Setting the Frontier of Aging Biology with Liquid Foundation Models

LongevityBench enables systematic evaluation of language models on interpreting aging- and longevity-related data across multiple biological modalities, such as clinical records, DNA methylation studies, transcriptomic profiles, plasma proteomics, and genetic evidence. In this post, we examine the performance of two LFM2 models after domain-specific fine-tuning, showing that compact models are competitive with much larger frontier models on tasks relevant to aging research.

The resulting LFM2-1.2B-Longevity and LFM2-2.6B-Longevity are available on Hugging Face alongside LongevityBench. The accompanying paper in Cell Press provides the complete methodology, benchmark results, and additional analyses.

Why aging biology needs a shared benchmark

Aging is a systemic biological process and the principal risk factor for diseases responsible for the majority of human morbidity and mortality [1]. Efforts to characterize its molecular and physiological basis have generated extensive datasets spanning clinical biomarkers, DNA methylation, gene expression, proteomics, and genetic evidence. These modalities capture distinct but complementary aspects of aging biology, from molecular regulation to organism-level phenotypes. Their heterogeneity, however, has largely limited computational models of aging to individual data types and narrowly defined prediction tasks.

Aging clocks, for example, are typically trained on a fixed set of features and return a numerical estimate of age or aging-related risk [2, 3, 4]. More recent biological foundation models can learn richer representations, but they also tend to depend on modality-specific architectures and preprocessing pipelines [5, 6, 7]. Extending these models to a new data modality often requires building and validating a new model.

A general-purpose language model offers a different approach. Measurements from several modalities can be represented as structured text and presented to the same model, allowing it to address multiple types of biological questions. Whether this approach works, however, must be established empirically. A fluent response does not demonstrate that a model has used the measurements correctly. A meaningful evaluation must therefore test whether a model can use the biological measurements provided in the prompt to reach the correct answer. LongevityBench was designed to address this evaluation gap.

^{How biological measurements are represented in a prompt. To make heterogeneous data compatible with a language model, we convert each biological record into a structured textual representation. The figure above shows sample-level representations from each biodata domain: clinical measurements, DNA methylation, plasma proteomics, transcriptomic profiles, and genetic evidence.}

What LongevityBench measures

LongevityBench comprises 17 tasks and 25,457 prompts drawn from five biodata domains: clinical records from NHANES [8, 9], DNA methylation studies from GEO [10], transcriptomic profiles from GTEx [11], plasma proteomics from three public Olink studies [12, 13, 14], and genetic evidence from OpenGenes [15], CellAge [16, 17], and SynergyAge [18]. Each task presents an experimental description together with structured biological or clinical data and asks for a defined prediction. The four question formats are binary classification, pairwise comparison, multiclass classification, and numeric regression.

Several source datasets appear in multiple formats. The same NHANES records, for example, support tasks that compare two participants’ ages, assign one participant to an age group, or estimate chronological age. Treating these as separate tasks allows us to examine how the requested output affects model performance while holding the underlying data domain fixed. 

We separate each evaluation task from its training counterpart using a split defined at the level of the underlying biology or study design. NHANES clinical data is split by survey wave, DNA methylation profiles by GEO study, GTEx transcriptomic and Olink proteomic data by individual, OpenGenes records by protein family, and SynergyAge records by species.

_{From biological records to benchmark prompts: Each task converts a source record into a prompt with a defined prediction target and answer format. The walkthrough above shows one benchmark example from each data domain, including input features, the correct answer, and the output answers from both LFM2 Longevity models. The complete prompts often contain additional measurements or annotations.}

Adapting LFM2 to structured aging data

The LFM2 models were trained on a multitask collection of aging-related prompts spanning clinical, epigenomic, transcriptomic, proteomic, and genetic data, together with gene- and aging-clock knowledge tasks.

