aminx weights β€” v0.1.0a1

JAX/Equinox reimplementation of ProteinMPNN and LigandMPNN.

Weights are converted from the original PyTorch checkpoints and verified to β‰₯0.999 Pearson correlation with the reference implementation (atol/rtol 1e-4 for all families; side-chain packer: atol 1e-4/rtol 1e-3).

Usage

pip install aminx   # huggingface-hub is a required dep; weights download on first use
from aminx.io.weights import load_model

model = load_model("proteinmpnn_v_48_020")
model = load_model("ligandmpnn_v_32_020_25")
model = load_model("solublempnn_v_48_020")
model = load_model("per_residue_label_membrane_mpnn_v_48_020")
model = load_model("ligandmpnn_sc_v_32_002_16")

Weights are cached to ~/.cache/huggingface/hub/ after first download.

Checkpoint topology

All families share: node/edge/hidden features = 128, encoder layers = 3, decoder layers = 3, vocab size = 21. The suffix digits in checkpoint names encode training noise level (e.g. _020 = 0.20 Γ… backbone noise, per upstream convention) and atom context count (e.g. _25, _16).

Family Checkpoints k_neighbors num_positional_embeddings Notes
ProteinMPNN proteinmpnn_v_48_{002,010,020,030} 48 32 Original backbone
SolubleMPNN solublempnn_v_48_{002,010,020,030} 48 32 Solubility-biased variant
LigandMPNN ligandmpnn_v_32_{005,010,020,030}_25 32 32 2 ligand-context layers; atom_context_num=25
Membrane {global,per_residue}_label_membrane_mpnn_v_48_020 48 32 physics_feature_dim=3
Side-chain packer ligandmpnn_sc_v_32_002_16 32 16 atom_context_num=16

Deprecation notice

The previous maraxen/prxteinmpnn repository is archived. Use this repo going forward.

Citation

If you use these weights, please cite the original works:

@article{dauparas2022robust,
  title={Robust deep learning-based protein sequence design using ProteinMPNN},
  author={Dauparas, Justas and others},
  journal={Science},
  year={2022}
}

@article{dauparas2025atomic,
  title={Atomic context-conditioned protein sequence design using LigandMPNN},
  author={Dauparas, Justas and others},
  journal={Nature Methods},
  volume={22},
  number={4},
  pages={717--723},
  year={2025}
}

Family checkpoints (aminx 0.2.1)

Converted copies of third-party weights, redistributed under their upstream MIT licences. Each upstream copyright and permission notice is in NOTICE-weights.md in the aminx source tree and reproduced below. These are the upstream parameters re-serialised as Equinox .eqx; they are not retrained. Parity of each against its upstream is measured per path and is not complete (see the aminx parity ledger).

checkpoint id upstream upstream file
pottsmpnn_vanilla_20 PottsMPNN, KeatingLab/PottsMPNN @ 0cb0a588 pottsmpnn_vanilla_20.eqx
pottsmpnn_vanilla_30 PottsMPNN, KeatingLab/PottsMPNN @ 0cb0a588 pottsmpnn_vanilla_30.eqx
pottsmpnn_sol_20 PottsMPNN, KeatingLab/PottsMPNN @ 0cb0a588 pottsmpnn_sol_20.eqx
pottsmpnn_sol_30 PottsMPNN, KeatingLab/PottsMPNN @ 0cb0a588 pottsmpnn_sol_30.eqx
pottsmpnn_ft PottsMPNN, KeatingLab/PottsMPNN @ 0cb0a588 pottsmpnn_ft.eqx
lasermpnn_nothing_heldout LASErMPNN, polizzilab/LASErMPNN @ e70f2c6d lasermpnn_nothing_heldout.eqx
lasermpnn_noise_ligandmpnn_split LASErMPNN, polizzilab/LASErMPNN @ e70f2c6d lasermpnn_noise_ligandmpnn_split.eqx
lasermpnn_soluble_65000 LASErMPNN, polizzilab/LASErMPNN @ e70f2c6d lasermpnn_soluble_65000.eqx
protonpottsmpnn_v6_30 Proton-PottsMPNN, epoch-0125 checkpoint, pin 09682abf protonpottsmpnn_v6_30.eqx

Load by id with aminx (--checkpoint-id <id>); every file is pinned by sha256 in aminx/io/family_weights.json. Each .eqx has a .json manifest next to it that must be kept with it (it records static fields the .eqx format does not store, for example LASErMPNN's build_hydrogens).

Upstreams: PottsMPNN (https://github.com/KeatingLab/PottsMPNN, MIT, (c) 2022 Justas Dauparas, (c) 2026 Foster Birnbaum, Amy E. Keating), LASErMPNN (https://github.com/polizzilab/LASErMPNN, MIT, (c) 2025 Benjamin Fry, Nicholas Polizzi), Proton-PottsMPNN (MIT, (c) 2026 Christian P. Jacobsen). Please cite the upstream publications if you use these weights.

