Functional lipid analysis via index-based lipidomics profile: a new computational module in LipidOne

TL;DR

This paper introduces a new computational tool in LipidOne that translates lipidomic data into functional insights, helping researchers understand lipid roles in health and disease.

Contribution

The novel Functional Lipid Analysis (FLA) module computes 42 lipid function indices for systems-level interpretation.

Findings

FLA computes 42 indices related to membrane structure, energy storage, and signaling from lipidomic data.

Indices are annotated and linked to predicted protein mediators for pathway-based interpretation.

Application to published datasets confirmed prior findings and revealed new functional insights.

Abstract

Understanding the functional roles of lipids is essential for interpreting metabolic phenotypes in health, disease, and dietary interventions. However, lipidomic analyses typically focus on individual lipid species, making it difficult to extract mechanistic and systems-level insights. We therefore asked how quantitative lipidomic data can be translated into biologically structured and function-oriented interpretations.

Here, we present a major update to LipidOne (lipidone.eu), introducing the novel analytical module: Functional Lipid Analysis (FLA). FLA computes 42 indices describing lipid functions related to membrane structure, energy storage, and signaling. Indices are derived from lipid classes and fatty acyl-, alkyl-, and alkenyl-chain composition, statistically compared across experimental groups, and explored using multivariate and visualization tools. Each index is semantically annotated and linked to predicted protein mediators, enabling pathway- and network-based interpretation. Application to published datasets confirmed previous conclusions while uncovering additional biologically coherent functional insights.

New FLA module is freely available through LipidOne.eu web platform. The LipidOne FLA core R script (v1.0.0) is archived on Zenodo (DOI: 10.5281/zenodo.18468230). The LipidOne web platform is available at https://lipidone.eu.

Figures

Examples of graphical outputs generated by the FLA module in LipidOne. (A) Bar plot comparing all indices; (B) Radar plot summarizing the overall functional landscape; (C) Functional Dominance plot showing a three-bar panel (Structural, Signaling, Energy) to study functional shifts within a study; (D) Functional Volcano plot; (E) Functional Bubble plot displaying a bubble heatmap of top index differences across groups; (F) Functional biomarker discovery table; (G) Functional Correlation Circos plot; (H) PCA score and loading plots; (I) PLS-DA score plot; (J) PLS-DA loading plot; (K) VIP score bar plots; (L) Functional heatmap of indices; (M) Predicted functional protein network.

Tables

Table 1. Functional lipid indices, biochemical classification, and calculation formula.

Functional Category	Index Name	Formula	Proteins	Reference
Structural	Cardiolipin Fraction (Note 1)	(CL/(Total − CL))	TAZ, CRLS1, ACAD9	(Kagan et al. 2016)
	CE/Total Index	(CE/(Total − CE))	ACAT1, SOAT1, CETP	(Barter et al. 2003, Chang et al. 2009, Bernstein et al. 2013)
	Chol/PLmem Index	(Cholesterol/(PC + PE + PS + PI + SM))	ABCA1, HMGCR, LCAT, LDLR, NPC1, NPC2, SGMS1, SGMS2, SMPD1, SOAT1	(Van Meer et al. 2008, Subczynski et al. 2017)
	Chol/SM Index	(Cholesterol/SM)	ABCA1, HMGCR, SGMS1, SGMS2, SMPD1	(Ramstedt and Slotte 1999, Slotte 1999)
	Double Bond Index (Note 2)	(Σ intensity × DB)/(Σ intensity)	FADS1, FADS2, SCD1, ELOVL5	(Stillwell and Wassall 2003, Weiss-Hersh et al. 2020, Baccouch et al. 2023)
	Mono/Poly Ratio	(MUFA/PUFA)	FADS1, FADS2, SCD1, SCD2, ELOVL5	(Ntambi 1999, Baccouch et al. 2023)
	Odd/Even Chain	(Odd‑chain/Even‑chain)	HACL1, ECHDC1, ACAT2, ACAD8, MBOAT1, DHDDS	(Escribá et al. 2015, Benzerouk et al. 2020)
	OxPL/PL Index (Note 3)	(OxPL/(PL − OxPL))	ALOX15, LTA4H, PLA2G4A, GPX4, PTGS1	(Fruhwirth et al. 2007, Itri et al. 2014)
	PC Unsat/Sat Index	(PC unsaturated/PC saturated)	FADS1, FADS2, SCD1, ELOVL5, PPARA	(Stillwell and Wassall 2003, Róg et al. 2004)
	PE/PC Index	(PE/PC)	LPCAT3, GPAM, PLD1, GPCPD1, GPX4	(Li et al. 2006, Gimeno and Cao 2008)
	Membrane Fluidity Index	PC/(PE + SM)	PCYT1A, CEPT1, PCYT2, SGMS1, SMPD1, PEMT	(Fajardo et al. 2011)
	PL/SM Index	(PL/SM)	SMPD1, SMS1, SMS2, SGMS1	(Simons and Ikonen 1997, Lingwood and Simons 2010, Fajardo et al. 2011)
	Saturation Index	(Saturated/Unsaturated)	SCD1, FADS1, FADS2, ELOVL5, SCD2	(Stillwell and Wassall 2003, Weiss-Hersh et al. 2020)
	SM/PC Index	(SM/PC)	SMase, SCD1, FASN, SGMS1, CDIPT	(Simons and Ikonen 1997, Lingwood and Simons 2010, Fajardo et al. 2011)
	Structural/Energetic (Note 4)	((PC + PE + SM)/(TG + DG))	CPT1A, SCD, ACADM, PDK1, FASN	(Fahy et al. 2009)
Signaling	(LPC+LPE)/PL Index	((LPC + LPE)/PL)	PLA2G4A, LPCAT3, PLD1	(Lin and Boyce 2006, Meyer Zu Heringdorf and Jakobs 2007, Shao et al. 2018, Yaginuma et al. 2023, Henze et al. 2025)
	AA/DHA Index	(AA/DHA)	PLA2G4A, ALOX5, PTGS2, PTGDS, CYP1B1	(Serhan et al. 2008, Calder 2013)
	BMP Fraction (Note 5)	(BMP/(Total − BMP))	NPC1, SMPD1, SCARB2	(Hullin-Matsuda et al. 2009)
	Cer/SM Index	(Ceramide/Sphingomyelin)	CerS1, CerS2, SMS1, SMS2, SPTLC1	(Hannun and Obeid 2018, Ogretmen 2018, Stith et al. 2019, Thakkar et al. 2025)
	Ceramide Fraction (Note 6)	(Ceramide/(Total − Ceramide))	CerS1, SMS1, CerS2, SGMS2, LASS1	(Stith et al. 2019, Thakkar et al. 2025)
	DG/PL Index	(DG/PL)	PLCB1, PLCD1, PLCG1, PRKCA, PRKCB	(Berridge 1987, Newton 2010)
	Ether Lipid Fraction (Note 7)	(Ether lipids/(Total − Ether))	ELOVL1, PPARA, PLA2G4A, GPD1L, LPCAT3	(Dean and Lodhi 2018, Fontaine et al. 2020, Rangholia et al. 2021)
	Ferroptosis Susceptibility Index (Note 8)	(PL containing 20:4, 22:4 or 22:6)/(Total − same PL)	GPX4, ACSL4, FTH1, TFRC, SLC7A11, ALOX12, NCOA4, FSP1	(Doll et al. 2017, Lee et al. 2021, Liu et al. 2024)
	GM3/GM2 Ratio (Note 9)	(GM3/GM2)	ST3GAL5, HEXA, HEXB	(Sachinidis et al. 1996, Tajima et al. 2023)
	LPC/PC Index	(LPC/PC)	PLA2, PLD1, LCAT, GPAM, GPCPD1	(Prokazova et al. 1998, Law et al. 2019)
	LPE/PE Index	(LPE/PE)	LPCAT3, GPCPD1, PLD2, GPD1, LPAAT	(Hisano et al. 2021, Yamamoto et al. 2022)
	Lyso-O/PL-O (matched)	((LPC‑O + LPE‑O)/(PC‑O + PE‑O)) with chain‑matching	AGPS, GNPAT, LPCAT2, LPCAT3, PLA2G6, PLA2G7	(Yang et al. 1996, Werner et al. 2020)
	Lyso-P/PL-P (matched)	((LPC‑P + LPE‑P)/(PC‑P + PE‑P)) with chain‑matching	AGPS, FAR1, GNPAT, PLA2G6, TMEM189	(Farooqui 2010, Werner et al. 2020)
	Lyso/PL (matched)	((LPC + LPE)/(PC + PE)) with chain‑matching	PLA2, LPL, ABCA1, LPL, ALOX5	(Yang et al. 1996, Farooqui 2010, Dennis et al. 2011, Murakami et al. 2020, Werner et al. 2020)
	PA/PL Index	(PA/PL)	DGKZ, LPIN1, MTOR, PLD1, PLD2	(Hornberger et al. 2006, Foster et al. 2014)
	PI/PL Index	((PI + PI‑O)/(PC + PE + PG + PS))	PIK3CA, PIK3CB, PLCG1, PLCG2, INPP5D, PTEN	(Di Paolo and De Camilli 2006)
	ω6/ω3 Index	((18:2 + 20:4)/(18:3 + 22:6))	FADS1, FADS2, ELOVL5, ELOVL2, ACAT2, ALOX5	(Simopoulos 2008, Calder 2017)
Energy	Acylcarnitine Fraction (Note 10)	(CAR/(Total − CAR))	CPT1A, CPT2, SLC25A20	(Bartlett and Eaton 2004)
	DG/TG Index	(DG/TG)	ACSL1, DGAT1 DGAT2, LPL, AGPAT2, LCAT	(Reue and Brindley 2008, Lass et al. 2011)
	Energy Load Index (Note 11)	((TG + DG + CE)/(Total − TG − DG − CE))	DGAT1, DGAT2, ACSL4, CPT1A	(Listenberger et al. 2003)
	Long/Medium Chain	(Long chains/Medium chains)	CPT1A, ACADM, ACADL, ACAT2, PDK4	(Papamandjaris et al. 1998, Nakamura et al. 2014, Mett et al. 2021)
	Neutral/Polar (Note 12)	(Neutral lipids/Polar lipids)	ABCA1, NPC1, FABP4, NPC2, AP2M1	(Harayama and Riezman 2018)
	Short/Long Chain	(Short chains/Long chains)	ACSL1, ACADM, ECHDC1, MBOAT1, LPL	(Schönfeld and Wojtczak 2016)
	Storage Index (Note 13)	(TG/(Total − TG))	DGAT1, ACAT1, FASN, ACSL1, LPL	(Samra 2000, Walther and Farese 2012, Czech et al. 2013)
	TG/CE Index	(TG/CE)	ACSL1, ACSL4, DGAT1, DGAT2, LPL	(Maxfield and Tabas 2005, Farese and Walther 2009, Walther and Farese 2012)
	TG/FA Index	(TG/FA)	ACSL1, DGAT1, DGAT2, LIPE, PNPLA2	(Zimmermann et al. 2004, Yen et al. 2008)
	TG/PL Index	(TG/PL)	DGAT1, DGAT2, LPL	(Zimmermann et al. 2004, Moessinger et al. 2014)

