PLASMA: Pluggable Local Alignment via Sinkhorn MAtrix

PLASMA is a tiny, pluggable head that turns any frozen protein-language-model (PLM) into a residue-level local aligner. It reformulates protein substructure alignment as a regularised optimal transport problem and runs ~50× faster than structure-based aligners (TM-Align, Foldseek) by operating on pre-computed embeddings.

This repository hosts the trained PLASMA heads for every (task, backbone) combination from the paper, plus instructions for the parameter-free PLASMA-PF baseline (which has no learned weights).

• Paper: https://arxiv.org/abs/2510.11752
• Code: https://github.com/ZW471/PLASMA-Protein-Local-Alignment
• License: MIT

What's in this repo

Each variant lives in its own subfolder and is loaded by the load_plasma helper from the GitHub package:

weights/
  active_site/
    prot_bert/                 # config.json + model.safetensors + metadata.json
    ankh-base/
    TM-Vec/
    ProstT5/
    prot_t5_xl_half_uniref50-enc/
    esm2_t33_650M_UR50D/
    ProtSSN/
  binding_site/
    ...
  motif/
    ...

All heads share the same architecture: a small LRL non-linearity (LazyLinear → ReLU → Linear → LayerNorm, hidden dim 512) followed by a parameter-free Sinkhorn iteration (temperature=0.1, n_iters=20). The checkpoint files are ~3 MB each.

How to use

PLASMA is a head: it consumes per-residue embeddings from a frozen protein language model and returns a soft alignment matrix between two sub-structures. The end-to-end pipeline is therefore three steps:

1. Embed each protein with the backbone the head was trained on (one of the seven listed above).
2. Run the PLASMA head on the (residue × residue) embeddings to get a soft alignment matrix M ∈ [0, 1]^{n_q × n_c}.
3. Optionally reduce M to a scalar similarity score with utils.alignment_score.

1. Install

git clone https://github.com/ZW471/PLASMA-Protein-Local-Alignment
cd PLASMA-Protein-Local-Alignment
uv sync

The Alignment class and the load_plasma helper live in the alignment package shipped by that repo.

2. Load a trained head

from alignment import load_plasma

# task ∈ {"active_site", "binding_site", "motif"}
# backbone is the PLM whose embeddings the head was trained on
model = load_plasma(task="active_site", backbone="esm2_t33_650M_UR50D")
model.eval()

load_plasma downloads the matching config.json + model.safetensors from this repo via huggingface_hub and rebuilds the Alignment module.

3. Compute embeddings with the matching backbone

PLASMA does not embed sequences itself. The example below shows how to do it with ESM-2 via transformers; the same pattern works for any other backbone (Ankh, ProstT5, ProtBERT, ProtT5, TM-Vec, ProtSSN — their loaders are documented in embed.py in the GitHub repo).

import torch
from transformers import AutoTokenizer, AutoModel

device = "cuda" if torch.cuda.is_available() else "cpu"

tokenizer = AutoTokenizer.from_pretrained("facebook/esm2_t33_650M_UR50D")
backbone = AutoModel.from_pretrained("facebook/esm2_t33_650M_UR50D").to(device).eval()

@torch.no_grad()
def embed(sequence: str) -> torch.Tensor:
    """Return per-residue embeddings of shape (L, 1280) — no special tokens."""
    tokens = tokenizer(sequence, return_tensors="pt", add_special_tokens=True).to(device)
    h = backbone(**tokens).last_hidden_state[0]   # (L+2, 1280): <cls> ... <eos>
    return h[1:-1].cpu()                          # drop <cls> and <eos>

seq_q = "MKTAYIAKQRQISFVKSHFSRQDILDLWIYHTQGYFP"
seq_c = "MKTVRQERLKSIVRILERSKEPVSGAQLAEELSVSRQVIVQDIAYLRSLGYNI"

H_q = embed(seq_q)            # (n_q, 1280)
H_c = embed(seq_c)            # (n_c, 1280)

4. Run PLASMA and read the alignment matrix

# `batch_q` / `batch_c` assign each residue to a sample. Use zeros for a
# single pair; use [0, 0, ..., 1, 1, ...] to score multiple pairs in one batch.
batch_q = torch.zeros(H_q.size(0), dtype=torch.long)
batch_c = torch.zeros(H_c.size(0), dtype=torch.long)

with torch.no_grad():
    M = model(H_q, H_c, batch_q, batch_c)        # (n_q, n_c) in [0, 1]

# Hard residue-residue assignment (top of column / row in the transport plan)
q_to_c = M.argmax(dim=1)        # for each query residue, the best candidate residue
c_to_q = M.argmax(dim=0)        # for each candidate residue, the best query residue

M is a (near-)doubly-stochastic transport plan: rows and columns each sum to ~1, so M[i, j] is the soft probability that query residue i aligns to candidate residue j. Thresholding at 0.5 gives a sparse local alignment; plotting M as a heatmap gives the canonical PLASMA visualisation (the diagonal stripe in the visual abstract above).