MPNN Variant & ThermoMPNN Checkpoints

Converted copies of third-party weights (AbMPNN, HyperMPNN, CyclicMPNN, CA-Only ProteinMPNN, ThermoMPNN v1/v2, and ThermoMPNN-D), redistributed under their upstream MIT / CC-BY-4.0 licences (see NOTICE-weights.md). Verified against upstream PyTorch checkpoints to < 1e-7 in float64 and < 1e-4 in float32.

Checkpoint ID Tier / Family Upstream Artifact
abmpnn_v_48_020 Tier 1 (AbMPNN) Dreyer et al., 2023 (Zenodo 10.5281/zenodo.8164693, CC BY 4.0) abmpnn_v_48_020.eqx.zst
hypermpnn_v_48_{002,010,020,030} Tier 1 (HyperMPNN) Ertelt et al., 2024 (meilerlab/HyperMPNN, MIT) hypermpnn_v_48_{002,010,020,030}.eqx.zst
cyclicmpnn_v_48_010 Tier 1 (CyclicMPNN) Beaudoin et al., 2025 (Graylab/CyclicMPNN, MIT) cyclicmpnn_v_48_010.eqx.zst
ca_v_48_{002,010,020} Tier 2 (CA-Only ProteinMPNN) Dauparas et al., 2022 (dauparas/ProteinMPNN, MIT) ca_v_48_{002,010,020}.eqx.zst
thermo_mpnn_v1_default Tier 3 (ThermoMPNN v1) Dieckhaus et al., 2024 (Kuhlman-Lab/ThermoMPNN, MIT) thermo_mpnn_v1_default.eqx.zst
thermo_mpnn_v2_ens{1,2,3} Tier 3 (ThermoMPNN v2) Dieckhaus et al., 2025 (Kuhlman-Lab/ThermoMPNN-D, MIT) thermo_mpnn_v2_ens{1,2,3}.eqx.zst
thermo_mpnn_d_ens{1,2,3} Tier 3 (ThermoMPNN-D) Dieckhaus et al., 2025 (Kuhlman-Lab/ThermoMPNN-D, MIT) thermo_mpnn_d_ens{1,2,3}.eqx.zst

Usage

from aminx.io.weights import load_model
from aminx.families.thermo_mpnn import (
  load_thermo_mpnn_checkpoint,
  load_thermo_mpnn_double_checkpoint,
)

# Tier 1 & Tier 2
ab_model = load_model("abmpnn_v_48_020")
hyper_model = load_model("hypermpnn_v_48_020")
cyclic_model = load_model("cyclicmpnn_v_48_010")
ca_model = load_model("ca_v_48_020")

# Tier 3 (ThermoMPNN & ThermoMPNN-D)
thermo_v1 = load_thermo_mpnn_checkpoint("thermo_mpnn_v1_default")
thermo_v2 = load_thermo_mpnn_checkpoint("thermo_mpnn_v2_ens1")
thermo_d = load_thermo_mpnn_double_checkpoint("thermo_mpnn_d_ens1")

Follow-Up Family Checkpoints (AntiBMPNN, Caliby, FAMPNN)

Checkpoint ID Tier / Family Upstream Artifact
antibmpnn_v_48_{000,010,020,030} Tier 1B (AntiBMPNN) Goudy et al., 2024 (oda-lab/AntiBMPNN, Zenodo 10.5281/zenodo.13387792, CC BY 4.0 / MIT) antibmpnn_v_48_{000,010,020,030}.eqx.zst
caliby, soluble_caliby, soluble_caliby_v1, caliby_distill, soluble_caliby_distill Tier 4 (Caliby Potts) Shuai, Widatalla, Huang et al., 2025/2026 (ProteinDesignLab/caliby, Zenodo 10.5281/zenodo.18644985, MIT) {caliby,soluble_caliby,soluble_caliby_v1,caliby_distill,soluble_caliby_distill}.eqx.zst
caliby_packer_{000,010,030} Tier 4 (Caliby Sidechain Diffusion Packer) Shuai, Widatalla, Huang et al., 2025/2026 (ProteinDesignLab/caliby, Zenodo 10.5281/zenodo.18644985, MIT) caliby_packer_{000,010,030}.eqx.zst
fampnn_{0_0,0_3,0_3_cath} Tier 4 (Full-Atom MPNN) Shuai, Widatalla, Possu et al., ICML 2025 (richardshuai/fampnn, MIT) fampnn_{0_0,0_3,0_3_cath}.eqx.zst
megascale_default_la-true_esm-False_split1 FrustraMPNN (RosettaCommons, 2026), https://github.com/RosettaCommons/frustraMPNN @ 3a03cdc300bfe24c4bb70e60207118532bc73b3b (BSD-3-Clause) model_weights/megascale_default_la-true_esm-False_split1.ckpt + v_48_020.pt
Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. πŸ™‹ Ask for provider support