Full text

1 Introduction

Lipids constitute a highly diverse group of biomolecules that are traditionally classified into their major categories based on their structure, each containing multiple classes and subclasses of structurally related species. This complexity arises from their modular architecture: lipid molecules are assembled from distinct building blocks—such as glycerol, acyl- alkyl- alkenyl-chains, sphingoid bases, phosphate groups, and saccharides—each with a specific metabolism, enzymatic regulation, and subcellular distribution. Despite their diversity, different biosynthetic routes can converge to produce identical or functionally similar lipid molecules, reflecting a high degree of redundancy and plasticity within lipid metabolic networks (Fahy et al. 2005).

1.1 From lipid complexity to biological insight: a computational framework for functional lipidomics

The biological roles of lipids are equally diverse. They act as essential components of cellular membranes, which makes them essential to cellular structures as well as regulating membrane protein function, as reservoirs of metabolic energy, and as powerful signaling mediators involved in virtually every physiological process. Importantly, lipid functions can be exerted at multiple levels: entire classes (e.g. triglycerides, phospholipids), individual molecular species (e.g. PE 16:0/22:6), or even specific substructures such as fatty acyl chains or headgroups may carry distinct and sometimes independent biochemical implications (Conroy et al. 2024).

Lipids are highly dynamic molecules that undergo continuous remodeling in response to environmental, physiological and pathological stimuli. Compared with aqueous (polar) metabolites, which can fluctuate rapidly and transiently, lipid species generally change on slower timescales when considering membrane composition and neutral-lipid storage/usage. Nevertheless, specific lipid signaling steps can be fast (e.g. COX- and LOX-mediated eicosanoid production) and also the localization of lipid in membranes can be highly dynamic. Because lipids frequently sit at entry and exit nodes of complex metabolic networks, they integrate upstream and downstream biochemical signals; the cumulative effects of these changes are particularly relevant to chronic diseases and long-term adaptation to physiological stress (Cho et al. 2023).

Moreover, lipid biochemistry is tightly interwoven with protein signaling pathways and the metabolism of water-soluble metabolites, forming a complex regulatory network that governs cell fate decisions, membrane organization, and energy homeostasis (Sych et al. 2022).

Over the past decade, lipidomics has emerged as a powerful omics approach to quantitatively map the lipid composition of biological systems. Typically, a lipidomic study focuses on one or more of the following applications:

Biomarker discovery using statistical and bioinformatic analysis, to identify lipid-based markers for early diagnosis, disease progression, or therapeutic response—especially valuable in cardiometabolic and neurodegenerative disorders (Hornemann 2022).Pathway analysis, which seeks to reconstruct lipid transformation routes and infer the involvement of specific genes, enzymes, and regulatory proteins by integrating lipidomic profiles with transcriptomic or proteomic data (Nguyen et al. 2017).Functional analysis, which focuses on interpreting the biological roles of lipids based on their variation across conditions, and on linking those changes to cellular behaviors such as membrane remodeling, energy shifts, or inflammatory signaling.