LFM2-1.2B and LFM2-2.6B were each fine-tuned for three epochs with a 32k-token context window. For each size, we trained three variants: longevity data alone, longevity data mixed with 10% general chat, and longevity data mixed with 20% general chat. We included general chat data to preserve the model’s broader conversational capabilities and avoid overfitting to the structure of the longevity training prompts. We then combined these three variants by equal-weight linear merging within that model size. The benchmark results shown here come from single-pass, zero-shot evaluation with thinking disabled.

Longevity-LLMs ranked among the top performers

We compared our Longevity-LLMs with 18 frontier LLMs. The domain-adapted models often matched or exceeded much larger frontier models, with the largest model in the family (L-Qwen3.5-9B) achieving the best overall rank.

This strong performance extended across all five domains represented in LongevityBench: clinical data, genetics, transcriptomics, DNA methylation, and proteomics.

LFM2-2.6B-Longevity ranked first across all models on the NHANES pairwise-age task, which asks the model to identify the older of two participants from their clinical profiles. On the masked OpenGenes binary classification task, which asks whether a gene’s expression increases or decreases with age, LFM2-1.2B-Longevity and LFM2-2.6B-Longevity held the top two positions across all models. On the GTEx multiclass age-group classification task, which predicts a donor’s age group from gene-expression data, LFM2-2.6B-Longevity ranked second overall and LFM2-1.2B-Longevity ranked fourth; both outperformed every frontier model evaluated. LFM2-2.6B-Longevity also ranked second overall on the GEO pairwise-age task, which compares two donors using their DNA methylation profiles, outperforming every frontier model. Finally, on the Olink proteomics pairwise-age task, which compares two plasma-protein profiles to identify the older donor, both LFM2 models outperformed every frontier model evaluated.

The panels below show the highest-performing models on a subset of benchmark tasks, along with the strongest frontier model.

^{Where did the two LFMs perform strongly? Results for clinical age, masked gene-expression direction, and transcriptomic age-group classification.}

Which biological features influence the models' predictions?

We were also interested in understanding which biological features the LFM2 models relied on when making their predictions. We therefore performed a matched ablation analysis: for each benchmark prompt, we removed one biologically defined group of measurements while every other part of the prompt remained unchanged.

For each intact and ablated prompt, we calculated the difference between the logits for the correct and incorrect answers. We use this difference as a proxy for the model’s confidence in the correct answer. A decrease after ablation indicates that the removed feature group had supported the correct answer; an increase indicates that the model favored the correct answer more strongly without it.

We also count answer flips, in which a prediction changes from correct to incorrect after ablation. The logit difference captures smaller shifts in the model’s preference, while a flip indicates that removing the feature group changed the final answer. Note that these results identify input features associated with a model’s prediction; they do not establish causal relationships in the underlying biology.

What happens when this information is removed?

^{How does removing a feature group change the prediction? Choose a task and a feature group to see whether the model became more or less confident in the correct answer, and how often removing that information changed a correct prediction to an incorrect one.}

Both LFM2 models were sensitive to the biological information provided in the prompt. Ablating selected feature groups reduced confidence in the correct answer and, in some cases, led to incorrect predictions, with the magnitude of these effects varying across tasks.

A first step toward specialized models for longevity research

These results show that domain-specific training can adapt compact language models to selected structured-omics tasks without requiring frontier-scale resources. This is an initial step toward deploying language models for specialized tasks in longevity research.

Compact models also create a practical option for local deployment when patient-level data cannot be sent to an external API. Such use will still require task-specific validation and appropriate safeguards for the full system.

LongevityBench is intended as a resource for researchers developing and comparing their own models. Considerable work remains: the benchmark should expand to additional tasks and data types, Longevity-LLM performance must improve across the full task suite, and reliability, fairness, latency, and cost should be evaluated in realistic research settings.

Acknowledgements

We are grateful to Insilico Medicine for the collaboration that made this work possible.

