ERA Directed Evolution

Efficient, few-shot protein discovery with energy rank alignment.

Sebastian Ibarraran, Shriram Chennakesavalu, Frank Hu, and Grant M. Rotskoff

ERA aligns learned sequence energies with experimental fitness rankings, enabling fast, data-efficient design loops. Use it to train geometric transformers, sample sequences, and map full fitness landscapes from limited data.

Run the code Download the data Read the paper
ERA schematic

Preference Optimization for Few-Shot Experimental Design

Method DHFR GB1 ParD2 ParD3 TrpB4
MLDE 0.96 0.67 0.99 0.90 0.78
ALDE 0.96 0.80 0.99 0.99 0.89
EVOLVEpro (ESM2) 0.95 0.81 1.0 0.99 0.88
EVOLVEpro (ESM3) 1.0 0.77 1.0 0.99 0.79
DPO 1.0 0.78 1.0 0.99 0.81
ERA 1.0 ± 0.00 0.83 ± 0.04 1.0 ± 0.00 1.0 ± 0.00 0.91 ± 0.02
Average maximum fitness achieved across multiple replicate runs with standard errors included. All methods used 384 total samples for training, with ALDE, DPO, and ERA doing so over four rounds (96 per round). For MLDE and ALDE, 50 replicates were averaged and values were obtained from Li et al. (2025), whereas 10 replicates were averaged for ERA, DPO, and EVOLVEpro.

Energy rank alignment reframes few-shot protein engineering as preference optimization. Let $f(x)$ be a learned energy for sequence $x$. For any observed pair $(x_i, x_j)$ where $x_i$ is experimentally preferred to $x_j$, we enforce $f(x_i) < f(x_j)$ using a smooth pairwise loss. We optimize $$ \mathcal{L}_{\mathrm{ERA}}(\pi_\theta) = \mathbb{E}_{x \sim \mathcal{D},\, y,y' \sim \pi_{\mathrm{ref}}(\cdot \mid x)} \!\left[ D_{\mathrm{KL}}\!\left( p_\gamma(y \succ y') \,\|\, p_\theta \right) \right]. $$ which penalizes inversions while remaining stable under limited data. Optimizing this ranking loss aligns energies with relative fitness ordering instead of absolute calibration, improving generalization when only a handful of measurements are available.

Train using your data

GitHub repository

Environment setup

uv sync
uv pip install -e .

Train a model

pera_train \
  train.data_root_path=./data \
  train.data=GB1 \
  train.trainer_args.max_epochs=100 \
  train.trainer_args.devices=1

Sample sequences

pera_sample \
  infer.target=GB1 \
  infer.num_samples=512 \
  infer.network_filename=/path/to/checkpoint.pt

Compute landscape

pera_compute_landscape \
  compute_landscape.data=GB1 \
  compute_landscape.data_root_path=./data