A collaborative study linking epigenomics to intervention outcomes

A new study published in Clinical Epigenetics addresses this gap directly. Led by Flavio Palmieri (B2SLab, IRIS — Universitat Politècnica de Catalunya) in close collaboration with Josep C. Jiménez-Chillarón (Biophysics Unit, Department of Physiological Sciences, School of Medicine, University of Barcelona, L’Hospitalet de Llobregat, Spain), the work investigates whether baseline DNA methylation profiles — measured before any intervention — can forecast a child’s response to a six-month moderate lifestyle programme.

The study analysed a cohort of 26 pre-pubertal children with obesity (7–10 years of age), recruited at the Sant Joan de Déu Barcelona Children’s Hospital. Children were classified as High Responders (HR) or Low Responders (LR) according to the change in their age- and sex-standardised BMI z-score (ΔzBMI) over the intervention period.

From 850,000 CpG sites to eight predictive markers

Using whole-blood DNA methylation data from the Infinium Methylation EPIC 850K array, the team profiled nearly 850,000 CpG sites per subject. After rigorous quality control and preprocessing — including correction for batch effects and estimated leukocyte composition — 788,373 CpGs were available for statistical analysis.

Differential methylation analysis using leave-one-out (LOO) regression identified 214 CpG sites consistently associated with the HR/LR classification at baseline. Crucially, the intervention itself did not alter the global DNA methylation landscape, confirming that these markers reflect pre-existing epigenetic differences and not adaptive responses to the programme.

To arrive at a clinically applicable model, the team iteratively reduced the 214 candidate markers to a minimal predictive subset using partial least squares (PLS) regression. The final model comprises eight CpG sites, including markers in or near the genes GSDMD, GFRA1, NRP2, NLRC5, SPTLC2, and LTBP3 — genes implicated in inflammatory regulation, metabolic signalling, and lipid metabolism.

A predictive model with 84% classification accuracy

The eight-marker PLS model achieves an area under the ROC curve (AUC) of 84%, with a sensitivity of 77% and a specificity of 85% at the optimal classification threshold. On the study cohort, the model correctly classified all children in the Low Responder group and misclassified only two High Responders — a strong performance given the modest cohort size.

Biological insight: sphingolipid metabolism as a key pathway

Beyond predictive accuracy, the study provides important biological context. An enrichment analysis of the 214 prioritised CpG sites consistently highlighted the sphingolipid metabolism pathway in both KEGG and Gene Ontology databases. Two of the eight final predictive markers — SPTLC2 (Serine Palmitoyltransferase Long Chain Base Subunit 2) and SGMS1 (Sphingomyelin Synthase) — are rate-limiting enzymes in the de novo synthesis of ceramides and sphingolipids.

An integrative epigenomics–metabolomics analysis further revealed that the methylation of the SPTLC2-associated CpG site correlates significantly with plasma ceramide levels, which in turn respond to the lifestyle intervention. These findings position sphingolipid regulation as a molecular link between epigenetic predisposition and individual metabolic responsiveness in childhood obesity.

Implications for personalised paediatric medicine

The results demonstrate, for the first time, that a compact DNA methylation signature measurable from peripheral blood can stratify pre-pubertal children with obesity according to their likely response to a lifestyle intervention — before the intervention is initiated. Because blood-based DNA methylation can be measured reliably, non-invasively, and at relatively low cost, this approach is compatible with clinical workflows.

The authors emphasise that broader, more diverse cohort studies are required to validate the generalisability of these eight markers and to refine the model for routine clinical use. Nevertheless, the study establishes a clear proof of concept and provides a foundation for the development of personalised treatment strategies in paediatric obesity.

This work is the result of a productive collaboration between B2SLab and the group of Josep C. Jiménez-Chillarón at the University of Barcelona, and reflects the laboratory’s ongoing commitment to integrating epigenomics, metabolomics, and machine learning for translational biomedical research.

