We adapt a graph grammar to incorporate long range dependencies. This is acchieved by coarsening a graph and combining the interface on the base and coarsened level.
We test out grammar on RNA molecules. Generated graphs are assigned a bit-score by a state of the art oracle. Human expers have determines that an instance with a bitscore of higher than 39 should be considered to belong to the RNA family in question.
This should take care of all dependencies.
pip3 install git+https://github.com/fabriziocosta/EDeN.git --user
pip3 install graphlearn structout
Lets first load some data, input musst be networkx graphs with 'label' annotation on every node/edge. The cycler will generate a coarsened graph that contracts cycles and store the original graph in g.graphp['original'].
from graphlearn.util import util as util_top
from graphlearn.test import cycler
g = util_top.test_get_circular_graph()
gplus=g.copy()
gplus.node[0]['label']='weird'
c=cycler.Cycler()
g=c.encode_single(g)
gplus = c.encode_single(gplus)Fitting a grammar and conducting one substitution
from graphlearn.lsgg_layered import lsgg_layered
decomposition_args={ "base_thickness_list":[2],
"radius_list": [0],
"thickness_list": [1]}
lsggg = lsgg_layered(decomposition_args=decomposition_args)
lsggg.fit([g, gplus, g,gplus])
neigh = lsggg.neighbors(gplus).__next__()Print original and generated graph to terminal
import structout as so
so.gprint([gplus, gplus.graph['original']], size =15)
neigh=c.encode_single(neigh)
so.gprint([neigh, neigh.graph['original']], size =15)The code below demonstrates how to use this technique on RNAs, if you are interested in the application on general graphs please see the modern implementation showcased above.
import graphlearn.abstract_graphs.RNA as rna
from graphlearn.estimator import Wrapper as estimatorwrapper
sampler=rna.AbstractSampler(radius_list=[0,1],
thickness_list=[2],
min_cip_count=1,
min_interface_count=2,
preprocessor=rna.PreProcessor(base_thickness_list=[1],ignore_inserts=True),
postprocessor=rna.PostProcessor(),
estimator=estimatorwrapper( nu=.5, cv=2, n_jobs=-1)
)
sequences = sampler.sample([sequences from an RNA family])A modern version of the grammar can be found at https://github.com/fabriziocosta/GraphLearn
Experiments are in code/experiments. HPC scripts are for the "sun grid engine".
In the paper we demonstrated GraphRNN applied to RNA grpahs. Our modified GraphRNN:
https://github.com/smautner/GraphRNN
| Meassure | aae | char_rnn | vae | organ | baseline | coarsened |
|---|---|---|---|---|---|---|
| valid | 0.604 | 0.351 | 0.001 | 0.108 | 1.0 | 1.0 |
| unique@1000 | 0.358 | 1.0 | 1.0 | 0.944 | 1.0 | 1.0 |
| unique@10000 | 0.358 | 1.0 | 1.0 | 0.944 | 1.0 | 1.0 |
| FCD/Test | 33.985 | 6.136 | nan | 33.185 | 34.845 | 35.778 |
| SNN/Test | 0.468 | 0.369 | 0.090 | 0.266 | 0.204 | 0.217 |
| Frag/Test | 0.908 | 0.982 | 0.0 | 0.558 | 0.919 | 0.893 |
| Scaf/Test | 0.0480 | 0.265 | nan | 0.0 | 0.0 | 0.0 |
| FCD/TestSF | 35.092 | 6.949 | nan | 33.994 | 35.994 | 36.377 |
| SNN/TestSF | 0.448 | 0.364 | 0.091 | 0.261 | 0.202 | 0.215 |
| Frag/TestSF | 0.905 | 0.975 | 0.0 | 0.558 | 0.910 | 0.891 |
| Scaf/TestSF | 0.0 | 0.022 | nan | 0.0 | 0.0 | 0.0 |
| IntDiv | 0.695 | 0.849 | 0.0 | 0.847 | 0.855 | 0.829 |
| IntDiv2 | 0.659 | 0.836 | 0.0 | 0.808 | 0.850 | 0.824 |
| Filters | 0.957 | 0.934 | 1.0 | 0.583 | 0.079 | 0.872 |
| logP | 0.0146 | 0.0839 | 8.752 | 0.388 | 1.186 | 3.260 |
| SA | 0.311 | 0.081 | 0.266 | 0.438 | 11.714 | 9.176 |
| QED | 0.002 | 0.0002 | 0.185 | 0.05145 | 0.315 | 0.279 |
| NP | 0.306 | 0.099 | 5.090 | 2.971 | 4.025 | 3.325 |
| weight | 323.785 | 52.685 | 76736.756 | 24619.107 | 66388.0129 | 140922.662 |