However, unlike biomarker discovery and pathway reconstruction, functional interpretation remains the main bottleneck in the utilization of lipidomics. The enormous diversity of lipid species and the limited coverage of curated knowledge bases (KEGG, Reactome, WikiPathways, HMDB) leave many measured species unmapped, hampering downstream functional inference. Existing platforms such as LipidOne (Pellegrino et al. 2022, Alabed et al. 2024) and BioPAN (Gaud et al. 2021) mitigate pathway gaps.

When it comes to functional analysis, it is essential to recognize that lipids perform three fundamental biochemical functions in living organisms: structural, energetic, and signaling/regulation (Uzman 2001). This triad is a foundational concept in lipid biochemistry, but it has not yet been fully translated into computational tools capable of resolving which of these functions are altered in a given experimental context (Shevchenko and Simons 2010). A few platforms, such as LION/web (Molenaar et al. 2019), MetaboAnalyst (Lu et al. 2023) and LipidSig (Lin et al. 2021), are excellent resources that offer, among other features, tools described as “functional analysis” for lipidomic datasets. However, in these contexts, “functional analysis” is most often implemented through category/ontology- or property-based annotations and enrichment-style summaries, which are highly useful but typically yield set-level outputs rather than a compact panel of directional, quantitative functional readouts.

This gap is especially relevant because lipid metabolites, due to their chemical heterogeneity, functional redundancy, and compartmental specificity, require dedicated analysis strategies. Interpreting lipidomic changes purely in terms of class enrichment or fold changes is often insufficient and may lead to misleading conclusions. This is particularly true for functional analyses that adopt enrichment-based strategies originally developed for gene expression data. As highlighted by Lee et al. such approaches can misrepresent the biology of small molecules like lipids, which differ substantially from genes in structure, dynamics, and regulation (Lee et al. 2025). In other words, most existing platforms excel at annotating and enriching lipid lists, but they do not provide a compact, quantitative representation of lipid functions that can be directly used as input for downstream statistical modelling.

What is needed is a framework that integrates quantitative lipid features, biochemical functions, and systems-level reasoning to generate interpretable biological outputs in the form of functional readouts, rather than category-level summaries.

To address this unmet need, we developed a dedicated module within the LipidOne platform that systematically translates quantitative lipidomic data into biochemically interpretable functional profiles. The new functional lipid analysis (FLA) module provides a structured and reproducible approach to identify whether specific lipid functions are enhanced or suppressed across experimental conditions by summarising each lipidomic profile into a small set of biochemically interpretable functional variables that can then be analysed with standard statistical tools. These statistical tools (differential analysis, PCA/PLS-DA, clustering) are similar to those implemented in other lipidomics platforms; the distinctive feature of FLA is the functional layer on which they operate.

2 Methods

2.1 Functional lipid analysis (FLA) module

The FLA module within LipidOne is based on a curated set of 42 mathematically defined indices, which are quantified from user-uploaded lipidomics datasets. These indices are calculated as ratios between lipid species, subclasses, or defined building blocks—including acyl, alkyl, or alkenyl chains, and headgroups. Each index is specifically designed to capture distinct biochemical alterations in lipid function, reflecting properties such as the saturated/unsaturated lipid ratio, triglyceride metabolism, or the relative abundance of signaling lipids like ceramides.

2.2 Construction and curation of the functional index library

The FLA module is built on a library of 42 functional lipid indices defined as ratios or weighted sums of lipid classes, molecular species or fatty-acyl building blocks (Table 1). The initial set of candidates was assembled by a targeted survey of the lipidomics and lipid biochemistry literature (PubMed) using combinations of terms such as “lipid index,” “lipid ratio,” “membrane fluidity,” “PUFA/SFA,” “n-6/n-3,” “lysophospholipid signalling” and “ceramide.” Indices were retained when they had a clear biochemical or biophysical rationale (e.g. membrane fluidity, mitochondrial inner-membrane composition, neutral lipid storage versus mobilization), could be unambiguously expressed in terms of standard lipid classes/species or chain descriptors, and were linked by at least one primary study or review to a biological function, phenotype or pathway. In other cases, the literature converged on qualitative contrasts such as “storage versus membrane lipids” or “ω-6 versus ω-3 tone” without specifying a unique formula; here we designed new composite indices that formalise these contrasts into explicit ratios, guided by expert knowledge in lipid biochemistry. All 42 indices were thus defined a priori on biochemical grounds, without data-driven feature selection or optimisation on the case studies analysed in this work.

Table 1: Functional lipid indices, biochemical classification, and calculation formula.

