GenOM: Robotics, Ontology & Omics

• GenOM is a framework combining robotics, ontology matching, and omics translation, emphasizing modular design, formal verification, and machine learning integration.
• It originated from GenoM for robotic functional layers and evolved to include LLM-driven semantic enrichment and conditional GAN models for omics data.
• GenOM's modular approaches ensure verifiability and adaptability across domains, demonstrating superior real-time performance and robust data translation.

GenOM encompasses a set of influential methodologies and frameworks at the intersection of robotics, ontological alignment, and generative omics, each advancing verifiability, semantic integration, and machine learning-based data translation. Originally stemming from robotic functional architecture (notably GenoM in the LAAS architecture), GenOM now includes advanced frameworks such as GenOM for ontology matching and GenerOM for high-dimensional omics data translation. Across its applications, GenOM stresses correct-by-construction system design, modularity, verifiable semantics, and the integration of formal reasoning or machine learning.

1. Robotic Functional Layer: GenoM and GenoM3

GenoM (Generator of Modules) originated as a methodology for building the functional level of autonomous robots, emphasizing modular decomposition, clear synchronization semantics, and real-time correctness. Each GenoM module encapsulates robotic capabilities—perception, actuation, data fusion—using a formal interface:

• $M = \langle S, T, P, IDS \rangle$ where $S$ are services, $T$ execution tasks, $P$ communication ports, and $IDS$ the internal data structure (Bensalem et al., 2013).
• Services are divided into control (fast, variable manipulation) and execution (potentially long activity, modeled by finite state machines with codels).
• Explicit port definitions (triggers, aborts, status, posters) standardize inter-module and inter-task communications.

GenoM3 extends this by introducing a middleware-independent template layer, allowing targets such as ROS or PocoLibs with zero modification to module logic. Component specifications are written in a mini-language and used to generate code, documentation, and formal models (Foughali et al., 2018).

2. Formal Verification and Correctness by Construction

A critical development in GenOM is the translation of component specifications into formal models for offline and online validation.

• Each control and execution task maps to a timed automaton, with WCET, period, and explicit synchronization and resource acquisition.
• Middleware-aware templates automatically generate code for platforms and for formal methods tools (Fiacre, RT-BIP) (Foughali et al., 2018).
• Safety properties (schedulability, bounded stop delay, mutual exclusion, data freshness) are encoded and verified using model checkers (TINA, D-Finder).

Experimental analysis demonstrates that, even for complex missions (e.g. Mars-rover analogs), deadlock-freedom and real-time deadlines can be certified, and the generated systems remain practical, albeit with an execution time overhead due to the BIP engine's connector evaluation (Bensalem et al., 2013).

3. Ontology Matching with LLMs: GenOM

In ontological alignment, GenOM denotes a LLM-based ontology matching pipeline for semantic interoperability, especially in the biomedical domain (Song et al., 14 Aug 2025). The pipeline addresses major challenges such as sparse definitions and heterogeneity in biomedical ontologies (e.g., SNOMED-CT, NCIT):

• Semantic Enrichment: A lightweight LLM (Qwen2.5-7B-Instruct) is prompted to generate succinct, expert-grade textual definitions for concepts.
• Embedding Retrieval: Augmented definitions, labels, and synonyms are embedded (OpenAI’s text-embedding-3-small) and FAISS+HNSW is used for efficient top-$k$ candidate selection under cosine similarity.
• Equivalence Judgment: Candidate pairs are classified via an LLM prompt, yielding $P(YES)$ scores. Only high-confidence ($P(YES)\geq0.99$) and high-similarity ($\mathrm{sim}\geq0.97$) pairs are retained.
• Exact-Matching Fusion: Integrates string-based matchers (LogMapLt, BERTMapLt) for precision/recall enhancement.

GenOM achieves a mean F1 of 0.769 on OAEI Bio-ML tasks, surpassing leading embedding-based (BERTMap) and LLM-based (LLM4OM) baselines (Song et al., 14 Aug 2025).

Ontology Matching Pipeline in GenOM

Stage	Role	Key Algorithms/Tools
Data Extraction	Collects labels, parents, axioms	DeepOnto, custom scripts
Definition Generation	LLM-based semantic enrichment	Qwen2.5-7B-Instruct
Embedding Retrieval	Candidate generation	OpenAI text-embedding-3-small, FAISS+HNSW
Equivalence Judgment	LLM binary classification	Prompting, softmax scoring
Result Fusion	Boosts precision/recall	LogMapLt, BERTMapLt

Ablation studies demonstrate that enriched definitions and few-shot prompting substantially improve retrieval and ranking scores. The modular design supports efficient adaptation across domains and ontology pairs.

4. Generative Omics Modeling: GenerOM (OmiTrans)

The GenerOM paradigm is exemplified by OmiTrans, a conditional GAN-based omics-to-omics translation framework that learns mappings between high-dimensional omics profiles (e.g., DNA methylation to gene expression) (Zhang et al., 2021). Core methodological features include:

• Conditional GAN objective combines classic adversarial loss with profile-level $L_1$ reconstruction for sharper, biologically faithful outputs.
• Architectures: Both fully connected encoder-decoder networks and 1D-convolutional U-Nets are supported.
• Training and Evaluation: Uses complex datasets (e.g., TCGA pan-cancer; >39k CpG sites, >60k transcripts). OmiTrans achieves MSE=0.1097, sample-wise mean $S$0, and robust generalization to new cohorts.

This approach demonstrates that deep generative models can capture regulatory relationships between omics layers which are inaccessible to traditional regression models, and can be extended to arbitrarily chosen omics modalities or translation targets (Zhang et al., 2021).

5. Integration, Modularity, and Comparative Analysis

GenOM methodologies consistently apply modularity and compositionality. In robotic systems, this enables independent verification, middleware targeting, and late-stage integration. For ontology matching, GenOM’s pipeline structure allows independent improvement of semantic enrichment, retrieval, and judgment submodules. In omics translation, the generator and discriminator architectures are interchangeable, supporting exploration and optimization for different translation pairs.

Performance comparisons within each subfield indicate:

• In robotics, GenoM3-generated components combined with formal-model synthesis provide uniquely strong correctness guarantees, verified on real hardware with practical overhead (Foughali et al., 2018, Bensalem et al., 2013).
• In ontology matching, GenOM outperforms both traditional and recent neural or LLM-based baselines, with definition enrichment and candidate fusion as major contributors to gains (Song et al., 14 Aug 2025).
• In omics translation, OmiTrans-trained models generalize to new biological datasets without retraining, and downstream analysis (e.g., classifiers) perform nearly as well on synthetic as real profiles (Zhang et al., 2021).

6. Limitations and Future Research Directions

While GenOM frameworks set benchmarks in their respective domains, several frontiers remain:

• In robotics, the runtime overhead imposed by runtime verification (e.g., BIP engine) may be ameliorated by adaptive polling, incremental guards, or real-time-aware engine improvements (Bensalem et al., 2013).
• In ontology matching, definition and calibration of equivalence is inherently task-dependent, and future work targets relation expansion (subsumption, relatedness) and prompt-generalization by adapter tuning (Song et al., 14 Aug 2025).
• For generative omics, enhancing biological interpretability (e.g., graph convolutions, domain adaptation, attention mechanisms) and extension to multi-modal or semi-supervised modes are key directions (Zhang et al., 2021).

A plausible implication is that the cross-domain emphasis on compositional templates, formal methods, and semantic or generative enrichment in GenOM anticipates further convergence of declarative system design and dynamic, learning-based integration mechanisms.