Multiple myeloma (MM) is a cancer of plasma cells in the bone marrow. It is usually preceded by monoclonal gammopathy of undetermined significance (MGUS), an asymptomatic precursor condition. Most people with MGUS do not rapidly develop MM, yet the two states share many genomic alterations. This makes it difficult to identify which genes and pathways are associated with malignant transformation.

Our study, “BDL-SP: A Bio-inspired DL model for the identification of altered Signaling Pathways in Multiple Myeloma using WES data,” introduces an explainable graph-based deep-learning framework for this problem. The work was published in the American Journal of Cancer Research in 2023 (PubMed; full text in PubMed Central). The associated code and supplementary results are available in the BDL-SP GitHub repository.

BDL-SP stands for Bio-inspired Deep Learning architecture for the identification of altered Signaling Pathways. Its central premise is that genes do not act independently. They participate in protein-interaction networks and signaling pathways. BDL-SP therefore combines patient-specific genomic measurements with a protein–protein interaction (PPI) graph, then uses post-hoc explainability to connect model predictions back to genes, genomic features, and biological pathways.

Figure 1. Overview of the BDL-SP architecture and analysis workflow. Source: BDL-SP repository.

In one sentence: BDL-SP learns from the genomic profiles and known interactions of 824 altered genes, distinguishes MGUS from MM, and uses SHAP and pathway enrichment to explain which molecular signals contributed to the distinction.

Background: Why study the transition from MGUS to MM?

MM develops through the accumulation and selection of genomic abnormalities. Some events may already exist during MGUS, whereas others emerge or become more important during progression. Studying these differences can help answer several biological questions:

1. Which genes are significantly altered in MGUS, MM, or both?
2. Which genomic feature types contribute most strongly to separating the two stages?
3. Do interacting groups of genes provide more information than isolated mutations?
4. Which pathways gain or lose statistical significance as disease advances?
5. Can an AI model be interpreted at both the cohort and individual-sample levels?

These questions are difficult because MGUS and MM are heterogeneous, the available cohorts are highly imbalanced, and most samples are cross-sectional rather than paired longitudinal specimens from the same people.

BDL-SP was designed as a biomarker- and pathway-discovery framework. It is not an approved clinical diagnostic and does not, by itself, predict whether an individual with MGUS will progress to MM.

The study data

The paper analyzed tumor–normal matched whole-exome sequencing (WES) data from three geographic cohorts:

This combined design increased cohort size and population coverage, but the MM and MGUS classes remained strongly imbalanced. It also means that disease stage, study source, sequencing workflow, and population background may be partly entangled and must be considered when interpreting the results.

From WES reads to an explainable graph model

The workflow begins with BAM files and proceeds through variant calling, annotation, gene selection, feature construction, graph learning, and interpretation.

flowchart TD
    A["Tumor-normal WES BAM files"] --> B["Four somatic variant callers"]
    B --> C["ANNOVAR annotation"]
    C --> D["dNdScv gene analysis"]
    D --> E["824-gene union"]
    E --> F["824 × 28 feature matrix"]
    G["STRING PPI network"] --> H["BDL-SP graph model"]
    F --> H
    H --> I["MGUS versus MM classification"]
    H --> J["SHAP interpretation"]
    J --> K["Gene and pathway ranking"]

Step 1: Variant calling with four tools

The study used four somatic-variant callers:

• MuSE,
• Mutect2,
• SomaticSniper, and
• VarScan2.

Using several callers can improve sensitivity to variants detected under different statistical assumptions. However, caller outputs are not interchangeable; reproducible use requires explicit filtering, quality thresholds, genome build, and normalization rules.

Step 2: Functional annotation

Variants were annotated with ANNOVAR to connect sequence-level changes with genes and functional categories.

Step 3: Identification of significantly altered genes

The workflow used dNdScv to identify genes showing evidence of selection. It found:

• 362 significantly altered genes in MGUS,
• 617 significantly altered genes in MM, and
• 155 genes shared by both stages.

Taking the union produced 824 genes for model construction. The union strategy was intended to retain stage-specific signals rather than restricting the model to genes common to both conditions.

Step 4: Construction of 28 genomic features

For each of the 824 genes, the workflow generated 28 genomic features. They summarize three broad SNV groups:

• non-synonymous SNVs,
• synonymous SNVs, and
• other SNVs, including variants outside those two coding categories.

The feature set includes mutation counts and distribution summaries involving measures such as variant allele fraction (VAF) and allele depth (AD). Each patient is therefore represented by a matrix with 824 gene rows and 28 genomic-feature columns.

Step 5: Addition of biological structure

The 824 genes were connected using protein-interaction information from STRING. This produces an adjacency matrix describing which gene products are connected in the prior interaction network.

BDL-SP Model Architecture

The released implementation uses a normalized STRING-derived adjacency matrix with self-loops. For every patient, the graph model processes the 824-by-28 feature matrix through two graph-convolution stages.

The repository code documents the following configuration:

At a high level, each graph-convolution stage performs three operations:

1. normalize the interaction graph and add self-connections;
2. aggregate each gene’s features with information from neighboring genes; and
3. apply a learned transformation and nonlinear activation.

