TOPMed Demo

Trans-Omics for Precision Medicine | TOPMed

Live Presentation of BioNeuralNet.

A step-by-step guide to BioNeuralNet.

• This demonstration was made specifically for the 2025 TOPMed Annual Meeting. Featuring artificial intelligence and machine learning.

• For more information on TOPMed and their mission, please visit TOPMed.

• This presentation uses an outdated version of BioNeuralNet, please refer to TCGA-BRCA_Dataset.ipynb for an up to date end-to-end example.

# In terminal use:
pip install bioneuralnet

# In Jupiter notebook use
!{sys.executable} -m pip install bioneuralnet

BioNeuralNet Components

from bioneuralnet.network_embedding import GNNEmbedding
from bioneuralnet.downstream_task import SubjectRepresentation
from bioneuralnet.downstream_task import DPMON
from bioneuralnet.clustering import CorrelatedPageRank
from bioneuralnet.clustering import CorrelatedLouvain
from bioneuralnet.clustering import HybridLouvain
from bioneuralnet.datasets import DatasetLoader

from bioneuralnet.metrics import omics_correlation
from bioneuralnet.metrics import cluster_correlation
from bioneuralnet.metrics import louvain_to_adjacency
from bioneuralnet.metrics import evaluate_rf
from bioneuralnet.metrics import plot_variance_distribution
from bioneuralnet.metrics import plot_variance_by_feature
from bioneuralnet.metrics import plot_performance
from bioneuralnet.metrics import plot_embeddings
from bioneuralnet.metrics import plot_network
from bioneuralnet.metrics import compare_clusters

from bioneuralnet.utils import rdata_to_df
from bioneuralnet.utils import get_logger

from bioneuralnet.external_tools import SmCCNet

Loading your data:

• If you data is stored in a csv file, it can be loaded by following the example below.

• After loading your data, the remaining steps will be the same.

import pandas as pd

genomic = pd.read_csv( "genes.csv")
proteomics = pd.read_csv("proteins.csv")
metabolites = pd.read_csv("metabolites.csv")
phenotype = pd.read_csv("phenotype.csv")
clinical_data = pd.read_csv("clinical.csv")

Ease of component exploration via DatasetLoader

• This component allows users to explore BioNeuralNet capabilities.

• DatasetLoader example1 is synthetic and purely for example purposes.

from bioneuralnet.datasets import DatasetLoader
loader = DatasetLoader("example1")
omics1, omics2, phenotype, clinical = loader.load_data()

display(omics1)
display(omics2)
display(phenotype)
display(clinical)

	Gene_1	Gene_2	Gene_3	Gene_4	Gene_5	Gene_6	Gene_7	Gene_8	Gene_9	Gene_10	...	Gene_491	Gene_492	Gene_493	Gene_494	Gene_495	Gene_496	Gene_497	Gene_498	Gene_499	Gene_500
Samp_1	22.485701	40.353720	31.025745	20.847206	26.697293	30.205449	23.512005	33.677622	19.430333	30.260153	...	12.967210	13.334132	13.262070	13.608147	13.338606	15.247613	13.603157	14.727795	13.400950	12.769172
Samp_2	37.058850	34.052233	33.487020	23.531461	26.754628	31.735945	22.795952	29.301536	14.936397	30.823015	...	11.886247	13.241386	12.445773	13.400288	12.587993	15.048676	12.018609	14.513958	12.066379	12.583460
Samp_3	20.530767	31.669623	35.189567	20.952544	25.018826	32.157235	25.069464	22.853719	18.220225	23.092805	...	12.198357	12.869640	12.841290	13.127437	12.607377	14.398177	12.554904	14.339382	12.891962	12.760553
Samp_4	33.186888	38.480880	18.897097	31.823300	34.049383	38.799887	24.106468	12.397175	13.724255	27.703085	...	12.197502	13.312536	11.645099	13.933684	12.103087	15.133432	12.436027	15.116888	12.810732	12.972879
Samp_5	28.961981	41.060494	28.494956	18.374495	30.815238	24.004535	29.616296	24.364045	11.409338	33.599828	...	14.157141	13.386978	13.007109	13.826532	12.343930	15.280175	13.363962	14.008675	12.479124	12.156407
...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...
Samp_354	24.520652	28.595409	31.299666	32.095379	33.659730	29.353948	29.871659	25.089732	14.437588	27.261916	...	11.748228	13.096453	12.768936	13.228560	11.595884	15.251601	12.055375	15.530606	13.644383	13.018032
Samp_355	31.252789	28.988087	29.574195	31.189288	32.098841	26.849138	22.619669	28.006670	8.615549	30.838843	...	11.460876	12.994504	13.325608	13.386970	12.335143	14.966097	13.998975	13.250821	12.947672	13.161434
Samp_356	24.894826	25.944887	30.852641	26.705158	30.102546	37.197952	35.032309	34.775576	13.354611	29.697348	...	12.884821	12.785372	12.954534	13.906209	12.074805	15.519179	12.742078	14.445011	12.129990	13.844271
Samp_357	17.034337	38.574705	25.095201	37.062442	35.417758	29.753145	25.855997	25.603043	13.180794	31.188598	...	13.278454	13.445404	12.457436	14.036252	13.412114	14.158416	13.284701	13.662274	12.943670	13.996352
Samp_358	20.839167	27.099788	31.038453	19.410859	31.818995	33.493625	22.769369	22.869481	16.839248	24.644476	...	12.519057	13.308512	12.261076	11.211687	12.058898	15.349873	13.166209	15.295902	13.257230	13.178058

