ARIES model weights

ARIES predicts enzyme-conditioned reaction sites on substrate atoms using exact GROVER atom and directed-bond representations, ESM-C 6B residue representations, and a reaction-template representation.

Source code, installation instructions, and the aries-inference command are available in the ARIES GitHub repository. Example data and precomputed embeddings are distributed separately in xuan-liu/ARIES-data.

Available weights

Every directory contains model_info.json, seed42.pt, seed43.pt, and seed44.pt. Keep the directory names unchanged: ARIES uses the metadata to validate architecture, ensemble seeds, and LoRA/base compatibility.

Model	Weight type	Required base	Directory
ARIES-XS random	Full	—	aries-random-xs/
ARIES-S random	Full	—	aries-random-s/
ARIES-M random	Full	—	aries-random-m/
ARIES-L random	Full	—	aries-random-l/
ARIES-XL random	Full	—	aries-random-xl/
ARIES-XL-Brenda	Full	—	aries-xl-brenda/
ARIES-XL-Hal-LoRA	LoRA adapter	aries-xl-brenda/	aries-xl-hal-lora/
P450 specialist	Random-split LoRA adapter	aries-random-xl/	specialist_hub/p450-random/
Glycosyltransferase specialist	Random-split LoRA adapter	aries-random-xl/	specialist_hub/gt-random/
Methyltransferase specialist	Random-split LoRA adapter	aries-random-xl/	specialist_hub/mt-random/
Fatty-acid-desaturase specialist	Random-split LoRA adapter	aries-random-xl/	specialist_hub/fad-random/
Prenyltransferase specialist	Random-split LoRA adapter	aries-random-xl/	specialist_hub/pt-random/

The five specialist-hub adapters are the canonical random-split models. They use rank 16 and alpha 32 and must be combined with the seed-matched ARIES-XL random checkpoints. ARIES-XL-Hal-LoRA must instead be combined with the seed-matched ARIES-XL-Brenda checkpoints.

Adapter files contain only LoRA A/B tensors, model configuration, seed, required-base metadata, rank, alpha, and target manifests. They do not duplicate the frozen base weights.

Installation

Clone ARIES, create the dedicated conda environment, and install the package:

git clone https://github.com/xuanliugit/ARIES.git
cd ARIES
conda create -n aries python=3.12 -y
conda activate aries
python -m pip install --upgrade pip
python -m pip install -e ./aries

Install the current Hugging Face CLI if it is not already available:

python -m pip install --upgrade huggingface-hub

Download

Download the complete repository into model_weights/:

hf download xuan-liu/ARIES --local-dir model_weights

Alternatively, download only the directories needed for one prediction task. For ARIES-XL-Brenda with the Halogenase adapter:

hf download xuan-liu/ARIES \
  --include "aries-xl-brenda/*" "aries-xl-hal-lora/*" \
  --local-dir model_weights

For ARIES-XL random with the P450 specialist adapter:

hf download xuan-liu/ARIES \
  --include "aries-random-xl/*" "specialist_hub/p450-random/*" \
  --local-dir model_weights

Inference

Base-only inference selects the model solely through --weights-dir:

conda activate aries
aries-inference \
  --input-csv data/example_data/experimental_inputs.csv \
  --embedding-dir data/example_data \
  --weights-dir model_weights/aries-xl-brenda \
  --output-csv predictions_brenda.csv \
  --device auto

Add an optional compatible adapter with --lora-weights-dir:

aries-inference \
  --input-csv data/example_data/experimental_inputs.csv \
  --embedding-dir data/example_data \
  --weights-dir model_weights/aries-xl-brenda \
  --lora-weights-dir model_weights/aries-xl-hal-lora \
  --output-csv predictions_hal.csv \
  --device auto

ARIES loads seeds 42–44 sequentially, applies the seed-matched adapter when provided, averages atom probabilities, and writes atom indices, probabilities, and ranks to the output CSV.

License

ARIES model weights are released under the MIT License.