After the second graph layer, each gene has a one-dimensional learned representation. The 824 gene values are flattened and passed to a final layer that produces MM/MGUS probabilities.

Why cost-sensitive learning?

The dataset contains 1,174 MM samples but only 61 MGUS samples. A naïve classifier could obtain high ordinary accuracy by predicting the majority class. BDL-SP therefore assigns greater loss to errors involving the minority class. The repository code uses a 1:20 class-weight ratio, close to the observed imbalance.

This makes balanced accuracy, area under the ROC curve, and area under the precision–recall curve more informative than ordinary accuracy alone.

Why five-fold evaluation?

In stratified five-fold cross-validation, each fold retains approximately the same class proportions. Every sample is used for testing once and training in the other folds. This provides a more stable estimate than a single split, although external validation on a separately collected cohort remains stronger evidence of generalization.

Model performance

The paper compared BDL-SP with six cost-sensitive machine-learning models, including Random Forest, CatBoost, XGBoost, logistic regression, support vector classification, and decision trees.

The three strongest reported models were:

The important lesson is not that a deep model automatically dominates conventional ML. Random Forest achieved similar predictive performance. The paper argues that BDL-SP’s additional value emerged in its application-specific interpretation: the graph architecture and SHAP analysis recovered more previously reported MM-relevant genes among the leading ranks.

Performance estimates should still be read in the context of the small MGUS group, five-fold reuse of one combined dataset, and potential cohort effects.

From classification to explanation with SHAP

BDL-SP used SHAP to rank both genes and genomic feature types according to their contribution to model predictions.

Group-level interpretation

At the group level, the analysis ranked all 824 genes across MM and MGUS. The paper reports KIR3DL2, IGLL5, and FCGR2A as the three highest-ranked genes by the study’s best-SHAP-score procedure.

Other leading genes included previously reported MM-associated drivers and biomarkers such as:

• HLA-A,
• KRAS,
• LTB,
• TP53,
• EGR1,
• FGFR3,
• NFKBIA,
• IRF1, and
• NRAS.

The ranked lists also contained oncogenes such as CARD11, NOTCH1, VAV1, IRS1, MGAM, and ABL2; tumor-suppressor genes such as HLA-B, HLA-C, and SDHA; and actionable-gene candidates including KRAS, NOTCH1, TP53, FGFR3, and ARID1B.

Among 31 previously reported oncogenes present in the 824-gene space, BDL-SP placed:

• 20 in its top 250 genes, and
• 27 in its top 500 genes.

These recovery counts support biological relevance, but they are not independent validation when prior databases are also used to assess the resulting ranking.

Sample-level interpretation

The analysis also generated gene rankings for individual samples. For each patient, SHAP was applied to fold models that correctly predicted the sample, and the contribution with the largest absolute value was used as the “best SHAP score.”

This helps answer a different question from group-level analysis: not “Which genes matter on average?” but “Which genomic attributes most influenced this particular prediction?”

Sample-level explanations are useful for studying heterogeneity, but they can be unstable across model seeds, correlated inputs, and background datasets. Stability analysis is therefore essential before interpreting a rank as patient-specific biology.

Ranking the 28 genomic features

The top three genomic features reported by the paper were:

1. total number of SNVs,
2. total number of SNVs in the “Other SNV” group, and
3. standard deviation of VAF among variants in the “Other SNV” group.

The finding that synonymous and non-canonical SNV summaries contributed to classification is biologically interesting because it moves attention beyond protein-changing mutations alone. It does not establish that synonymous variants are causal; these signals may reflect regulatory effects, mutation processes, linkage, sequencing properties, or broader genomic instability.

Connecting genes to altered signaling pathways

The top 500 BDL-SP-ranked genes were subjected to enrichment analysis using KEGG and Reactome. The study organized pathways into four disease-stage patterns:

This is more informative than a single list of enriched pathways because it asks how pathway-level evidence changes between precursor and malignant states.

The paper reported:

• 5 KEGG and 9 Reactome pathways that gained significance from MGUS to MM;
• 108 KEGG pathways observed as significantly altered only in MM; and
• 134 Reactome pathways observed as significantly altered only in MM.

Some pathways that lost significance during progression were associated with other disease types, whereas many pathways unique to or strengthened in MM reflected cancer-related processes.

Pathway enrichment remains a statistical summary. Results depend on the selected gene set, gene universe, database release, overlapping pathway definitions, and multiple-testing correction. “Significant pathway” should not be interpreted as proof that every pathway component is functionally active in every patient.

Population-related observations

The study compared the number of known oncogenes, tumor-suppressor genes, oncogene/driver genes, and actionable genes across disease stages and geographic cohorts.

It reported significant differences for several gene groups between MM and MGUS in the combined analysis. Differences were also observed between the Indian and European MGUS cohorts, while the compared Indian and American MM cohorts did not show significant differences across all four groups.

These results motivate ancestry-aware follow-up, but “population effect” cannot be separated cleanly from cohort source, sample handling, sequencing protocol, and disease-stage composition in this study design. Future analyses should include genetic ancestry estimates and harmonized processing rather than using study geography as a proxy for ancestry.