358 rows × 500 columns

	Mir_1	Mir_2	Mir_3	Mir_4	Mir_5	Mir_6	Mir_7	Mir_8	Mir_9	Mir_10	...	Mir_91	Mir_92	Mir_93	Mir_94	Mir_95	Mir_96	Mir_97	Mir_98	Mir_99	Mir_100
Samp_1	15.223913	17.545826	15.784719	14.891983	10.348205	9.689755	11.093630	13.189246	12.526561	10.311609	...	7.551462	5.407878	13.933045	10.287521	10.213316	10.609973	14.554996	10.955650	11.422531	10.862970
Samp_2	16.306965	16.672830	13.361529	14.488549	12.660905	11.333613	11.254930	13.457435	13.279747	10.428848	...	6.862413	7.309226	13.586180	11.975544	11.496937	10.653742	14.414225	11.811004	12.413667	10.719110
Samp_3	16.545119	16.735005	14.617472	17.845267	13.822790	11.329333	11.171749	12.715107	13.787771	10.364577	...	6.874958	7.754733	13.847029	12.424403	10.930177	10.255484	13.570352	11.311925	11.072915	11.418794
Samp_4	13.986899	16.207432	16.293078	17.725286	12.300565	9.844108	11.586551	11.878702	14.594392	10.936651	...	9.615623	6.693593	13.840728	12.245466	10.836894	11.502232	15.483399	10.812811	10.121957	11.039089
Samp_5	16.338332	17.393869	16.397925	15.853725	13.387675	10.599414	12.355466	12.450426	13.778779	10.405327	...	9.146865	8.104206	13.903452	11.423350	9.997107	10.744586	14.465583	11.730166	12.206151	10.724849
...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...
Samp_354	15.065065	16.079830	14.635616	17.013845	11.612843	9.850100	11.375670	12.870558	13.873811	11.369083	...	8.110691	8.124900	14.316491	12.599296	10.302683	11.286541	14.478041	11.141163	11.102427	11.993050
Samp_355	15.997576	15.448951	15.355566	16.501752	11.701778	11.323521	11.610163	12.733208	13.650662	10.770903	...	8.711232	6.977555	14.805837	12.283549	10.847157	9.833187	14.037476	11.082880	11.708466	10.654141
Samp_356	15.206862	14.395378	16.218001	16.044955	13.650741	11.057462	13.324846	13.070428	16.109746	11.253359	...	7.497926	6.840175	15.048896	12.615673	10.583973	12.708186	16.325254	10.789635	10.830833	10.983455
Samp_357	14.474129	15.482863	15.512549	15.136613	14.531277	10.713539	11.632242	13.792981	13.486404	10.870028	...	7.902283	7.339226	14.295151	11.853855	10.439473	11.897782	16.684790	10.847185	11.491449	11.684467
Samp_358	15.094188	16.047304	15.298871	17.022665	15.046043	10.931386	12.289466	14.327210	13.533348	9.792684	...	8.017328	7.814296	14.267007	12.117150	11.832061	11.021241	14.533734	12.015799	11.551237	11.221372

358 rows × 100 columns

	phenotype
Samp_1	235.067423
Samp_2	253.544991
Samp_3	234.204994
Samp_4	281.035429
Samp_5	245.447781
...	...
Samp_354	236.120451
Samp_355	222.572359
Samp_356	268.472285
Samp_357	235.808167
Samp_358	213.886123

358 rows × 1 columns

	Age	Gender	BMI	Chronic_Bronchitis	Emphysema	Asthma
PatientID
Samp_1	78	0	31.2	1	1	0
Samp_2	68	1	19.2	1	0	0
Samp_3	54	1	19.3	0	1	1
Samp_4	47	1	36.2	0	0	1
Samp_5	60	1	26.2	0	1	1
...	...	...	...	...	...	...
Samp_354	71	0	23.0	1	0	1
Samp_355	62	1	25.5	0	1	1
Samp_356	61	0	21.1	1	0	0
Samp_357	64	0	37.6	0	0	1
Samp_358	61	1	31.3	1	0	0

358 rows × 6 columns

Generating a Multi-Omics Network using SmCCNet

• SmCCNet stands for: Sparse Multiple Canonical Correlation Network Analysis Tool.