<table><colgroup><col align="left"/><col align="left"/><col align="left"/><col align="left"/><col align="char"/></colgroup><thead><tr><th rowspan="1" colspan="1">Functional Category</th><th rowspan="1" colspan="1">Index Name</th><th rowspan="1" colspan="1">Formula</th><th rowspan="1" colspan="1">Proteins</th><th rowspan="1" colspan="1">Reference</th></tr></thead><tbody><tr><td rowspan="15" colspan="1">Structural</td><td rowspan="1" colspan="1">Cardiolipin Fraction (Note 1)</td><td rowspan="1" colspan="1">(CL/(Total − CL))</td><td rowspan="1" colspan="1">TAZ, CRLS1, ACAD9</td><td rowspan="1" colspan="1">(<xref>Kagan <italic>et al.</italic> 2016</xref>)</td></tr><tr><td rowspan="1" colspan="1">CE/Total Index</td><td rowspan="1" colspan="1">(CE/(Total − CE))</td><td rowspan="1" colspan="1">ACAT1, SOAT1, CETP</td><td rowspan="1" colspan="1">(<xref>Barter <italic>et al.</italic> 2003</xref>, <xref>Chang <italic>et al.</italic> 2009</xref>, <xref>Bernstein <italic>et al.</italic> 2013</xref>)</td></tr><tr><td rowspan="1" colspan="1">Chol/PLmem Index</td><td rowspan="1" colspan="1">(Cholesterol/(PC + PE + PS + PI + SM))</td><td rowspan="1" colspan="1">ABCA1, HMGCR, LCAT, LDLR, NPC1, NPC2, SGMS1, SGMS2, SMPD1, SOAT1</td><td rowspan="1" colspan="1">(<xref>Van Meer <italic>et al.</italic> 2008</xref>, <xref>Subczynski <italic>et al.</italic> 2017</xref>)</td></tr><tr><td rowspan="1" colspan="1">Chol/SM Index</td><td rowspan="1" colspan="1">(Cholesterol/SM)</td><td rowspan="1" colspan="1">ABCA1, HMGCR, SGMS1, SGMS2, SMPD1</td><td rowspan="1" colspan="1">(<xref>Ramstedt and Slotte 1999</xref>, <xref>Slotte 1999</xref>)</td></tr><tr><td rowspan="1" colspan="1">Double Bond Index (Note 2)</td><td rowspan="1" colspan="1">(Σ intensity × DB)/(Σ intensity)</td><td rowspan="1" colspan="1">FADS1, FADS2, SCD1, ELOVL5</td><td rowspan="1" colspan="1">(<xref>Stillwell and Wassall 2003</xref>, <xref>Weiss-Hersh <italic>et al.</italic> 2020</xref>, <xref>Baccouch <italic>et al.</italic> 2023</xref>)</td></tr><tr><td rowspan="1" colspan="1">Mono/Poly Ratio</td><td rowspan="1" colspan="1">(MUFA/PUFA)</td><td rowspan="1" colspan="1">FADS1, FADS2, SCD1, SCD2, ELOVL5</td><td rowspan="1" colspan="1">(<xref>Ntambi 1999</xref>, <xref>Baccouch <italic>et al.</italic> 2023</xref>)</td></tr><tr><td rowspan="1" colspan="1">Odd/Even Chain</td><td rowspan="1" colspan="1">(Odd‑chain/Even‑chain)</td><td rowspan="1" colspan="1">HACL1, ECHDC1, ACAT2, ACAD8, MBOAT1, DHDDS</td><td rowspan="1" colspan="1">(<xref>Escribá <italic>et al.</italic> 2015</xref>, <xref>Benzerouk <italic>et al.</italic> 2020</xref>)</td></tr><tr><td rowspan="1" colspan="1">OxPL/PL Index (Note 3)</td><td rowspan="1" colspan="1">(OxPL/(PL − OxPL))</td><td rowspan="1" colspan="1">ALOX15, LTA4H, PLA2G4A, GPX4, PTGS1</td><td rowspan="1" colspan="1">(<xref>Fruhwirth <italic>et al.</italic> 2007</xref>, <xref>Itri <italic>et al.</italic> 2014</xref>)</td></tr><tr><td rowspan="1" colspan="1">PC Unsat/Sat Index</td><td rowspan="1" colspan="1">(PC unsaturated/PC saturated)</td><td rowspan="1" colspan="1">FADS1, FADS2, SCD1, ELOVL5, PPARA</td><td rowspan="1" colspan="1">(<xref>Stillwell and Wassall 2003</xref>, <xref>Róg <italic>et al.</italic> 2004</xref>)</td></tr><tr><td rowspan="1" colspan="1">PE/PC Index</td><td rowspan="1" colspan="1">(PE/PC)</td><td rowspan="1" colspan="1">LPCAT3, GPAM, PLD1, GPCPD1, GPX4</td><td rowspan="1" colspan="1">(<xref>Li <italic>et al.</italic> 2006</xref>, <xref>Gimeno and Cao 2008</xref>)</td></tr><tr><td rowspan="1" colspan="1">Membrane Fluidity Index</td><td rowspan="1" colspan="1">PC/(PE + SM)</td><td rowspan="1" colspan="1">PCYT1A, CEPT1, PCYT2, SGMS1, SMPD1, PEMT</td><td rowspan="1" colspan="1">(<xref>Fajardo <italic>et al.</italic> 2011</xref>)</td></tr><tr><td rowspan="1" colspan="1">PL/SM Index</td><td rowspan="1" colspan="1">(PL/SM)</td><td rowspan="1" colspan="1">SMPD1, SMS1, SMS2, SGMS1</td><td rowspan="1" colspan="1">(<xref>Simons and Ikonen 1997</xref>, <xref>Lingwood and Simons 2010</xref>, <xref>Fajardo <italic>et al.</italic> 2011</xref>)</td></tr><tr><td rowspan="1" colspan="1">Saturation Index</td><td rowspan="1" colspan="1">(Saturated/Unsaturated)</td><td rowspan="1" colspan="1">SCD1, FADS1, FADS2, ELOVL5, SCD2</td><td rowspan="1" colspan="1">(<xref>Stillwell and Wassall 2003</xref>, <xref>Weiss-Hersh <italic>et al.</italic> 2020</xref>)</td></tr><tr><td rowspan="1" colspan="1">SM/PC Index</td><td rowspan="1" colspan="1">(SM/PC)</td><td rowspan="1" colspan="1">SMase, SCD1, FASN, SGMS1, CDIPT</td><td rowspan="1" colspan="1">(<xref>Simons and Ikonen 1997</xref>, <xref>Lingwood and Simons 2010</xref>, <xref>Fajardo <italic>et al.</italic> 2011</xref>)</td></tr><tr><td rowspan="1" colspan="1">Structural/Energetic (Note 4)</td><td rowspan="1" colspan="1">((PC + PE + SM)/(TG + DG))</td><td rowspan="1" colspan="1">CPT1A, SCD, ACADM, PDK1, FASN</td><td rowspan="1" colspan="1">(<xref>Fahy <italic>et al.</italic> 2009</xref>)</td></tr><tr><td rowspan="17" colspan="1">Signaling</td><td rowspan="1" colspan="1">(LPC+LPE)/PL Index</td><td rowspan="1" colspan="1">((LPC + LPE)/PL)</td><td rowspan="1" colspan="1">PLA2G4A, LPCAT3, PLD1</td><td rowspan="1" colspan="1">(<xref>Lin and Boyce 2006</xref>, <xref>Meyer Zu Heringdorf and Jakobs 2007</xref>, <xref>Shao <italic>et al.</italic> 2018</xref>, <xref>Yaginuma <italic>et al.</italic> 2023</xref>, <xref>Henze <italic>et al.</italic> 2025</xref>)</td></tr><tr><td rowspan="1" colspan="1">AA/DHA Index</td><td rowspan="1" colspan="1">(AA/DHA)</td><td rowspan="1" colspan="1">PLA2G4A, ALOX5, PTGS2, PTGDS, CYP1B1</td><td rowspan="1" colspan="1">(<xref>Serhan <italic>et al.</italic> 2008</xref>, <xref>Calder 2013</xref>)</td></tr><tr><td rowspan="1" colspan="1">BMP Fraction (Note 5)</td><td rowspan="1" colspan="1">(BMP/(Total − BMP))</td><td rowspan="1" colspan="1">NPC1, SMPD1, SCARB2</td><td rowspan="1" colspan="1">(<xref>Hullin-Matsuda <italic>et al.</italic> 2009</xref>)</td></tr><tr><td rowspan="1" colspan="1">Cer/SM Index</td><td rowspan="1" colspan="1">(Ceramide/Sphingomyelin)</td><td rowspan="1" colspan="1">CerS1, CerS2, SMS1, SMS2, SPTLC1</td><td rowspan="1" colspan="1">(<xref>Hannun and Obeid 2018</xref>, <xref>Ogretmen 2018</xref>, <xref>Stith <italic>et al.</italic> 2019</xref>, <xref>Thakkar <italic>et al.</italic> 2025</xref>)</td></tr><tr><td rowspan="1" colspan="1">Ceramide Fraction (Note 6)</td><td rowspan="1" colspan="1">(Ceramide/(Total − Ceramide))</td><td rowspan="1" colspan="1">CerS1, SMS1, CerS2, SGMS2, LASS1</td><td rowspan="1" colspan="1">(<xref>Stith <italic>et al.</italic> 2019</xref>, <xref>Thakkar <italic>et al.