Reference: Palmieri F, Castellano-Escuder P, Parra-Vargas M, Leal-Witt MJ, Ramón Krauel M, Lerin C, Perera A, Jiménez-Chillarón JC. Predictive epigenetic biomarkers of successful weight-loss intervention in pre-pubertal children with obesity. Clinical Epigenetics (2026) 18:108. https://doi.org/10.1186/s13148-026-02104-1

We suffered (quite a lot, to be honest)

Eleven floors of stairs. No elevator. No shortcuts. Lungs burning from floor three onwards. Legs questioning every life decision by floor seven. And yet, everyone made it to the top — which, given the average daily step count of a research lab, is something to be genuinely proud of.

The team showed up, laced up, and gave it everything they had. The flower leis and the badges reading “Jo he pujat 12 plantes” (“I climbed 12 floors”) tell the whole story.

Blanca wins the female category

The highlight of the day: Blanca took first place in the female classification, earning a well-deserved diploma and a round of applause in the ETSEIB lobby. An outstanding result and a proud moment for the whole team.

Beyond the race

The Cursa d’Escales was just one part of a full day at the school’s annual festival. The programme also included a popular lunch (dinar popular), an ETSEIB building tour, and a gymkhana — a full afternoon of activities that brought together students, staff, and researchers from across the school.

Congratulations to the organizers, to all participants, and especially to Blanca. Same time next year?

Research in computational metabolomics

Maria’s doctoral work spans the full arc of modern metabolomics: from the algorithmic challenge of annotating unknown molecules, to the clinical application of metabolite panels for disease prediction, to the use of deep learning on molecular structure itself.

What lies ahead: DTU and the green transition

Maria joins the Biotechnology Research Institute for the Green Transition (BRIGHT) at the Technical University of Denmark (DTU), where she will work within the Data Science Platform and the Multi-omics Network Analytics group, in collaboration with the DTU Microbes Initiative.

It is a fitting destination. The challenges of the green transition — understanding microbial communities, engineering biological systems, interpreting complex multi-omic data — call for exactly the computational metabolomics expertise that Maria has built. We have no doubt she will make a strong contribution there.

Thank you, Maria

Maria has been a cornerstone of B2SLab’s metabolomics research for years. Her rigour, her willingness to tackle hard problems, and her generosity in sharing knowledge with the rest of the team have made her an irreplaceable colleague. Watching her grow from a talented master’s student into the accomplished scientist who defended today has been a privilege.

The field of computational metabolomics is better for the work she has done. And wherever this next chapter takes her, we will be watching with great pride.

Congratulations, Dr. Barranco Altirriba. The best is yet to come.

/Àlex

On deep learning for chronic disease modeling

Enrico’s doctoral work focused on deep learning for Electronic Health Records, with a particular emphasis on modeling the progression of chronic diseases such as Type 2 Diabetes Mellitus (T2DM) and Chronic Obstructive Pulmonary Disease (COPD).

His research addressed a fundamental challenge in clinical AI: how to extract meaningful, actionable predictions from the messy, irregular, and heterogeneous data generated by routine clinical care.

Avoiding the use of text-based or LLM models, Enrico’s approach embraced the full longitudinal richness of EHRs — sequences of diagnoses, prescriptions, lab results, and clinical events spanning years — and applied modern deep learning architectures (transformers, attention mechanisms, recurrent networks) to model disease trajectories and forecast clinically relevant outcomes. This approach has pioneered b2slab first actions into building Foundational Models in Health, aiming to a full Diabetes Foundational Models in which other team members are currently working, direct consequence of Enrico’s work.

The thesis built upon a series of published contributions with scientific depth and practical relevance:

Longitudinal deep learning clustering of Type 2 Diabetes Mellitus trajectories using routinely collected health records Journal of Biomedical Informatics, 135, 104218, 2022. One of the foundational papers of the thesis, introducing unsupervised deep learning to identify patient subgroups in T2DM from real-world health records.

Mapping layperson medical terminology into the Human Phenotype Ontology using neural machine translation models Expert Systems with Applications, 204, 117446, 2022. A methodological contribution showing how NLP and neural machine translation can bridge the gap between patient language and formal clinical ontologies.

A deep attention-based encoder for the prediction of Type 2 Diabetes longitudinal outcomes from routinely collected health care data Expert Systems with Applications, 274, 126876, 2025. The DARE model, a transformer-based architecture trained on data from over 200,000 individuals, capable of predicting comorbidity onset, treatment changes, and glycemic control targets with high accuracy.