• Is one of our external tools offered through external_toolscomponent.

• For more information on SmCCNet please visit docs SmCCNet

from bioneuralnet.external_tools import SmCCNet
merged_omics = pd.concat([omics1, omics2, clinical, phenotype], axis=1)

smccnet = SmCCNet(
    phenotype_df=phenotype,
    omics_dfs=[omics1, omics2],
    data_types=["genes","proteins"],
)
global_network, smccnet_clusters = smccnet.run()

SmCCNet Output

• SmCCNet returned a 600x600 network and 3 Sub Netwoks or Clusters.

display(global_network.shape)
display(len(smccnet_clusters))

(600, 600)

3

Generating Low-Dimensional Embeddings using Graph Neural Networks to capture meaningful biological interactions.

• This will return node embeddings that reflect both local connectivity and supervised signals

• Each node (omics feature) is associated with a numeric label (e.g., Pearson correlation with phenotype) that guides learning.

import pandas as pd
from bioneuralnet.network_embedding import GNNEmbedding

merged_omics = pd.concat([omics1, omics2], axis=1)

embeddings = GNNEmbedding(
    adjacency_matrix=global_network,
    omics_data=merged_omics,
    phenotype_data=phenotype,
    clinical_data=clinical,
    tune=True,
)
embeddings.fit()
embeddings_output = embeddings.embed(as_df=True)

Output

For each Omic or Node in the network, our GNNEmbedding function generated a 64 dimmensional representation of that Omics.

display(embeddings_output.shape)

(600, 64)

from bioneuralnet.metrics import plot_embeddings

# Using our embeddings instance we get the necessary labels for the graph.
global_node_labels = embeddings._prepare_node_labels()
embeddings_array = embeddings_output.values

Embeddings visualization

By visulizing the Emebedding Space in a 2 dimensional space. We can notice some cluster of Nodes/Omics. Highlighting close relationships between them.

fig1 = plot_embeddings(embeddings_array, global_node_labels.to_frame(name="phenotype"), method="tsne")

Using the Embeddings

• Let use this omics to enrich the representation of the original dataset.

• The SubjectRepresentation function takes our previously generated embeddings and our original Omics Dataset and associated Phenotype

• This function will use the embeddings to enrich the orignal dataset. For more details and how this is performed please view our GNN Embeddings for Multi-Omics tab.

from bioneuralnet.downstream_task  import SubjectRepresentation

enhanced_omics = SubjectRepresentation(
    omics_data=merged_omics,
    embeddings=embeddings_output,
    phenotype_data=phenotype,
    tune=True,
)
enhanced_omics_df = enhanced_omics.run()

display(enhanced_omics_df)