</italic> 2025</xref>)</td></tr><tr><td rowspan="1" colspan="1">DG/PL Index</td><td rowspan="1" colspan="1">(DG/PL)</td><td rowspan="1" colspan="1">PLCB1, PLCD1, PLCG1, PRKCA, PRKCB</td><td rowspan="1" colspan="1">(<xref>Berridge 1987</xref>, <xref>Newton 2010</xref>)</td></tr><tr><td rowspan="1" colspan="1">Ether Lipid Fraction (Note 7)</td><td rowspan="1" colspan="1">(Ether lipids/(Total − Ether))</td><td rowspan="1" colspan="1">ELOVL1, PPARA, PLA2G4A, GPD1L, LPCAT3</td><td rowspan="1" colspan="1">(<xref>Dean and Lodhi 2018</xref>, <xref>Fontaine <italic>et al.</italic> 2020</xref>, <xref>Rangholia <italic>et al.</italic> 2021</xref>)</td></tr><tr><td rowspan="1" colspan="1">Ferroptosis Susceptibility Index (Note 8)</td><td rowspan="1" colspan="1">(PL containing 20:4, 22:4 or 22:6)/(Total − same PL)</td><td rowspan="1" colspan="1">GPX4, ACSL4, FTH1, TFRC, SLC7A11, ALOX12, NCOA4, FSP1</td><td rowspan="1" colspan="1">(<xref>Doll <italic>et al.</italic> 2017</xref>, <xref>Lee <italic>et al.</italic> 2021</xref>, <xref>Liu <italic>et al.</italic> 2024</xref>)</td></tr><tr><td rowspan="1" colspan="1">GM3/GM2 Ratio (Note 9)</td><td rowspan="1" colspan="1">(GM3/GM2)</td><td rowspan="1" colspan="1">ST3GAL5, HEXA, HEXB</td><td rowspan="1" colspan="1">(<xref>Sachinidis <italic>et al.</italic> 1996</xref>, <xref>Tajima <italic>et al.</italic> 2023</xref>)</td></tr><tr><td rowspan="1" colspan="1">LPC/PC Index</td><td rowspan="1" colspan="1">(LPC/PC)</td><td rowspan="1" colspan="1">PLA2, PLD1, LCAT, GPAM, GPCPD1</td><td rowspan="1" colspan="1">(<xref>Prokazova <italic>et al.</italic> 1998</xref>, <xref>Law <italic>et al.</italic> 2019</xref>)</td></tr><tr><td rowspan="1" colspan="1">LPE/PE Index</td><td rowspan="1" colspan="1">(LPE/PE)</td><td rowspan="1" colspan="1">LPCAT3, GPCPD1, PLD2, GPD1, LPAAT</td><td rowspan="1" colspan="1">(<xref>Hisano <italic>et al.</italic> 2021</xref>, <xref>Yamamoto <italic>et al.</italic> 2022</xref>)</td></tr><tr><td rowspan="1" colspan="1">Lyso-O/PL-O (matched)</td><td rowspan="1" colspan="1">((LPC‑O + LPE‑O)/(PC‑O + PE‑O)) with chain‑matching</td><td rowspan="1" colspan="1">AGPS, GNPAT, LPCAT2, LPCAT3, PLA2G6, PLA2G7</td><td rowspan="1" colspan="1">(<xref>Yang <italic>et al.</italic> 1996</xref>, <xref>Werner <italic>et al.</italic> 2020</xref>)</td></tr><tr><td rowspan="1" colspan="1">Lyso-P/PL-P (matched)</td><td rowspan="1" colspan="1">((LPC‑P + LPE‑P)/(PC‑P + PE‑P)) with chain‑matching</td><td rowspan="1" colspan="1">AGPS, FAR1, GNPAT, PLA2G6, TMEM189</td><td rowspan="1" colspan="1">(<xref>Farooqui 2010</xref>, <xref>Werner <italic>et al.</italic> 2020</xref>)</td></tr><tr><td rowspan="1" colspan="1">Lyso/PL (matched)</td><td rowspan="1" colspan="1">((LPC + LPE)/(PC + PE)) with chain‑matching</td><td rowspan="1" colspan="1">PLA2, LPL, ABCA1, LPL, ALOX5</td><td rowspan="1" colspan="1">(<xref>Yang <italic>et al.</italic> 1996</xref>, <xref>Farooqui 2010</xref>, <xref>Dennis <italic>et al.</italic> 2011</xref>, <xref>Murakami <italic>et al.</italic> 2020</xref>, <xref>Werner <italic>et al.</italic> 2020</xref>)</td></tr><tr><td rowspan="1" colspan="1">PA/PL Index</td><td rowspan="1" colspan="1">(PA/PL)</td><td rowspan="1" colspan="1">DGKZ, LPIN1, MTOR, PLD1, PLD2</td><td rowspan="1" colspan="1">(<xref>Hornberger <italic>et al.</italic> 2006</xref>, <xref>Foster <italic>et al.</italic> 2014</xref>)</td></tr><tr><td rowspan="1" colspan="1">PI/PL Index</td><td rowspan="1" colspan="1">((PI + PI‑O)/(PC + PE + PG + PS))</td><td rowspan="1" colspan="1">PIK3CA, PIK3CB, PLCG1, PLCG2, INPP5D, PTEN</td><td rowspan="1" colspan="1">(<xref>Di Paolo and De Camilli 2006</xref>)</td></tr><tr><td rowspan="1" colspan="1">ω6/ω3 Index</td><td rowspan="1" colspan="1">((18:2 + 20:4)/(18:3 + 22:6))</td><td rowspan="1" colspan="1">FADS1, FADS2, ELOVL5, ELOVL2, ACAT2, ALOX5</td><td rowspan="1" colspan="1">(<xref>Simopoulos 2008</xref>, <xref>Calder 2017</xref>)</td></tr><tr><td rowspan="10" colspan="1">Energy</td><td rowspan="1" colspan="1">Acylcarnitine Fraction (Note 10)</td><td rowspan="1" colspan="1">(CAR/(Total − CAR))</td><td rowspan="1" colspan="1">CPT1A, CPT2, SLC25A20</td><td rowspan="1" colspan="1">(<xref>Bartlett and Eaton 2004</xref>)</td></tr><tr><td rowspan="1" colspan="1">DG/TG Index</td><td rowspan="1" colspan="1">(DG/TG)</td><td rowspan="1" colspan="1">ACSL1, DGAT1 DGAT2, LPL, AGPAT2, LCAT</td><td rowspan="1" colspan="1">(<xref>Reue and Brindley 2008</xref>, <xref>Lass <italic>et al.</italic> 2011</xref>)</td></tr><tr><td rowspan="1" colspan="1">Energy Load Index (Note 11)</td><td rowspan="1" colspan="1">((TG + DG + CE)/(Total − TG − DG − CE))</td><td rowspan="1" colspan="1">DGAT1, DGAT2, ACSL4, CPT1A</td><td rowspan="1" colspan="1">(<xref>Listenberger <italic>et al.</italic> 2003</xref>)</td></tr><tr><td rowspan="1" colspan="1">Long/Medium Chain</td><td rowspan="1" colspan="1">(Long chains/Medium chains)</td><td rowspan="1" colspan="1">CPT1A, ACADM, ACADL, ACAT2, PDK4</td><td rowspan="1" colspan="1">(<xref>Papamandjaris <italic>et al.</italic> 1998</xref>, <xref>Nakamura <italic>et al.</italic> 2014</xref>, <xref>Mett <italic>et al.</italic> 2021</xref>)</td></tr><tr><td rowspan="1" colspan="1">Neutral/Polar (Note 12)</td><td rowspan="1" colspan="1">(Neutral lipids/Polar lipids)</td><td rowspan="1" colspan="1">ABCA1, NPC1, FABP4, NPC2, AP2M1</td><td rowspan="1" colspan="1">(<xref>Harayama and Riezman 2018</xref>)</td></tr><tr><td rowspan="1" colspan="1">Short/Long Chain</td><td rowspan="1" colspan="1">(Short chains/Long chains)</td><td rowspan="1" colspan="1">ACSL1, ACADM, ECHDC1, MBOAT1, LPL</td><td rowspan="1" colspan="1">(<xref>Schönfeld and Wojtczak 2016</xref>)</td></tr><tr><td rowspan="1" colspan="1">Storage Index (Note 13)</td><td rowspan="1" colspan="1">(TG/(Total − TG))</td><td rowspan="1" colspan="1">DGAT1, ACAT1, FASN, ACSL1, LPL</td><td rowspan="1" colspan="1">(<xref>Samra 2000</xref>, <xref>Walther and Farese 2012</xref>, <xref>Czech <italic>et al.</italic> 2013</xref>)</td></tr><tr><td rowspan="1" colspan="1">TG/CE Index</td><td rowspan="1" colspan="1">(TG/CE)</td><td rowspan="1" colspan="1">ACSL1, ACSL4, DGAT1, DGAT2, LPL</td><td rowspan="1" colspan="1">(<xref>Maxfield and Tabas 2005</xref>, <xref>Farese and Walther 2009</xref>, <xref>Walther and Farese 2012</xref>)</td></tr><tr><td rowspan="1" colspan="1">TG/FA Index</td><td rowspan="1" colspan="1">(TG/FA)</td><td rowspan="1" colspan="1">ACSL1, DGAT1, DGAT2, LIPE, PNPLA2</td><td rowspan="1" colspan="1">(<xref>Zimmermann <italic>et al.</italic> 2004</xref>, <xref>Yen <italic>et al.</italic> 2008</xref>)</td></tr><tr><td rowspan="1" colspan="1">TG/PL Index</td><td rowspan="1" colspan="1">(TG/PL)</td><td rowspan="1" colspan="1">DGAT1, DGAT2, LPL</td><td rowspan="1" colspan="1">(<xref>Zimmermann <italic>et al.</italic> 2004</xref>, <xref>Moessinger <italic>et al.</italic> 2014</xref>)</td></tr></tbody></table>