A BERT base model for the analysis of Electronic Health Records from diabetic patients 2024 IEEE Engineering in Medicine and Biology Conference (EMBC), 2024. Demonstrating the power of large language model pretraining on clinical sequences to build general-purpose representations of patient health.

Deep Survival Analysis of Longitudinal EHR Data for Joint Prediction of Hospitalization and Death in COPD Patients arXiv preprint arXiv:2511.05960, 2025. Extending the framework to COPD — a complex, multi-morbid condition — and tackling survival analysis with competing risks from longitudinal records.

Smile-to-BERT: A BERT architecture trained for physicochemical properties prediction and SMILES embeddings generation (with M. Barranco-Altirriba, V. Würf, J.K. Pauling, A. Perera-Lluna) — bioRxiv, 2024. A contribution to molecular property prediction, showing the breadth of Enrico’s interest in applying language model ideas beyond the clinical domain.

Thank you, Enrico

Watching a PhD student grow from their first steps in the lab to the moment they stand before a committee and defend years of original research is one of the most rewarding experiences I can have.

Enrico approached every challenge with rigor, intellectual honesty, and … good humor. He has contributed not only excellent science but also a great deal of energy and warmth to the lab.

The field of clinical AI is better for the work he has done. We have no doubt that wherever he goes next, he will continue to make an impact.

Congratulations, Dr. Manzini. The Legends were right — it does end someday.

/Àlex

Nearly 10,000 genes respond to a single race

The scale of the response was striking. Immediately after finishing, we found 9,874 differentially expressed genes compared to the pre-race baseline. This is a substantial fraction of the expressed genome — the body is not making small adjustments; it is undergoing a broad molecular reorganisation.

The major pathways activated immediately after the race were:

• Immune system mobilisation — strong up-regulation of inflammatory and innate immune pathways, consistent with the physical trauma of sustained exertion.
• Oxidative stress response — genes involved in counteracting reactive oxygen species, produced in large amounts by working muscles.
• Lipid metabolism — shifts in how the body processes fat, reflecting the switch to fat oxidation that occurs during prolonged endurance exercise.

These responses are not surprising in isolation — we knew exercise activates them. What the study adds is a genome-wide, unbiased view of their magnitude and co-occurrence in real-world non-professional athletes.

The body hasn’t recovered 24 hours later

Perhaps the most interesting finding is what happens the day after. Twenty-four hours post-race, most gene expression had returned toward baseline — but 279 genes were still significantly altered. These lingering changes were predominantly linked to mitochondrial function and energy production pathways.

Mitochondria are the powerhouses of cells, and after a marathon they have been working at or near capacity for several hours. The persistent transcriptomic signal suggests that mitochondrial repair and adaptation continues well beyond the acute post-race phase. For non-elite athletes, whose bodies are not as adapted to this level of effort as professional runners, this recovery window may have practical implications for training scheduling and injury prevention.

A window into exercise biology in everyday athletes

Most exercise transcriptomics studies focus on elite or highly trained athletes, whose molecular responses may differ substantially from the population at large. By focusing on non-elite runners — the kind who train regularly but compete recreationally — this study provides a more representative picture of what marathon running does to the human body.

The findings also highlight whole blood as an accessible and information-rich tissue for monitoring physiological state: no biopsies needed, just a blood draw.

The paper is available at: Ezquerra-Condeminas P, Martin-Fernandez L, Cardenas A, Sibila O, Borràs N, Vidal F, Perera-Lluna A, Soria JM. Gene expression profiling of whole blood samples following marathon running in non-elite athletes. Biology of Sport, 2026. https://doi.org/10.5114/biolsport.2026.158303

How the study was done

We used an untargeted metabolomics approach — casting a wide net across thousands of blood metabolites — on plasma samples from 352 individuals in a discovery cohort. These samples were collected on average 7.4 years before the subjects either developed T2D (143 cases) or remained healthy (209 controls). This design — a nested case-control within a prospective cohort — is key: it means the metabolite signal precedes the diagnosis, not just reflects it.

From the discovery phase, six candidate metabolites emerged as significantly associated with incident T2D: guanine, ecgonine, adenine, pregnenolone sulfate, phenyl sulfate, and citrulline. We then validated the most promising ones using targeted metabolomics in a larger, independent cohort of 2,044 individuals.