Why BDL-SP is significant

1. It combines genomic features with biological networks

The model does not treat genes as unrelated columns. It uses a PPI graph to structure information flow and represent pathway-level relationships.

2. It addresses severe class imbalance explicitly

Cost-sensitive learning and balanced metrics reduce the risk of mistaking majority-class prediction for useful MGUS/MM discrimination.

3. It connects prediction with biological interpretation

SHAP is applied to genes and genomic feature types at group and sample levels, then gene rankings are linked to known cancer-gene categories and pathway enrichment.

4. It studies pathways as changing disease-stage signals

The four-category pathway framework distinguishes pathways that gain significance, lose significance, or appear stage-specific.

5. It provides a bridge to later network models

BDL-SP established a static STRING-based GCN and explainability workflow. The later BIO-DGI study extended the broader research direction using attention-based edge weighting and multiple PPI resources to derive a focused 295-gene panel.

Potential research applications

BDL-SP can support several research directions:

1. Candidate-biomarker discovery
   Prioritize genes for functional experiments or independent cohort validation.

2. Pathway-level comparison of MGUS and MM
   Investigate biological processes that become stronger, weaker, or stage specific.

3. Patient-heterogeneity analysis
   Compare sample-level SHAP profiles and assess whether recurrent explanatory patterns define molecular subgroups.

4. Longitudinal progression studies
   Apply a locked model to serial MGUS, smoldering-MM, and MM specimens.

5. Population-aware genomics
   Test whether gene and pathway rankings replicate across ancestries under harmonized sequencing and preprocessing.

6. Functional prioritization
   Select influential genes and network neighborhoods for CRISPR, cell-line, organoid, or animal-model experiments.

7. Alternative network experiments
   Compare STRING with tissue-specific, pathway-specific, or experimentally measured interaction graphs.

8. Method transfer
   Adapt the gene-feature-plus-network design to other precursor-to-malignancy transitions.

BDL-SP compared with related approaches

The paper’s benchmarking shows why both quantitative and qualitative assessment matter. Similar AUC values do not imply that models learn equally useful biological representations.

Exploring the BDL-SP repository

The public repository contains the model script, SHAP notebooks, supplementary results, figure, dependency list, and an Apache-2.0 license.

Take-home message

BDL-SP demonstrates how prior biological knowledge can be incorporated into deep learning without abandoning interpretability. Starting from WES data, the workflow integrates four variant callers, dNdScv gene selection, 28 genomic features, and a STRING-derived network to model 824 genes across MGUS and MM.

The model achieved a reported balanced accuracy of 96.26% and an AUC of 0.99, while its strongest contribution was the post-hoc biological analysis: recovery of known MM-relevant genes, sample-level explanations, genomic-feature ranking, and classification of pathways according to how their significance changes across disease stages.

Resources and citation

• Ruhela V, Jena L, Kaur G, Gupta R, Gupta A. BDL-SP: A Bio-inspired DL model for the identification of altered Signaling Pathways in Multiple Myeloma using WES data. American Journal of Cancer Research. 2023;13(4):1155–1187. PubMed · PubMed Central
• BDL-SP source code and supplementary results
• BDL-SP architecture figure

Suggested repository citation

Ruhela V, Jena L, Kaur G, Gupta R, Gupta A. BDL-SP: Bio-inspired deep-learning
architecture for identifying altered signaling pathways in multiple myeloma.
https://github.com/vivekruhela/BDL-SP-Bio-inspired-DL-architecture-for-Identification-of-altered-Signaling-Pathways-in-MM

How BIO-DGI connects genomic variation, gene networks, and explainable AI in multiple myeloma

Multiple myeloma (MM) is a cancer of antibody-producing plasma cells. It is usually preceded by a clinically silent precursor condition called monoclonal gammopathy of undetermined significance (MGUS). Although MGUS is common and only a subset of cases progresses to active disease, the molecular events that distinguish stable precursor disease from malignant transformation remain difficult to resolve.

Our study, “A comprehensive targeted panel of 295 genes: Unveiling key disease initiating and transformative biomarkers in multiple myeloma,” addresses this problem with a biologically informed artificial-intelligence framework called BIO-DGI—Bio-Inspired Graph Network Learning-based Gene–Gene Interaction. The work was published in Computers in Biology and Medicine in 2025 (PubMed; DOI). The accompanying research code and analysis notebooks are available in the BIO-DGI GitHub repository.

The central idea is simple: a gene should not be studied only as an isolated column in a data table. Genes operate in pathways and molecular networks. BIO-DGI therefore combines patient-level genomic features with prior knowledge of protein–protein interactions, allowing the model to learn not only which genes matter, but also which biological relationships may help distinguish MGUS from MM.

Figure 1. Overview of the BIO-DGI workflow. Source: BIO-DGI repository.

In one sentence: BIO-DGI treats genes as an interacting biological system, uses explainable graph learning to identify signals separating MGUS from MM, and translates those signals into a focused 295-gene research panel.