Each index is specified at a defined structural level, and lipid membership in its numerator and denominator is determined by rule-based mapping of user-provided identifiers to an internal LIPID MAPS-compatible ontology. Class labels (e.g. PC versus LPC, TG versus DG), acyl/alkyl/alkenyl chains for chain-level indices, and the annotation level chosen by the user (molecular species versus sum composition) jointly determine whether a lipid contributes to a given index. When only sum compositions are available, indices requiring full chain resolution are automatically disabled, while class-level indices remain computable. These rules are encoded in the LipidOne back-end and applied systematically to all datasets.

For each index, we also compiled a set of proteins that directly catalyse, regulate or consistently respond to the lipid species involved, including acyltransferases, lipases, desaturases, elongases, transporters and key regulatory enzymes. Protein lists were curated by combining pathway resources (e.g. KEGG, Reactome, LIPID MAPS pathway annotations) with the primary literature cited in Table 1, and proteins were included only when a direct mechanistic link to the index lipids could be established.

Index–protein associations were curated in human and then extended to the ten model organisms supported by LipidOne using STRING orthology mapping, propagating links only when at least one clear ortholog was available (summarised in Table S2). Because lipid metabolism diverges across kingdoms, especially between animals, plants and fungi, we recommend interpreting protein-level suggestions more cautiously in non-mammalian systems and primarily as hypotheses for downstream validation.

LipidOne input conventions, including supported shorthand nomenclature, lipid class coverage, and annotation levels, have been described previously (Alabed et al. 2024) and are referenced from the LipidOne homepage. All datasets must be successfully uploaded to LipidOne and validated using these standard conventions prior to FLA. For datasets generated using vendor-specific nomenclature that differs from LIPID MAPS shorthand, LipidOne provides a Shorthand Lipid Translator utility to convert common identifiers into the required format. For FLA, only lipid classes and acyl chain compositions listed in Table 1 contribute to each functional index; indices are not computed when the required lipid species are absent.

To facilitate interpretation, the indices are grouped into three major functional categories, corresponding to the principal roles of lipids in biological systems (Gurr et al. 2016):

Structural indices, which describe both quantitative and qualitative features of membrane-associated lipids. These indices capture variations in the abundance of entire lipid classes (e.g. phosphatidylcholines, cardiolipins), but also reflect molecular-level properties such as the degree of unsaturation (linked to membrane fluidity), average chain length (affecting membrane curvature), and the presence of ether bonds. While certain structural lipids like cardiolipins are localized in mitochondria and have multifunctional roles, the structural indices in this group aim to represent the overarching biophysical properties of lipid assemblies across different cellular compartments.

Signaling indices, which track lipid species and subclasses involved in cell signaling, inflammation, and intercellular communication. These include not only individual bioactive molecules—such as omega-3 and omega-6 polyunsaturated fatty acids and eicosanoid precursors—but also broader lipid classes or sub-classes known for their signaling functions, including ceramides, lysolipids, and ether lipids. These molecules act as metabolic messengers, modulating immune responses, stress signaling, and pathways related to cell growth, differentiation, or apoptosis.

Energy-related indices, which capture lipid species involved in metabolic energy storage, mobilization, and mitochondrial function. These include not only neutral lipids such as triglycerides and diglycerides, but also lipid classes and acyl chain features linked to β-oxidation. For instance, changes in acyl chain length may reflect enhanced or impaired fatty acid catabolism, while specific mitochondrial lipids—such as cardiolipins and acyl-carnitines—are functionally associated with energy production and transport (Grevengoed et al. 2014, Paradies et al. 2019). Although some of these, like cardiolipins, may also contribute to structural integrity, their metabolic implications justify their partial inclusion in this group.

Table 1 lists all indices along with their associated functional category and the formula used to calculate their numerical values.

Table S1 provides the full specification of the FLA index library, including each index formula, the associated enzyme list, the biochemical description, and the standardized interpretative phrases used by LipidOne for EXP ≪ CTRL and EXP ≫ CTRL.