What survived validation

Three metabolites held up in the independent cohort: guanine, pregnenolone sulfate, and citrulline.

• Guanine is a purine nucleobase involved in nucleotide metabolism. Its association with T2D risk connects to emerging evidence that purine metabolism is disrupted in the early stages of diabetes development.
• Pregnenolone sulfate is a neurosteroid precursor that may reflect alterations in steroid hormone metabolism, which is increasingly recognised as playing a role in metabolic disease.
• Citrulline is an amino acid linked to arginine and nitric oxide metabolism, with potential ties to vascular function and insulin signalling.

Beyond individual metabolites, the study also identified nucleotide metabolism and ABC transporter pathways as consistently altered in pre-diabetic individuals, pointing to biological processes worth investigating for early intervention.

What this means for prevention

The ability to detect T2D risk years before diagnosis is clinically meaningful. Current screening relies heavily on fasting glucose and HbA1c, which typically become abnormal only once significant beta-cell dysfunction has occurred. Metabolomics markers like those identified here could complement standard screening, helping identify individuals who would benefit most from preventive lifestyle interventions while there is still time to act.

The study represents another step in B2SLab’s ongoing work at the intersection of metabolomics and diabetes, building on a body of research that includes EHR-based trajectory prediction and lipidomics in diabetes complications.

The paper is available at: Barranco M, Granado M, Yanes Ó, et al. Guanine and pregnenolone sulfate are associated with incident type 2 diabetes in two independent populations. Frontiers in Endocrinology, 2025. https://doi.org/10.3389/fendo.2025.1706886

A large paediatric dataset and a new model

In a new study published in Clinical Nutrition, our team — in collaboration with Hospital Sant Joan de Déu Barcelona — analysed 4,368 paediatric serum samples to build a more accurate framework for interpreting copper measurements in children.

Using the same continuous reference interval methodology we developed for cerebrospinal fluid amino acids and neurotransmitter metabolites, we modelled copper concentration as a function of age using nonlinear regression, then assessed the contribution of inflammatory markers — specifically erythrocyte sedimentation rate (ESR), fibrinogen, and C-reactive protein (CRP) — to residual variability.

The results were clear: inflammation elevated measured copper by approximately 24% on average. Children with active inflammation would therefore appear to have higher copper than they truly do after accounting for their biological state. Using uncorrected reference intervals in such patients risks under-diagnosing copper deficiency.

The solution: an inflammation-adjusted score

To address this, we built a composite inflammation score using partial least squares regression across the three inflammatory markers. This score captures the combined inflammatory load of a patient and allows copper measurements to be adjusted accordingly before comparing them against the reference interval.

The result is a proof-of-concept model that produces age- and inflammation-adjusted reference intervals — the first of their kind for paediatric serum copper.

Practical impact

For clinicians managing children with suspected metabolic disorders, chronic diseases, or nutritional deficiencies, this work offers a more reliable interpretive tool. It also extends our group’s broader programme of establishing statistically robust, age-continuous reference intervals for paediatric laboratory biomarkers — a methodology we have previously applied to cerebrospinal fluid amino acids and biogenic amines.

The paper is available at: Rodriguez-Gonzalez H, Arias A, Poyatos E, et al. Proof of concept for an age- and inflammation-adjusted model for the establishment of pediatric serum copper reference intervals. Clinical Nutrition, 2026. https://doi.org/10.1016/j.clnu.2026.106586

A lipidomics approach to vascular risk

In a study published in Cardiovascular Diabetology, our team analysed the serum lipidomic profile of 513 individuals — 151 with type 1 diabetes (T1D), 155 with T2D, and 207 non-diabetic controls — all of whom underwent carotid ultrasound to assess SCA. Lipidomics is the systematic measurement of lipid species in a biological sample; instead of measuring total cholesterol or triglycerides, it profiles hundreds of individual lipid molecules with chemical specificity.

All participants underwent ultrahigh-performance liquid chromatography-electrospray ionisation tandem mass spectrometry (UHPLC-ESI-MS/MS), producing detailed lipid profiles that were then associated with the presence and extent of carotid atherosclerotic plaques.