	Gene_45	Gene_197	Gene_493	Gene_432	Gene_357	Gene_382	Gene_152	Gene_427	Gene_41	Gene_130	...	Gene_202	Gene_249	Gene_280	Gene_62	Mir_21	Gene_495	Mir_59	Gene_340	Gene_291	Gene_81
Samp_1	-2.917084	-13.297334	-30.571794	-29.143930	42.021213	-12.279085	-6.925227	-2.520610	10.380707	-6.079539	...	10.465338	1.922512	5.955446	-2.626381	13.947213	-2.806483	-0.553959	1.726373	-2.651651	-2.406555
Samp_2	-3.102466	-12.954265	-28.690063	-27.614380	41.053950	-12.463960	-6.747382	-2.712280	10.841243	-6.340821	...	10.551755	1.800501	5.912429	-2.548239	12.697251	-2.648551	-0.624136	1.708398	-2.596500	-2.342931
Samp_3	-2.837912	-13.429165	-29.601810	-28.448539	39.378358	-12.837721	-6.436741	-2.377240	10.565243	-6.044507	...	10.307515	1.991552	6.038545	-2.479833	13.769784	-2.652630	-0.592947	1.583839	-2.653498	-2.446277
Samp_4	-2.859205	-12.821326	-26.844345	-29.029980	36.148898	-12.449477	-7.229847	-2.609629	11.160280	-6.466545	...	9.915145	2.137159	5.804663	-2.629407	13.458064	-2.546526	-0.499689	1.648553	-2.520416	-2.530145
Samp_5	-2.982405	-12.612553	-29.984057	-29.110915	46.903902	-12.758489	-6.868394	-2.698423	10.211682	-6.265595	...	10.592827	2.231166	5.314841	-2.462155	12.673634	-2.597200	-0.638230	1.709379	-2.481453	-2.397491
...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...
Samp_354	-2.935778	-13.323266	-29.435019	-27.890044	43.240876	-12.561521	-6.723627	-2.290536	10.847353	-6.341079	...	10.186933	2.077744	5.888022	-2.566317	13.658852	-2.439809	-0.526501	1.692860	-2.711183	-2.466377
Samp_355	-3.065084	-13.295135	-30.718263	-29.112126	44.613763	-12.331960	-6.955723	-2.971791	10.294231	-6.077254	...	9.624207	2.077401	5.623264	-2.663072	14.311062	-2.595351	-0.584500	1.673485	-2.578159	-2.418925
Samp_356	-3.069497	-14.104972	-29.862862	-30.838439	39.916304	-12.129115	-6.739620	-2.504380	10.865619	-6.296479	...	10.497804	1.984929	5.467676	-2.442813	14.854524	-2.540575	-0.527353	1.648124	-2.674988	-2.482001
Samp_357	-3.112670	-14.041763	-28.716949	-28.512552	48.270441	-12.605819	-6.424579	-2.768089	11.026826	-6.542514	...	10.566136	2.141326	6.178911	-2.667312	13.193234	-2.821949	-0.585463	1.659515	-2.625582	-2.425626
Samp_358	-2.983813	-13.657581	-28.264297	-28.283759	41.077029	-12.508963	-6.906523	-2.443945	11.551972	-6.444730	...	10.345094	2.132126	5.578996	-2.523854	14.942860	-2.537228	-0.534264	1.689256	-2.751487	-2.255327

358 rows × 600 columns

Comparing results

Lets compare the enriched dataset vs the raw dataset by using it with a popular machine learning technique, Random Forest. In this example we are simply using the high-dimensional omics data to make prediction on the phenotype.

• Our entire codebase is publicly avaible at the Big Data Management and Mining Laboratory github repository.

• For specific details on this code, please visit: BioNeuralNet

from bioneuralnet.metrics import evaluate_rf, plot_performance

y_phenotype = phenotype.values
X_enriched_omics = enhanced_omics_df.values
X_raw_omics = merged_omics.values

accuracy_with_embeddings = evaluate_rf(X_enriched_omics, y_phenotype, mode="regression")
accuracy_alone = evaluate_rf(X_enriched_omics, y_phenotype, mode="regression")

plot_performance(accuracy_with_embeddings, accuracy_alone, "Raw Omics vs Enriched Omics")

Network Visulization

• As part of this demonstration, I have developed a number of graphic and plotting tools.

• Please note that these tools are not part of the BioNeuralNet core, and it is not my intent to present them this way.

• This examples highlight how external libraies and your own code can be easily integrated into our workflow. Allowing users to further explore the omics-data.

• All this code and visulization aid components will be availble after the presentation.

# Lets take a quick look at some of those clusters from SmCCNet
from bioneuralnet.metrics import plot_network

cluster1, cluster2, cluster3 = smccnet_clusters

cluster1_mapping = plot_network(cluster2, weight_threshold=0.1, show_labels=True, show_edge_weights=True)
display(cluster1_mapping.head())

cluster2_mapping = plot_network(cluster1, weight_threshold=0.009, show_labels=True, show_edge_weights=True)
display(cluster2_mapping.head())

	Omic	Degree
Index
15	Mir_2	14
1	Gene_1	4
2	Gene_2	4
3	Gene_6	4
4	Gene_7	4

	Omic	Degree
Index
7	Mir_53	6
1	Gene_5	4
2	Gene_54	4
4	Gene_88	3
5	Gene_174	3

Correlated Clustering

• BioNeuralNet includes internal modules for performing correlated clustering on complex networks.

• These methods modify and extend the traditional community detection by integrating phenotype correlation, allowing users to extract biologically relevant, phenotype-associated modules from any network.

