SGNM

Predict RNA chemical probing reactivity (SHAPE/DMS) from 3D structure.

Overview

sgnm predicts per-residue SHAPE and DMS reactivity profiles from RNA 3D structures. Two model architectures:

• SGNM: Learnable Gaussian Network Model. Computes k independent adjacency matrices from distance and orientation embeddings (via radial basis functions), runs batched GNM variance computation, and projects to per-residue (SHAPE, DMS) predictions.
• EquivariantReactivityModel: SO(3)-equivariant transformer using flash-eq. Operates at the atom level with k-NN graphs and ciffy's PolymerEmbedding.

Both models output (N, 2) for SHAPE and DMS channels, accept ciffy Polymer objects via .ciffy(polymer), and are differentiable.

Installation

pip install -e .

# For the equivariant model (requires CUDA):
pip install -e '.[equivariant]'

Pre-trained Weights

Download from GitHub Releases:

Model	Params	Val Correlation (SHAPE / DMS)	File
GNM	4,626	+0.39 / +0.35	gnm-checkpoint.pth
Equivariant	78,798	+0.63 / +0.75	equivariant-checkpoint.pth

Usage

Prediction

import ciffy
import sgnm

# Load any model from a checkpoint (auto-detects model type)
model = sgnm.load("checkpoint.pth")

# Predict
poly = ciffy.load("structure.cif", backend="torch")
reactivity = model.ciffy(poly)  # (N, 2) for [SHAPE, DMS]

Scoring

Score a structure against an experimental reactivity profile:

import torch
import sgnm
from sgnm.config import ScoringConfig

model = sgnm.load("checkpoint.pth")
scorer = sgnm.StructureScorer(model, ScoringConfig(metric="correlation"))

poly = ciffy.load("structure.cif", backend="torch")
score = scorer.score(experimental_profile, poly)

Available metrics: mae, mse, correlation.

Use channels to score specific output channels (e.g. SHAPE only):

scorer = sgnm.StructureScorer(model, ScoringConfig(metric="correlation", channels=[0]))

Ranking

Rank decoy structures by agreement with experimental reactivity:

import sgnm

model = sgnm.load("checkpoint.pth")
result = sgnm.rank(model, "decoys/", reactivity)

print(f"Best: {result.best.file} (score={result.best.score:.4f})")
result.to_csv("rankings.csv")

Or via the command line:

python scripts/rank.py decoys/ \
    --model gnm --weights checkpoint.pth \
    --profile profiles.h5 --fasta ref.fasta --id 8UYS_A

Training

Training is configured via TOML files. See configs/example.toml for a template.

python scripts/train.py config.toml

Data format

Reactivity data is provided as one HDF5 file per condition, each containing a reactivity dataset of shape (N, L). Paths support an optional :key suffix for datasets within a file (e.g. profiles.h5:PDB130-2A3/reactivity).

[data]
reactivity_paths = ["2a3.h5", "dms.h5"]
fasta_path = "sequences.fasta"
structures_dir = "structures/"

[train]
learning_rate = 1e-3
max_epochs = 100
loss_type = "correlation"

[gnm]
dim = 32
layers = 2

Contact

Email: hmblair@stanford.edu