The findings

In type 2 diabetes, 27 unique lipid species were associated with SCA — a specificity that was not seen with the same clarity in T1D or controls. The main classes involved were:

• Phosphatidylcholines (PCs): Ten species were up-regulated in T2D patients with SCA, while four PC species containing polyunsaturated fatty acids (PUFAs) were down-regulated. PUFAs are known to have anti-inflammatory properties; their reduction may reflect a lipid environment more conducive to plaque formation.
• Diacylglycerols (DAGs): One was down-regulated, while three others — particularly in T2D patients without dyslipidaemia — were positively associated with SCA. DAGs are signalling lipids that can activate protein kinase C pathways involved in endothelial dysfunction.

Particularly notable was the subgroup analysis: among T2D patients who smoke and those without dyslipidaemia (who would not typically be flagged as high cardiovascular risk), specific lipid associations emerged that were not visible in the overall population. This suggests that lipidomics could identify at-risk individuals within groups that standard clinical markers would miss.

Toward lipid-informed precision medicine in diabetes

These results add to the growing body of evidence that standard lipid panels — total cholesterol, LDL, HDL, triglycerides — are an incomplete picture of cardiovascular risk. The species-level resolution of lipidomics reveals distinctions that aggregate measures obscure.

For T2D patients, where cardiovascular risk management is a central clinical challenge, lipid signatures like those identified here could eventually complement standard risk calculators, helping clinicians prioritise patients for intensive preventive therapy.

The paper is available at: Barranco M, Rossell J, Alonso N, et al. Lipidomic analysis reveals metabolism alteration associated with subclinical carotid atherosclerosis in type 2 diabetes. Cardiovascular Diabetology, 2025. https://doi.org/10.1186/s12933-025-02701-z

What singIST does

Our new method, singIST (single-cell Integrative Similarity Tool), addresses this directly. Published in PLOS Computational Biology, singIST takes single-cell RNA sequencing (scRNA-seq) data from both a disease model and human patients and computes how well the model captures the human disease at three nested levels:

• Pathway level — are the same biological processes disrupted?
• Cell type level — are the same cell populations involved?
• Gene level — are the same individual genes differentially expressed?

Critically, singIST accounts for two biological realities that simpler methods ignore: gene conservation between species (not every human gene has a functional mouse equivalent), and differences in cell type composition between species (a cell type present in human skin may be rare or absent in mouse skin).

The method produces interpretable similarity scores at each level, making it possible to say not just “this model is broadly similar” but “this model captures keratinocyte biology well but misses the dendritic cell response entirely”.

Testing on skin disease models

We validated singIST on three mouse models of atopic dermatitis (AD), a common inflammatory skin disease, and additionally applied it to human hidradenitis suppurativa (HS) data. These diseases share some features but differ mechanistically, making them a good test of the method’s discriminative power.

singIST reproduced established knowledge — for instance, correctly identifying which mouse model best recapitulates the Th2-skewed immune response characteristic of human AD — while also generating new hypotheses about specific pathways and cell types that each model captures or misses.

Why it matters for drug development

Choosing the wrong preclinical model is a major driver of clinical trial failure. singIST offers a systematic, reproducible way to audit candidate models before committing to expensive and time-consuming experiments. Rather than discovering post-hoc that a model missed a key pathway, researchers can now make that comparison up front.

The tool is freely available and designed to work with standard scRNA-seq pipelines, meaning it can be integrated into existing workflows without requiring additional data collection.

This work was led by Aitor Moruno-Cuenca in collaboration with Francesc Fernández-Albert and colleagues at B2SLab (IRIS-UPC) and the University of Michigan.

The paper is available at: Moruno-Cuenca A, Picart-Armada S, Bogle R, Fox J, Tsoi LC, Gudjonsson JE, Perera-Lluna A, Fernández-Albert F. singIST: An integrative method for comparative single-cell transcriptomics between disease models and humans. PLOS Computational Biology, 2026. https://doi.org/10.1371/journal.pcbi.1014002

Not all polyps are the same

Adenomas come in three main histological subtypes — tubular, tubulovillous, and serrated — which are known to carry different risks of malignant transformation. Tubular adenomas are the most common and generally lower-risk; serrated adenomas follow a distinct molecular pathway and are more likely to be missed during colonoscopy; tubulovillous adenomas combine features of both.