Background: Why this problem matters

MGUS and MM share many genomic alterations. This makes the transition between the two states more subtle than a simple “mutation present versus mutation absent” comparison. Several questions arise:

1. Which alterations are already present in the precursor state?
2. Which genes become selectively altered as malignant disease emerges?
3. Do combinations of interacting genes contain more information than single genes considered independently?
4. Can a model remain accurate while also providing biologically interpretable evidence?
5. Can the resulting discoveries be compressed into a panel that is more practical to investigate than whole-exome or whole-genome sequencing?

These questions motivated two useful biological concepts in the paper:

• Disease-initiating genes: genes with alterations detectable in both MGUS and MM, making them candidates for early events in disease development.
• Disease-transformative genes: genes whose alterations are associated more specifically with the malignant MM state and may help characterize the transition from precursor disease.

These labels describe evidence from the analyzed cohorts. They do not, by themselves, prove that an alteration causes disease or that it can predict progression in an individual patient.

The challenge of turning sequencing data into biological insight

Whole-exome and whole-genome studies produce high-dimensional data. A single patient can carry many types of genomic change, including:

• single-nucleotide variants (SNVs),
• small insertions and deletions,
• copy-number variants (CNVs),
• structural variants (SVs), and
• loss-of-function events.

Traditional machine-learning methods often represent every gene as an independent feature. This is convenient, but biologically incomplete. For example, a modest alteration in one gene may become important when several of its interaction partners are also disrupted. Conversely, a frequently altered gene may contribute little to the specific distinction being modeled.

BIO-DGI introduces biological structure before learning begins. Genes are represented as nodes, known or predicted protein interactions as edges, and genomic measurements as node features. The model can then propagate information through the graph and learn which relationships are useful for classifying MM and MGUS profiles.

flowchart TD
    A["Patient genomic profiles"] --> B["Gene-level genomic features"]
    C["Nine PPI resources"] --> D["Integrated gene network"]
    B --> E["BIO-DGI graph model"]
    D --> E
    E --> F["MM versus MGUS classification"]
    E --> G["SHAP and network interpretation"]
    G --> H["295-gene research panel"]

Study data at a glance

The study analyzed genomic profiles from 1,154 MM samples and 61 MGUS samples collected across American, European, and Indian populations. Starting from sequencing-derived alterations, the workflow identified 798 significantly altered genes for graph modeling.

For each sample, the repository’s graphical abstract describes 26 genomic features per gene. These feature vectors summarize multiple aspects of the gene’s alteration profile rather than relying on a single mutation count.

The nine protein-interaction resources integrated in the workflow were BioGRID, BioPlex, FunCoup, HIPPIE, IAS, HumanNet, ProteomeHD, Reactome, and STRING. Combining several resources reduces dependence on a single database, although it can also introduce duplicated, predicted, context-independent, or unevenly supported interactions.

BIO-DGI Model Architecture

For a given patient, the genomic input can be viewed as a feature matrix (X), where rows correspond to genes and columns correspond to genomic measurements. The prior interaction network is represented by an adjacency matrix (A).

BIO-DGI applies a multi-head attention mechanism to learn context-relevant weights for gene–gene relationships. Conceptually, the learned adjacency can be written as:

A graph-convolutional layer then aggregates each gene’s features with information from its neighbors:

where (\hat{A}) includes the learned or updated graph structure, (\hat{D}) is its degree matrix, (H^{(l)}) is the representation at layer (l), (W^{(l)}) contains learned parameters, and (\sigma) is a nonlinear activation.

The graph representation passes through a fully connected layer and softmax output for supervised MM-versus-MGUS classification. The essential design choice is that the model does not use a static PPI graph as mere decoration: attention is used to modify the relative importance of edges before graph aggregation.

Why multi-head attention?

Different attention heads can learn different relationship patterns. One head might prioritize interactions involving mutation burden, while another could emphasize copy-number or loss-of-function context. The heads do not automatically correspond to clean biological mechanisms, but they increase the capacity to represent more than one interaction pattern.

Why graph convolution?

Graph convolution encodes the assumption that a gene’s relevance can depend partly on its neighborhood. This is well suited to signaling and protein-interaction systems, where coordinated perturbation may be more informative than isolated events.

Why a supervised endpoint?

The MM/MGUS label provides a task-specific learning signal. It directs the network toward patterns that discriminate the malignant and precursor cohorts, after which post-hoc interpretation is used to investigate the model’s biological reasoning.

From prediction to explanation

High classification performance alone is not sufficient for biomarker discovery. A model can exploit technical artifacts, cohort imbalance, or correlates with no mechanistic significance. The study therefore used several complementary interpretation steps.

1. SHAP-based feature interpretation

SHAP assigns a contribution score to a feature relative to a model prediction. In this setting, it helps answer questions such as:

• Which genes contribute most strongly to MM-versus-MGUS discrimination?
• Which genomic feature types drive the importance of a gene?
• Are important signals consistent across samples and model folds?

SHAP describes the fitted model; it does not establish causal biology. A high SHAP value means that a feature influenced the model’s prediction under the learned representation.

2. Gene-community analysis

Biological networks contain communities—groups of densely connected genes. Community analysis helps place a highly ranked gene in a functional neighborhood and can expose influential genes that connect multiple modules.