5. Reduce to a similarity score

To collapse the alignment matrix into a single number per protein pair (the quantity used to compute ROC-AUC / F1-Max in the tables above), use utils.alignment_score from the GitHub repo. It applies the diagonal convolution + thresholding described in the paper:

from utils.alignment_utils import alignment_score

score = alignment_score(
    H_q, H_c, M, batch_c,
    threshold=0.5,           # gating on max-row / max-col residues
    K=10,                    # diagonal-convolution window
)                            # -> shape (num_pairs_in_batch,), here (1,)
print(float(score))

PLASMA-PF (parameter-free)

PLASMA-PF is a hinge / Sinkhorn baseline with no learned weights. Use it when you want a strong zero-training baseline on top of any backbone — there is nothing to download:

from alignment import load_plasma_pf

model = load_plasma_pf()         # Alignment(eta='hinge', omega='sinkhorn', ...)

with torch.no_grad():
    M_pf = model(H_q, H_c, batch_q, batch_c)

It accepts the same forward signature as the trained heads above and pairs with any of the seven supported backbones.

Available variants & evaluation results

Numbers below are 3-seed averages reported in the paper. The seven backbone columns correspond to the seven subfolders under each task. Bold marks the best backbone for each row.

Interpolation (in-distribution test split)

Task	Metric	Ankh	ESM-2	ProstT5	ProtBERT	ProtSSN	ProtT5	TM-Vec
Motif	ROC-AUC	.925	.933	.954	.854	.922	.972	.910
	F1-Max	.885	.877	.885	.784	.866	.918	.853
	PR-AUC	.921	.931	.953	.872	.920	.971	.914
	Label Match Score	.921	.890	.929	.746	.767	.937	.792
Binding Site	ROC-AUC	.995	.992	.993	.981	.992	.993	.980
	F1-Max	.987	.986	.983	.948	.982	.988	.970
	PR-AUC	.996	.994	.995	.985	.993	.995	.984
	Label Match Score	.951	.950	.951	.880	.872	.951	.900
Active Site	ROC-AUC	.994	.991	.993	.986	.992	.994	.991
	F1-Max	.989	.985	.987	.967	.987	.987	.982
	PR-AUC	.994	.992	.994	.988	.994	.994	.992
	Label Match Score	.975	.969	.975	.904	.885	.972	.938

Extrapolation (held-out hard test split)

Task	Metric	Ankh	ESM-2	ProstT5	ProtBERT	ProtSSN	ProtT5	TM-Vec
Motif	ROC-AUC	.960	.972	.975	.870	.949	.968	.954
	F1-Max	.915	.931	.926	.799	.896	.922	.903
	PR-AUC	.948	.970	.969	.873	.940	.962	.944
	Label Match Score	.842	.786	.801	.541	.537	.738	.704
Binding Site	ROC-AUC	.995	.999	.993	.951	.999	.999	.990
	F1-Max	.992	.991	.985	.896	.988	.996	.983
	PR-AUC	.997	.999	.995	.958	.998	.999	.992
	Label Match Score	.894	.851	.891	.603	.753	.902	.824
Active Site	ROC-AUC	.995	.996	.996	.980	.997	.999	.995
	F1-Max	.992	.986	.991	.950	.991	.991	.985
	PR-AUC	.995	.997	.997	.984	.998	.999	.996
	Label Match Score	.938	.882	.931	.697	.737	.893	.880

Each subfolder also contains a metadata.json with the full hyperparameter config in machine-readable form.

Training details

• Architecture: Alignment(eta='lrl', omega='sinkhorn', eta_kwargs={'hidden_dim': 512}, omega_kwargs={'temperature': 0.1, 'n_iters': 20}).
• Score head: K=10, threshold=0.5 (used by utils.alignment_score to reduce the transport plan to a scalar).
• Optimiser / loss: Adam (lr=1e-4), BCEWithLogitsLoss on the alignment score plus a label-match auxiliary loss (target_loss_weight=1.0).
• Data: the InterPro-derived motif / binding-site / active-site datasets shipped under data/raw/ in the GitHub repo, split into train / validation / test / test-hard with dataset_fraction=0.1 (default sweep) and dataset_fraction=1.0 (full sweep — checkpoints here are from the full sweep).
• Selection metric: validation loss (early stopping, patience=3).

Citation

If you use these weights, please cite the PLASMA paper:

@inproceedings{wang2026plasma,
  title     = {Fast and Interpretable Protein Substructure Alignment via Optimal Transport},
  author    = {Wang, Zhiyu and Zhou, Bingxin and Wang, Jing and Tan, Yang and Zhao, Weishu and Li{\`o}, Pietro and Hong, Liang},
  booktitle = {International Conference on Learning Representations (ICLR)},
  year      = {2026},
  url       = {https://arxiv.org/abs/2510.11752},
}