Each index in Table 1 has been systematically curated based on evidence from the scientific literature and reflects established relationships between lipid composition and cellular functions. Indices are also linked to specific enzymes or regulatory proteins that determine, modulate or respond to lipid variation. For instance, the Ceramide/Sphingomyelin Index is associated with enzymes involved in apoptotic signaling and stress responses such CerS1, CerS2, SMS1, SMS2, SPTLC1, while the DG/TG Index reflects lipid mobilization and energy storage dynamics by activation of ACSL1, DGAT2, LPL, AGPAT2, LCAT proteins. By linking these indices with the associated proteins, LipidOne the opportunity to make inferences about the mechanisms that drive the differences between the analyzed samples.

As the indices are directly related to specific proteins the FLA module facilitates the construction of protein interaction networks, based on predicted activations or deactivations of these proteins, allowing researchers to visualize how specific lipid alterations might influence broader biological pathways. The network-based interpretation extends to protein clusters, where specific protein families—such as kinases or phosphatases—may be identified as central mediators of lipid-driven pathways. By visualizing these interactions, users can generate hypotheses regarding the functional consequences of lipidomic changes. These insights can then inform experimental designs or suggest potential therapeutic interventions.

Finally, since LipidOne has been designed to build protein interaction networks for 10 model organisms, we provide homologous proteins from these species based on their short names in the STRING database. Table S2 reports the curated index–protein mapping in human and its orthology-based propagation across the supported model organisms (STRING ortholog mapping), indicating which indices retain protein support in each species.

The full set of 42 curated functional indices is automatically computed for each sample in the lipidomics dataset when the user accesses the FLA module of LipidOne. FLA operates on the lipid names contained in the user input data matrix and extracts class and chain information using a rule-based parser. Class-level indices (e.g. LPC/PC, TG/CE, Structural/Energetic) are computed whenever the relevant lipid classes are present, irrespective of whether lipids are reported as molecular species or sum compositions. Indices that require explicit fatty-acyl chains (e.g. AA/DHA, n-6/n-3, specific ferroptosis-related chains) are computed only for samples in which the corresponding chains can be identified in the lipid names; when the required components are absent or their denominator is zero, the index value is set to missing (NA) and will not be displayed. Thus, datasets reported purely at sum-composition level yield a reduced subset of FLA indices, whereas molecular-species and mixed datasets allow a larger fraction of indices to be calculated, without inferring any chain composition beyond what is provided.

The use of ratio-based definitions ensures that the indices remain interpretable, scalable, and robust across experimental platforms, minimizing dependency on absolute quantification. The resulting numerical values serve as variables for statistical analysis and group comparison enabling the identification of biologically meaningful differences across experimental conditions, as illustrated in the following sections.

2.3 From functional indices to predicted protein networks

In FLA, the link between functional lipid indices and enzyme-level hypotheses is explicitly reaction-based and follows the same philosophy as the predicted protein network that we previously described for pathway-based analyses of lipid classes, molecular species and lipid building blocks in LipidOne (Alabed et al. 2024). In that setting, significantly altered lipids are mapped to the enzymes that catalyse reactions among lipids, and these enzymes are then used as seeds for a STRING interaction network. Here we generalise this framework by using functional indices as the starting variables. For each enzyme, we consider the set of lipid reactions it catalyses and, for each group, compute aggregate “substrate” and “product” weights by summing the abundances of the corresponding lipids. For a simple reaction A + B → C + D, this corresponds to comparing, between experimental and control groups, the combined abundance of A and B with that of C and D. When the product-to-substrate balance shifts toward products in the experimental group (relative to control), we interpret this pattern as a net forward drive through the reaction and assign the enzyme an activity-oriented increase; the opposite pattern suggests reduced net flux. Importantly, FLA does not estimate protein concentrations and does not provide a direct kinetic measurement of enzyme activity. Instead, it uses curated stoichiometric relationships between lipids and the observed lipid phenotype to map, for each enzyme, an activity-oriented direction of change that is mechanistically consistent with the underlying reaction network. The same reaction-based mapping has been implemented in LipidOne for class-, species- and building block-based analyses since its initial release; the FLA module reuses this framework but replaces individual lipids with functional indices as the input layer, after assigning each index to its curated set of enzymes as described above.

2.4 FLA workflow

The FLA module integrates statistical and visualization tools to detect and interpret alterations in lipid function across experimental conditions. All analyses are implemented in R and executed server-side, ensuring reproducibility, transparency, and scalability. Starting from the user’s lipidomics dataset, the platform computes a panel of 42 mathematically defined functional indices, which act as the primary variables. These indices are statistically compared across groups (typically using t-tests or ANOVA, depending on the number of groups), and the resulting log_2_ fold changes and P-values drive the plots and summaries. Unless otherwise stated, P-values reported for FLA indices are unadjusted; given the limited number of pre-defined indices tested (≤42), we report raw P-values for interpretability. An application of these analyses is shown further below in the text.

As the general philosophy of LipidOne, the tools of FLA module, have been grouped into four categories:

2.4.1 Univariate statistical analysis

Functional summary: Compares all indices between two selected groups. Bar Plot show the group mean log2 fold change, while the error bars depict the standard error of the log2 fold change estimated by first-order error propagation (delta method) from within-group variability and sample size. This Standard Error-based approach is coherent with the Welch test used for significance and provides a compact, distribution-light estimate of uncertainty. Bars are colored according to the predominant lipid function of each index (Fig. 1A). Radar Plot show a companion radar chart summarizes the overall functional landscape (Fig. 1B). Functional Dominance show a three-bar panel (Structural, Signaling, Energy) that reports, at a glance, the prevailing functional shift (Fig. 1C). For each function, we aggregate the magnitude of change by summing the absolute log2FC of indices that pass significance and normalizing by the number of indices effectively evaluated in that function, to avoid bias from missing/filtered indices. Larger bars indicate a stronger net deviation of that function in the experimental group. Functional Summary outputs are accompanied by CSV exports reporting per-index statistics (log2FC, SE, P-value from Welch’s test, and other metrics when available). For indices that meet significance thresholds, the CSVs also include a standardized Interpretation field with concise phrases tailored to the index identity and direction of change, enabling immediate reuse in figure tables and narrative summaries.

Examples of graphical outputs generated by the FLA module in LipidOne. (A) Bar plot comparing all indices; (B) Radar plot summarizing the overall functional landscape; (C) Functional Dominance plot showing a three-bar panel (Structural, Signaling, Energy) to study functional shifts within a study; (D) Functional Volcano plot; (E) Functional Bubble plot displaying a bubble heatmap of top index differences across groups; (F) Functional biomarker discovery table; (G) Functional Correlation Circos plot; (H) PCA score and loading plots; (I) PLS-DA score plot; (J) PLS-DA loading plot; (K) VIP score bar plots; (L) Functional heatmap of indices; (M) Predicted functional protein network.

Functional volcano plot: Highlights the magnitude of change (log_2_ fold change) and significance for each index. In FLA, a three-panel volcano separates indices by functional category. A results table lists indices that are both significant and beyond the user-defined variation threshold (Fig. 1D). The significant indices are presented together with brief sentences explaining the biochemical meaning of the differences.