Despite these clinical differences, the molecular signatures that underlie them had not been comprehensively mapped. In a study published in Cancers, our group — together with collaborators at UniversalDx — performed a detailed multi-omic characterisation of these three subtypes.

Three molecular layers, three stories

We analysed methylation, copy-number alterations (CNA), and somatic mutations across the adenoma subtypes, in the context of the progression from normal colon tissue through advanced precancerous lesions (APLs) to early-stage CRC.

The results reveal distinct molecular identities at each layer:

Methylation showed 2,321 significantly altered regions across subtypes. Serrated adenomas were enriched for changes in cAMP signalling and stem cell pluripotency pathways — consistent with their unique biological behaviour. Tubular and tubulovillous adenomas showed enrichment for WNT signalling, one of the most studied cancer pathways.

Copy-number alterations were predominantly found in tubular and tubulovillous adenomas, with recurrent signals in chromosomes 7, 12, 19, and 20. Importantly, the CNA profile of early-stage CRC differed — chromosomes 7, 8, and 20 — suggesting that the progression from APL to carcinoma involves distinct genomic events, not just more of the same.

Mutations reinforced the subtype-level differences, with specific alterations characteristic of each histological class.

Why this matters for early detection

These molecular signatures are precisely the kind of information that early detection tests need. Whether blood-based, stool-based, or tissue-based, a test that can distinguish adenoma subtypes and estimate progression risk could dramatically improve how we triage patients — reducing unnecessary colonoscopies for low-risk individuals while prioritising those with high-risk adenomas.

The findings also contribute to the emerging picture of colorectal cancer as a collection of molecularly distinct diseases rather than a single entity, each subtype potentially requiring different surveillance strategies.

The paper is available at: Mattia F, Higareda J, Canal P, Bertossi A, Perera A, Herbert M, Kruusmaa K. Colorectal adenoma subtypes exhibit signature molecular profiles: unique insights into the microenvironment of advanced precancerous lesions for early detection applications. Cancers, 2025. https://doi.org/10.3390/cancers17040654

A systematic measurement campaign

In a study published in Indoor Air, our team conducted a careful measurement campaign across primary schools and university classrooms in Catalonia. Over five months, we performed 33 individual measurements in rooms with different topologies and ventilation systems — natural ventilation, mechanical ventilation, and mixed systems.

Rather than simply comparing sensor readings to a reference, we used the rate of change of CO2 (dCO2/dt) as a diagnostic tool. This approach allows tendencies in CO2 evolution to be tracked in a way that is sensitive to the local airflow environment around a sensor — not just the average room concentration.

What the data showed

The results were striking in their practical implications. Sensor position and ventilation strategy caused spatial CO2 discrepancies exceeding 100 ppm in many measurement scenarios — a difference that in many guidelines would shift a room from acceptable to concerning ventilation status.

Critically, these differences were larger and more systematic than the variation between sensor models from different manufacturers. In other words, buying a more expensive sensor does not help you if it is placed in the wrong spot.

Specific findings included:

• Sensors placed near supply air inlets (where fresh air enters) consistently read lower concentrations than those near return air vents or occupied zones.
• In naturally ventilated rooms, CO2 distribution was highly non-uniform and dependent on window openings — making single-point measurements particularly unreliable.
• In mechanically ventilated rooms, sensor position relative to the airflow path had a systematic and predictable effect on readings.

Practical recommendations

The study derives concrete guidelines for future measurements:

1. Avoid placing sensors near air supply inlets or directly in the path of draught from open windows.
2. In mechanically ventilated rooms, sensor position relative to supply and return vents should be documented and reported alongside measurements.
3. For classroom monitoring intended to represent occupant exposure, sensors should be positioned in the occupied zone at breathing height.

These recommendations are particularly relevant for schools, where CO2 monitoring is increasingly mandated by health authorities — but standards for sensor placement remain inconsistent or absent.

The paper is available at: Marín D, Ruiz de Alegria A, Canals Casals L, Macarulla M, Fonollosa J. The reliability of CO2 measurements using low-cost sensors: a study of sensor positioning and ventilation strategies in classrooms. Indoor Air, 2025. https://doi.org/10.1155/ina/5517242