This matters because cancer phenotypes often arise from pathway disruption rather than a single lesion. A gene with moderate individual importance may still be valuable if it links a strongly altered community.

3. Pathway enrichment

Pathway analysis tests whether the prioritized genes are over-represented in known biological processes. The paper reports that the selected genes strongly recover MM-related pathways, providing an orthogonal biological check on the model-derived ranking.

Enrichment should be interpreted carefully: it depends on the gene universe, pathway-database version, overlapping annotations, and multiple-testing procedure.

4. Survival analysis

The clinical relevance of the selected genes was examined using a two-fold univariate survival-analysis strategy. Association with survival strengthens the argument that the panel contains disease-relevant information, but it is not equivalent to prospective clinical validation or proof that the panel improves patient management.

5. Comparison with prior panels

The 295-gene design was evaluated against previously published MM panels for its ability to cover genomic and transformative events. The reported design contains 9,417 coding regions spanning 2.630 Mb.

Comparisons between panels require context. A larger number of covered genes is not automatically better: sequencing depth, bait design, translocation coverage, copy-number performance, sample quality, and intended clinical use all influence utility.

How the 295-gene panel emerges

The panel is best understood as the final product of several evidence filters rather than a list generated by one importance score.

flowchart TD
    A["Sequencing-derived alterations"] --> B["798 altered genes"]
    B --> C["BIO-DGI ranking"]
    C --> D["SHAP evidence"]
    C --> E["Gene communities"]
    C --> F["Pathway evidence"]
    C --> G["Survival associations"]
    D --> H["Evidence integration"]
    E --> H
    F --> H
    G --> H
    H --> I["295-gene panel"]

This layered approach is important. It combines:

• predictive evidence from MM/MGUS discrimination,
• model-explanation evidence,
• prior biological-network context,
• pathway-level coherence,
• disease-stage interpretation, and
• outcome association.

The result is a hypothesis-driven targeted panel intended to capture both early and transformation-associated biology.

What makes this work significant?

It moves from isolated markers to interacting systems

Many biomarker pipelines rank genes independently. BIO-DGI explicitly incorporates relationships among genes and learns which connections matter for the classification task.

It integrates multiple variant classes

The framework summarizes more than one genomic event type. This is important in MM, where SNVs, CNVs, structural changes, and loss-of-function events can jointly shape disease behavior.

It connects model performance with interpretability

The paper reports that BIO-DGI outperformed baseline machine-learning and deep-learning approaches on quantitative metrics and recovered the largest number of MM-relevant genes in its post-hoc comparison. More importantly, the analysis attempts to connect model decisions to genes, genomic features, communities, pathways, and survival.

It produces a focused experimental object

A 2.630 Mb targeted design is substantially more focused than an exome or genome. In principle, a focused panel can support deeper sequencing, lower per-sample sequencing cost, and repeated evaluation of selected regions. Actual performance and cost depend on assay design and laboratory validation.

It introduces a disease-stage perspective

Distinguishing alterations shared by MGUS and MM from those enriched in malignant disease encourages a temporal interpretation of myeloma genomics. This is a useful framework for studying initiation and transformation, even though cross-sectional cohorts cannot fully reconstruct the evolution of an individual patient.

Potential applications

The panel and BIO-DGI framework can support several research applications:

1. MGUS-to-MM progression studies
   Test whether panel alterations are associated with longitudinal progression in independent precursor cohorts.

2. Biomarker prioritization
   Identify genes and network communities for functional follow-up using cell lines, organoids, CRISPR perturbation, or other experimental systems.

3. Targeted cohort profiling
   Generate deeper coverage of selected regions in larger research cohorts when WES or WGS is impractical.

4. Pathway-focused studies
   Investigate whether groups of panel genes converge on recurrent myeloma mechanisms.

5. Longitudinal clonal analysis
   Profile serial samples to study when candidate initiating and transformative events appear.

6. Drug-target hypothesis generation
   Cross-reference prioritized genes and communities with functional or drug–gene resources. Such matches remain hypotheses until experimentally and clinically validated.

7. Transfer to other diseases
   The broader methodology—sample-specific genomic features plus prior interaction networks, attention, graph convolution, and explainability—can be adapted to other precursor-to-malignancy questions.

Targeted panel versus WES and WGS

The approaches are complementary. WES and WGS are valuable for discovery; a targeted panel is useful when a stable set of regions has already been prioritized and deep, scalable interrogation is the objective.

Exploring the BIO-DGI repository

The public repository contains the graphical abstract, source scripts, analysis notebooks, adjacency matrices, fold-level outputs, and GEO-based validation materials.

Recommended next steps

The study creates several opportunities for follow-up:

1. Validate the 295 genes prospectively in independent MGUS and smoldering-MM cohorts.
2. Evaluate time-to-progression using longitudinal samples and multivariable clinical models.
3. Compare BIO-DGI with models that account explicitly for ancestry, cohort, and sequencing-platform effects.
4. Test the stability of gene rankings across alternative PPI resources and network-confidence thresholds.
5. Perform perturbation studies on influential genes and gene communities.
6. Convert the reported coding regions into a locked assay design and establish analytical performance.
7. Assess incremental value over established clinical, cytogenetic, and molecular risk models.
8. Examine whether serial targeted sequencing can track emerging transformative events.