• For more details on how this performed, please visit our Correlated Clustering Methods tab

from bioneuralnet.clustering import CorrelatedLouvain
import networkx as nx

merged_omics = pd.concat([omics1, omics2], axis=1)
G_network = nx.from_pandas_adjacency(global_network)

louvain_instance = CorrelatedLouvain(
    G=G_network,
    B=merged_omics,
    Y=phenotype,
    tune=True
)

louvain_clusters = louvain_instance.run(as_dfs=True)

print(len(louvain_clusters))

3

# Lets comapre these clusters against SmCCNet clusters
from bioneuralnet.metrics import compare_clusters

compare_clusters(louvain_clusters, smccnet_clusters, phenotype, merged_omics)

	Cluster	Louvain Size	Louvain Correlation	SMCCNET Size	SMCCNET Correlation
0	Cluster_1	77	0.724029	67	0.134853
1	Cluster_2	76	-0.124891	63	0.731769
2	Cluster_3	6	0.069064	15	-0.085057

from bioneuralnet.metrics import plot_network, louvain_to_adjacency

cluster1 = louvain_clusters[0]
cluster2 = louvain_clusters[1]

# Convert Louvain clusters into adjacency matrices
louvain_adj1 = louvain_to_adjacency(cluster1)
louvain_adj2 = louvain_to_adjacency(cluster2)

# Plot using the converted adjacency matrices
cluster1_mapping = plot_network(louvain_adj1, weight_threshold=0.1, show_labels=True, show_edge_weights=True)
display(cluster1_mapping.head())

cluster2_mapping = plot_network(louvain_adj2, weight_threshold=0.1, show_labels=True, show_edge_weights=True)
display(cluster2_mapping.head())

	Omic	Degree
Index
34	Gene_222	7
30	Mir_2	7
66	Gene_1	7
45	Gene_6	6
53	Gene_7	6

	Omic	Degree
Index
30	Gene_187	9
44	Gene_464	6
21	Gene_83	6
16	Gene_354	6
9	Gene_300	5

from bioneuralnet.datasets import DatasetLoader
import numpy as np

loader = DatasetLoader("example1")
omics1, omics2, phenotype, clinical = loader.load_data()
display(phenotype)

min_val = phenotype["phenotype"].min()
max_val = phenotype["phenotype"].max()

# linspace creates an array of evenly spaced values
bins = np.linspace(min_val, max_val, 5)

phenotype["phenotype"] = pd.cut(phenotype["phenotype"], bins=bins, labels=[0, 1, 2, 3], include_lowest=True)
count_values = phenotype["phenotype"].value_counts(sort=False)

display(phenotype)
display(count_values)

	phenotype
Samp_1	235.067423
Samp_2	253.544991
Samp_3	234.204994
Samp_4	281.035429
Samp_5	245.447781
...	...
Samp_354	236.120451
Samp_355	222.572359
Samp_356	268.472285
Samp_357	235.808167
Samp_358	213.886123

358 rows × 1 columns

	phenotype
Samp_1	1
Samp_2	2
Samp_3	1
Samp_4	3
Samp_5	2
...	...
Samp_354	1
Samp_355	1
Samp_356	2
Samp_357	1
Samp_358	1

358 rows × 1 columns

phenotype
0     38
1    158
2    141
3     21
Name: count, dtype: int64

DPMON (Disease Prediction using Multi-Omics Networks) reuses the same GNN architectures but with a different objective:

• DPMON aggregates node embeddings with patient-level omics data.

• A downstream classification head is applied for sample-level disease prediction.

• This end-to-end approach leverages both local (node-level) and global (patient-level) network information.

• This single cell bellow captures the entire workflow demonstrated earlier (Generating GNN Embeddins + Integrating these Embeddings back into Omics Dataset) in an end-to-end iterative pipeline.

from bioneuralnet.downstream_task import DPMON
dpmon = DPMON(
    adjacency_matrix=global_network,
    omics_list=[omics1, omics2],
    phenotype_data=phenotype,
    clinical_data=clinical,
    tune=True,
)
dpmon_predictions = dpmon.run()

display(dpmon_predictions[0])

	Actual	Predicted
0	1	1
1	2	2
2	1	1
3	3	3
4	2	2
...	...	...
353	1	1
354	1	1
355	2	2
356	1	1
357	1	1

358 rows × 2 columns

DPMON allows BioNeuralNet users to significatly improve phenotype predictions with a few lines of code.

from bioneuralnet.metrics import evaluate_rf

X_raw = merged_omics.values
y_global = phenotype.values
raw_rf_acc = evaluate_rf(X_raw, y_global, mode="classification")

from bioneuralnet.utils import plot_performance
plot_performance(dpmon_predictions[1], raw_rf_acc, "Raw Omics, vs DPMON Omics")