choice for structure-intensive design applications where prediction throughput matters as much as absolute accuracy.
EquiFold's architecture begins with a sequence encoder that produces per-residue embeddings from the input amino acid sequence. These embeddings, combined with positional information, are used to initialize a coarse-grained structure representation in which each residue is associated with a small set of geometrically defined frames rather than a full set of atomic coordinates. This coarse-grained representation is processed by a series of SE(3)-equivariant graph neural network layers, where edges connect residues within a spatial cutoff distance and messages are passed between residues using operations that respect the rotational and translational symmetry of three-dimensional space. The network iteratively refines the positions and orientations of the coarse-grained frames through multiple rounds of message passing, gradually converging on a self-consistent geometric configuration.
After the coarse-grained structure converges, a lightweight all-atom reconstruction module maps the refined frames back to explicit atomic coordinates by predicting backbone and side-chain torsion angles, producing a complete all-atom protein structure suitable for direct use in downstream applications. The model does not use protein language model embeddings as input — a deliberate design choice that avoids both the computational cost of running a language model and any potential distributional mismatch between language model training data and the structure prediction task. The absence of MSA inputs means that EquiFold relies entirely on the geometric inductive biases of SE(3)-equivariant operations and the sequence-to-structure information encoded in the training dataset.
Training data was drawn from the Protein Data Bank (PDB), following standard practices for structure prediction model training. Benchmark evaluations on CAMEO and CASP targets demonstrated that EquiFold achieves accuracy comparable to AlphaFold 2 on many protein families, with the largest accuracy gaps appearing on proteins where MSA depth provides critical co-evolutionary information. The model excels on globular proteins and antibody structures, where the coarse-grained representation captures the essential geometry without loss of accuracy, and where the speed advantage is most valuable given the high throughput requirements of antibody design campaigns.
EquiFold's primary use case is high-throughput structure prediction in protein design workflows, where a design algorithm generates large numbers of candidate sequences that need to be filtered by predicted structural quality, stability, or target-binding pose. In computational antibody design, for example, a genetic algorithm or Monte Carlo sampler might generate thousands of CDR sequences per hour; EquiFold can evaluate the predicted structure of each candidate in seconds, enabling real-time structural filtering. The model is also valuable for protein property prediction tasks where structural features are better predictors than sequence features alone: by running EquiFold on all variants in a dataset, researchers can extract structural embeddings or compute geometric descriptors that serve as inputs to downstream predictors of solubility, thermostability, or aggregation propensity. Because EquiFold is end-to-end differentiable, it can be chained with a property predictor and optimized jointly through gradient-based sequence design, a modality that is not accessible with non-differentiable structure prediction tools. Researchers building protein-protein docking workflows also benefit from EquiFold's speed when predicting receptor structures for use as rigid docking targets against large numbers of designed binders.
EquiFold established that the MSA preprocessing step — widely assumed to be essential for competitive structure prediction accuracy — could be eliminated without catastrophic accuracy loss if the architecture incorporated sufficient geometric inductive biases. This result was influential in the subsequent development of MSA-free structure prediction methods and contributed to a broader recognition that co-evolutionary information, while valuable, is not strictly necessary when the model architecture is designed to maximize its use of sequence-to-structure relationships. The model's differentiability and speed have made it a practical tool for structure-in-the-loop protein design, a modality that was theoretically attractive but computationally impractical before EquiFold's release. A key limitation is that accuracy on sequences with distant homologs — where MSA depth provides irreplaceable co-evolutionary signal — lags behind AlphaFold 2, and the model may not generalize as well to novel fold families that are underrepresented in the PDB training set. The coarse-grained representation also introduces a small accuracy penalty on very precise side-chain placement tasks compared to models that operate directly on full atomic coordinates.