Take-home message

BIO-DGI reframes biomarker discovery as a network problem. Instead of asking only which genes are altered, it asks how genomic alterations interact across a biologically structured system and which parts of that system distinguish MGUS from multiple myeloma.

The resulting 295-gene panel is a focused research resource spanning 9,417 coding regions and 2.630 Mb. Its strongest contribution is not merely the number of genes: it is the integration of multi-variant genomic profiles, nine sources of interaction knowledge, attention-based graph learning, and several layers of post-hoc biological interpretation.

The next challenge is translation—independent replication, longitudinal validation, assay development, and prospective testing. Until then, the panel provides a carefully motivated foundation for investigating the genomic events and gene communities that may shape multiple-myeloma initiation and transformation.

Resources and citation

• Ruhela V, Gupta R, Oberoi R, Gupta A. A comprehensive targeted panel of 295 genes: Unveiling key disease initiating and transformative biomarkers in multiple myeloma. Computers in Biology and Medicine. 2025;196(Pt A):110619. PubMed · DOI
• BIO-DGI source code and analysis notebooks
• Graphical abstract
• Preprint record

Suggested repository citation

Ruhela V, Gupta R, Oberoi R, Gupta A. BIO-DGI: Bio-Inspired Graph Network
Learning-based Gene–Gene Interaction framework for multiple myeloma.
https://github.com/vivekruhela/BIO-DGI

An illustrated guide to the purpose, workflow, use, interpretation, applications, benchmarking, and limitations of miRPipe.

Contents

• Start here: what miRPipe does
• Why small-RNA analysis is difficult
• The complete workflow
• Beginner quick start
• Inputs and outputs
• Advanced-user guide
• Scientific significance
• Applications
• Published benchmark
• Comparison with other pipelines
• Limitations and responsible interpretation
• Troubleshooting
• Reproducibility checklist
• Citation

Start here: what miRPipe does

miRPipe is an end-to-end workflow for analyzing small-RNA sequencing data. It was designed to identify and quantify:

• known microRNAs (miRNAs);
• candidate novel miRNAs;
• miRNA paralogues and other candidates that share functional seed sequences;
• reverse-complement sequences that could otherwise be misclassified; and
• previously annotated piwi-interacting RNAs (piRNAs).

For a cohort containing two biological groups, miRPipe can also build count matrices and run differential-expression analysis with DESeq2. The published implementation is delivered as an interactive Jupyter notebook inside a Docker image.

In plain language, miRPipe turns raw sequencing reads into annotated small-RNA counts and, when the study design permits, lists of small RNAs whose abundance differs between groups.

Figure: miRPipe workflow supplied in the project repository. See the peer-reviewed article for the formal method description.

Who should use this guide?

Why small-RNA analysis is difficult

Small RNAs are short, abundant, and often highly similar. A few nucleotides can determine whether two reads represent the same molecule, an isomiR, a paralogue, a sequencing artifact, or a different RNA class. Reads may also map to more than one genomic locus.

The following concepts are central to miRPipe:

miRPipe addresses these issues by combining alignment, database searches, sequence clustering, genomic-position checks, count aggregation, and statistical analysis.

The complete workflow

flowchart TD
    A["FASTQ files + sample sheet"] --> B["QC and adapter trimming"]
    B --> C{"Read length"}
    C -->|"17–24 nt"| D["miRDeep* miRNA branch"]
    C -->|"25–31 nt intended"| E["Bowtie + piRNAdb branch"]
    D --> F["DASHR re-annotation"]
    F --> G["Seed/xseed clustering"]
    E --> H["piRNA count matrix"]
    G --> I["miRNA count matrix"]
    H --> J["DESeq2"]
    I --> J
    J --> K["Annotated and differential-expression outputs"]

Stage-by-stage explanation

Beginner quick start

1. Confirm that your experiment is appropriate

The published implementation is most directly suited to human, single-end small-RNA FASTQ data. For differential expression, prepare biological replicates in two groups. Do not treat technical replicates as independent biological samples.

Before running the pipeline, record:

• library-preparation kit and adapter sequence;
• sequencing orientation and quality encoding;
• reference genome build (hg19 or hg38);
• miRNA annotation version;
• biological condition for each sample; and
• any known batch, sex, age, or clinical covariates.

The supplied DESeq2 script does not model those additional covariates. If they matter, export the count matrix and run a customized DESeq2 design outside the notebook.

2. Prepare the data directory

Place the FASTQ files and sample_list.csv in one directory. A minimal sample sheet is:

sample,file,condition
Sample_A,Sample_A.fastq.gz,treated
Sample_B,Sample_B.fastq.gz,treated
Sample_C,Sample_C.fastq.gz,untreated
Sample_D,Sample_D.fastq.gz,untreated

For the current notebook and R script:

• use the exact column names shown above;
• keep sample IDs unique;
• use the exact group values treated and untreated;
• verify that every listed file exists; and
• sort sample rows alphanumerically, because the published implementation relies on the sample order matching the final count-matrix columns.