Functional bubble plot: Displays a bubble heatmap of the top ten indices across groups; bubble size reflects significance and color encodes fold change. Each index is accompanied by a concise interpretive sentence (Fig. 1E).

Functional biomarker discovery: Ranks indices by discriminative performance (P-value, ROC AUC—polarity-corrected when needed—statistical power, Cohen’s d). Outputs a Top-20 table (CSV + PNG) for rapid screening and prioritization (Fig. 1F).

Functional correlation circoplot: Visualizes pairwise correlations between indices to reveal co-regulation and functional synergy (Fig. 1G).

2.4.2 Multivariate statistical analysis

Instead of analyzing hundreds or thousands of molecular lipid species, FLA works with a compact set of functional indices (currently 42) that explicitly encode lipid roles (structural, energetic, signaling). This biochemistry-driven reduction in data complexity makes principal components and latent variables easier to interpret, because each feature reflects a function rather than an individual species and reduces co-linearity that often occurs in lipidomics data. By aggregating related species, the indices improve signal-to-noise and often sharpen group separation in PCA and PLS-DA. (see details in in the results section). Finally, VIP scores and loadings calculated on indices map directly onto functional biochemical mechanisms, revealing unexpected links between lipid functions.

Functional PCA: PCA on the matrix of functional indices (centered and auto-scaled by default; scaling options are user-configurable). Produces score plots (sample clustering) and loading plots (index contributions), enabling dimensionality reduction and identification of dominant functional axes (Fig. 1H).

Functional PLS-DA: Performs Partial Least Squares Discriminant Analysis on the functional index matrix. Outputs include score plots with confidence ellipses (Fig. 1I), loading plot (Fig. 1J), and VIP score bar plots (Fig. 1K). This supervised method improves discrimination between experimental groups and highlights key functional biomarkers.

2.4.3 Clustering analysis

Functional heatmap: Creates a clustered heatmap of indices across all samples of groups selected (Fig. 1L). Hierarchical clustering is applied to both samples and indices. User can select the number of indices to include and apply one of seven diverse algorithms of distances, aiding in the identification of group-specific functional or nutritional profiles and outlier patterns.

2.4.4 Lipid system biology

Predicted functional proteins network: This tool links each lipid functional index to relevant proteins and enzymes using the STRING database (Fig. 1M). The resulting network graph highlights biological pathways and regulatory modules, with nodes colored according to cluster annotations, thus providing a mechanistic link between lipidomic alterations and the proteins potentially involved in their regulation or response. To enhance interpretability and flexibility, the user can adjust key parameters of the analysis, including the statistical thresholds used to define significant indices (P-value and log_2_ fold change), the type of STRING interaction network to be queried (full network or physical interactions only), and the confidence score threshold for STRING interactions. Additionally, users can specify the number of neighboring proteins to retrieve for each lipid-associated hit and choose the annotation system used to color the protein nodes in the network. Available annotation sources include Gene Ontology (Biological Process, Molecular Function, and Cellular Component), KEGG Pathways, Pfam Domains, InterPro Families, UniProt Keywords, and Reactome Pathways. This high degree of configurability allows users to tailor the network reconstruction to the specific biological context of their study, supporting hypothesis generation and functional interpretation of lipidomic data.

3 Results

To demonstrate the ability of the FLA module to detect biologically meaningful alterations in lipid metabolism, we applied it in two independent lipidomic datasets: one investigating the role of PDK1 in cardiomyocyte lipid utilization (Atser et al. 2025) and another characterizing lipidomic signatures associated with hepatocellular carcinoma and chronic hepatitis C virus–related conditions (Caponigro et al. 2023). Detailed analysis reports are provided in the Supplementary Materials (see the Case studies using Functional Lipid Analysis section).

FLA turns untargeted lipidomes into mechanistically interpretable readouts with organellar context. Rather than listing hundreds of species, FLA quantifies a curated set of functional indices spanning signaling, structural, and energy domains and ties each index to a concise biochemical interpretation. The result is an analysis that tells not only what changed, but what it may imply functionally. Where enzyme or organelle processes are mentioned, these should be interpreted as hypotheses inferred from index behavior and STRING-contextualization, not direct measurements. FLA is not an enrichment analysis; it is a hypothesis-generating, mechanistically informed interpretation based on quantitative lipid features and prior biochemical knowledge.

Across two independent case studies, FLA both replicated the key findings of the original works and extended them with testable hypotheses that are difficult to obtain from class-level summaries alone. In cardiomyocytes, FLA separated storage-positive from mobilization-negative indices and mapped the phenotype onto a Pdk1–Pdk4–Pdha1 network interfacing with Pnpla2 (ATGL) and FAO enzymes. In HCV/HCC plasma, FLA confirmed lysophospholipid depletion in HCC and newly highlighted ether-lysophospholipid turnover, putative mitochondrial/inner-membrane remodeling signals (LPE/PE), and ceramide-axis differences. Because indices are ratio-based and statistics are attached to each readout (effect size, AUC, power), the outputs are interpretable, portable, and biomarker-ready.

Importantly, FLA links statistical outputs to standardized biochemical and mechanistically oriented statements, converting lipidomic ratios into actionable interpretations that are accessible to clinical and translational teams without deep lipid biochemistry expertise.

To our knowledge, FLA is the first general-purpose bioinformatics module that goes beyond lipid ontology/characteristic enrichment by (i) organizing lipidomes into directional functional axes supported by quantitative indices, (ii) overlaying significant functional signals onto protein interaction networks to propose enzyme-level hypotheses, and (iii) delivering a compact suite of publication-grade visuals (summary bars, bubble heatmap, functional volcano, top-index table, protein network) in minutes. This positions lipidomics not only as a descriptive assay but as a strategic tool for mechanism discovery and translational biomarker development.

4 Limitations and future directions

Accurate computation of several indices requires molecular-species–level annotation (chain length/unsaturation and, where relevant, O-/P-linkages). Datasets reported only as sum compositions limit index coverage.

Index accuracy depends on class coverage and identification quality. Ratio-based design mitigates but does not eliminate variability due to extraction, chromatography, adduct handling, or missing values (zero handling can inflate ratios at low abundance).

Some indices are correlated by construction; they should be interpreted as a panel rather than independent tests. Statistical control remains essential.

FLA infers function from correlations in lipidomes and maps results onto STRING; it does not measure enzyme activity. Network-level hypotheses (e.g. ATGL, SCD1, ELOVLs) require orthogonal validation (proteomics, flux assays).

Finally, the index library is curated and finite: rare contexts or novel lipid chemistries (e.g. oxidized species, cardiolipin microheterogeneity) may not yet be captured.

Future work will expand the index library (including oxidized lipids and pathway-specific composites), add data-driven index discovery and uncertainty quantification, tighten cross-study harmonization (QC rules, FDR defaults, zero-handling), and deepen multi-omics integration (proteome, transcriptome, 13C-flux) so that FLA can progress from functional readouts to testable, enzyme-level mechanisms at scale.

Supplementary Material

btag090_Supplementary_Data