Recommended layout:

mirpipe-run/
└── data/
    ├── sample_list.csv
    ├── Sample_A.fastq.gz
    ├── Sample_B.fastq.gz
    ├── Sample_C.fastq.gz
    └── Sample_D.fastq.gz

3. Pull and run the Docker image

docker pull docker.io/vivekruhela/mirpipe

docker run --rm \
  -p 127.0.0.1:8880:8888 \
  -e PASSWORD='replace-with-a-strong-local-password' \
  -e USE_HTTP=1 \
  -v "$(pwd)/data:/miRPipe/data" \
  docker.io/vivekruhela/mirpipe

Open http://localhost:8880/mirpipe and enter the password chosen in the command. Open the repository’s actual pipeline notebook, mirpipe_pipeline.ipynb.

The 127.0.0.1 binding keeps the unencrypted Jupyter service local to the computer. On a remote Linux server, use an SSH tunnel rather than exposing port 8880 publicly:

ssh -L 8880:127.0.0.1:8880 username@server.example.edu

Then open the same localhost address on your own computer.

Security note: never reuse an institutional, GitHub, Docker Hub, or email password as the notebook password. Do not expose this HTTP notebook directly to the internet.

4. Run the notebook carefully

Run one section at a time and inspect its output before continuing.

1. Confirm the printed home, data, tools, and reference directories.
2. Review FastQC/MultiQC reports.
3. Select the correct adapter option.
4. Select a compatible genome-build and miRBase combination.
5. Run the miRNA and/or piRNA branch.
6. Check the number of known and candidate novel features.
7. Review re-annotation and seed-clustering results.
8. Confirm the count-matrix column order against sample_list.csv.
9. Run DESeq2 only after those checks pass.
10. Copy final results and all run metadata to a permanent project directory.

Inputs and outputs

Required inputs

Five final output files documented in the repository

Also retain the QC reports, complete DESeq2 result tables, normalized counts, software versions, logs, and intermediate annotation tables. A significant-only table is insufficient for reproducible review.

Advanced-user guide

Read-length boundary

The paper describes 17–24 nt reads for miRNA analysis and 25–31 nt reads for piRNA analysis. The current scripts/fastq_split.sh, however, sets minlength=24 and maxlength=31 for the piRNA branch. That creates a 24-nt overlap between the two branches.

Before a new study, decide whether to reproduce the published intention or the repository implementation. Record the decision and benchmark it because a one-nucleotide boundary can change RNA-class assignment.

piRNA mapping behavior

The supplied piRNA script uses Bowtie with exact matching (-v 0), disables reverse-complement alignment (--norc), and quantifies overlap with piRNAdb annotations. These choices are stringent. They can reduce false-positive mappings but can also miss legitimate reads containing sequencing errors, biological variation, or opposite-strand mappings.

For a new production study, document:

• unique versus multi-mapping policy;
• allowed mismatches;
• strand rule;
• annotation database and release;
• genome build; and
• how overlapping annotations are counted.

miRNA alignment and re-annotation

miRDeep* performs the initial known/novel miRNA classification. miRPipe then checks candidate novel sequences against DASHR. If a sequence or its reverse complement matches a known miRNA at a compatible locus, its label and counts can be reassigned.

This is one of the method’s distinctive ideas, but the notebook automates a public website through an older Firefox/Splinter/geckodriver setup. For stable modern operation, replace browser automation with a versioned, locally downloadable annotation resource or a documented API if licensing and availability permit.

Seed/xseed clustering logic

The published method uses the following concepts:

1. Group sequences with identical seed regions.
2. Compare the sequence outside the seed.
3. Treat candidates with identical seed, no more than two xseed changes, and genomic positions within two nucleotides as one novel miRNA; merge their counts.
4. Treat identical-seed candidates at different loci as paralogue candidates.
5. Associate a novel candidate with a known family when it shares the known miRNA seed but maps to a different locus.

flowchart TD
    A["Candidate sequences"] --> B{"Identical seed?"}
    B -->|No| C["Separate candidates"]
    B -->|Yes| D{"xseed difference ≤2 nt?"}
    D -->|No| C
    D -->|Yes| E{"Genomic positions within 2 nt?"}
    E -->|Yes| F["Merge as one candidate"]
    E -->|No| G["Report paralogue relationship"]

These operational rules are useful for reproducibility, but “same seed” does not by itself prove identical biological function. Target context, arm selection, expression, localization, and experimental validation still matter.

Differential-expression model

The supplied scripts/norm_diff_exp.R uses:

design = ~ condition
contrast = c("condition", "treated", "untreated")

It applies the Benjamini–Hochberg procedure and selects features with adjusted p-value ≤ 0.05. The script does not include batch, age, sex, paired design, repeated measures, surrogate variables, or fold-change shrinkage.

For complex studies, use the exported raw-count matrix with an appropriate design, for example:

design = ~ batch + sex + age + condition

The exact model must follow the experimental design; the example is not universally appropriate. Also consider low-count prefiltering, independent filtering, effect-size thresholds, unwanted-variation diagnostics, and shrinkage of log2 fold changes.

Scaling to cohorts and HPC

miRPipe batches samples and runs miRDeep* work in Python threads, while the piRNA branch runs concurrently. This can accelerate multi-sample processing, but resource control is implemented inside a notebook rather than through a scheduler-aware workflow engine.

For an HPC deployment:

• run one container image identified by immutable digest;
• allocate memory per concurrent miRDeep* task;
• avoid oversubscribing CPUs inside a SLURM allocation;
• separate temporary and permanent storage;
• retain per-sample logs and exit codes;
• use a manifest rather than depending on directory glob order; and
• consider refactoring stages into Nextflow, Snakemake, or CWL while preserving miRPipe’s re-annotation and clustering logic.

Minimum QC report for a publication

Report at least:

• samples received, excluded, and analyzed;
• reads before and after adapter/quality filtering;
• read-length distribution;
• fraction assigned to miRNA, piRNA, other small RNA, and unmapped;
• mapping-rate and multi-mapping distribution;
• known versus candidate novel counts;
• number of candidates re-annotated through DASHR;
• number of clusters merged and paralogue groups reported;
• PCA/sample-distance plots from final counts;
• DE model, contrasts, covariates, multiple-testing rule, and effect-size rule; and
• exact software, container, reference, and database versions.

Scientific significance

Methodological significance

miRPipe combines tasks that were often handled separately: known and novel miRNA discovery, piRNA quantification, reverse-complement rescue, seed-based candidate reconciliation, cohort count-matrix construction, and differential expression. Its companion simulator, miRSim, provides ground truth for measuring false positives and false negatives.

Biological significance

Small RNAs regulate gene expression and may reflect disease mechanisms, cell states, or treatment responses. Reliable classification matters because an incorrect label or duplicated count can redirect downstream target prediction, enrichment, biomarker selection, and experimental validation.

Reproducibility significance

Containerization and a notebook lower the installation barrier and make the intended analysis visible. The design was especially valuable for users who needed an integrated workflow in 2022. A modern refresh should retain that accessibility while adding immutable images, command-line configuration, automated tests, and scheduler-aware execution.

Applications

miRPipe can support:

• exploratory small-RNA profiling in cancer and other diseases;
• identification of known and candidate novel miRNAs;
• cohort-level differential-expression analysis;
• piRNA quantification from compatible small-RNA libraries;
• paralogue-aware review of miRNA candidates;
• re-evaluation of candidates affected by reverse-complement labeling;
• benchmarking of small-RNA pipelines using miRSim ground truth;
• method development and sensitivity analysis; and
• non-human analyses after validated replacement of the genome and annotation resources.

It should not be presented as a clinical diagnostic system without independent analytical validation, biological validation, version-controlled references, quality-management procedures, and appropriate regulatory review.

Published benchmark

The 2022 article compared miRPipe with seven pipelines using miRSim synthetic data at 50,000, 100,000, and 1,000,000 reads. The following values reproduce the averages reported in Table 2 of the paper.

The reported overall average was approximately 95.22% accuracy and 94.17% F1 across RNA types and depths. The abstract reports 95.23% for accuracy, a small rounding/reporting difference.

50K reads	0.1M reads	1M reads

Comparison with other pipelines

The table below is a capability-oriented current-use guide, not a new head-to-head performance benchmark.

Practical selection guide

• Choose miRPipe when the scientific question specifically needs the published reverse-complement, seed/xseed, paralogue, and joint miRNA/piRNA logic.
• Choose nf-core/smrnaseq when production-grade portability, automated testing, provenance, HPC execution, and broad QC are the primary needs.
• Choose miRge3.0 when UMI, isomiR, tRF, or A-to-I editing analysis is central.
• Choose sRNAtoolbox/sRNAbench for broad small-RNA profiling, web access, Docker deployment, or non-model organisms.
• Choose miRDeep2 for a focused and widely recognized known/novel miRNA discovery workflow.

For a new methods paper, benchmark multiple tools under the same inputs, reference versions, compute limits, parameter-tuning policy, and independent ground truth.

Troubleshooting

Citation

If miRPipe is used, cite:

Ruhela V, Gupta A, Sriram K, Ahuja G, Kaur G, Gupta R. A Unified Computational Framework for a Robust, Reliable, and Reproducible Identification of Novel miRNAs From the RNA Sequencing Data. Frontiers in Bioinformatics. 2022;2:842051. https://doi.org/10.3389/fbinf.2022.842051

BibTeX:

@article{ruhela2022mirpipe,
  title   = {A Unified Computational Framework for a Robust, Reliable, and Reproducible Identification of Novel miRNAs From the RNA Sequencing Data},
  author  = {Ruhela, Vivek and Gupta, Anubha and Sriram, K. and Ahuja, Gaurav and Kaur, Gurvinder and Gupta, Ritu},
  journal = {Frontiers in Bioinformatics},
  volume  = {2},
  pages   = {842051},
  year    = {2022},
  doi     = {10.3389/fbinf.2022.842051}
}

Project and comparison resources

• miRPipe source repository
• miRPipe article
• miRSim source repository
• miRDeep2 documentation
• miRge3.0 source and documentation
• sRNAtoolbox source and documentation
• nf-core/smrnaseq