Citation

If you use this work, please cite the manuscript:

@article{zhavoronkov2026longevitybench,
  title   = {An Open Benchmark and Language Models for AI in Aging Biology},
  author  = {Zhavoronkov, Alex and Naumov, Vladimir and Sidorenko, Denis and Aliper, Alex and Aladinskiy, Vladimir and Hasani, Ramin and Amini, Alexander and Nasto, Katerina and Reymond, Mathieu and Shayakhmetov, Rim and Miftakhutdinov, Zulfat and Gladyshev, Vadim N. and Galkin, Fedor},
  journal = {Cell},
  volume  = {189},
  pages   = {5980--5994},
  year    = {2026},
  doi     = {10.1016/j.cell.2026.08.026},
  url     = {https://www.cell.com/cell/fulltext/S0092-8674(26)00999-2},
}

References

1. [1]López-Otín, C., Blasco, M. A., Partridge, L., Serrano, M. & Kroemer, G. (2023). Hallmarks of aging: An expanding universe. Cell 186, 243–278. https://doi.org/10.1016/j.cell.2022.11.001
2. [2]Horvath, S. (2013). DNA methylation age of human tissues and cell types. Genome Biology 14, R115. https://doi.org/10.1186/gb-2013-14-10-r115
3. [3]Belsky, D. W. et al. (2022). DunedinPACE, a DNA methylation biomarker of the pace of aging. eLife 11, e73420. https://doi.org/10.7554/eLife.73420
4. [4]Lu, A. T. et al. (2022). DNA methylation GrimAge version 2. Aging 14, 9484–9549. https://doi.org/10.18632/aging.204434
5. [5]Cui, H. et al. (2024). scGPT: toward building a foundation model for single-cell multi-omics using generative AI. Nature Methods 21, 1470–1480. https://doi.org/10.1038/s41592-024-02201-0
6. [6]de Lima Camillo, L. P. et al. (2024). CpGPT: a foundation model for DNA methylation. bioRxiv. https://doi.org/10.1101/2024.10.24.619766
7. [7]Brixi, G. et al. (2026). Genome modelling and design across all domains of life with Evo 2. Nature 652, 1349–1361. https://doi.org/10.1038/s41586-026-10176-5
8. [8]Nguyen, V. K. et al. (2023). Harmonized US National Health and Nutrition Examination Survey 1988–2018 for high throughput exposome-health discovery. medRxiv. https://doi.org/10.1101/2023.02.06.23284573
9. [9]Paulose-Ram, R., Graber, J. E., Woodwell, D. & Ahluwalia, N. (2021). The National Health and Nutrition Examination Survey (NHANES), 2021–2022: Adapting data collection in a COVID-19 environment. American Journal of Public Health 111, 2149–2156. https://doi.org/10.2105/AJPH.2021.306517
10. [10]Edgar, R., Domrachev, M. & Lash, A. E. (2002). Gene Expression Omnibus: NCBI gene expression and hybridization array data repository. Nucleic Acids Research 30, 207–210. https://doi.org/10.1093/nar/30.1.207
11. [11]GTEx Consortium. (2013). The Genotype-Tissue Expression (GTEx) project. Nature Genetics 45, 580–585. https://doi.org/10.1038/ng.2653
12. [12]Chia, R. et al. (2025). A plasma proteomics-based candidate biomarker panel predictive of amyotrophic lateral sclerosis. Nature Medicine 31, 3440–3450. https://doi.org/10.1038/s41591-025-03890-6
13. [13]Gong, Q. et al. (2025). Multi-omic profiling reveals age-related immune dynamics in healthy adults. Nature 648, 696–706. https://doi.org/10.1038/s41586-025-09686-5
14. [14]Kirsher, D. Y. et al. (2025). Current landscape of plasma proteomics from technical innovations to biological insights and biomarker discovery. Communications Chemistry 8, 279. https://doi.org/10.1038/s42004-025-01665-1
15. [15]Rafikova, E. et al. (2024). Open Genes—a new comprehensive database of human genes associated with aging and longevity. Nucleic Acids Research 52, D950–D962. https://doi.org/10.1093/nar/gkad712
16. [16]de Magalhães, J. P. et al. (2024). Human Ageing Genomic Resources: updates on key databases in ageing research. Nucleic Acids Research 52, D900–D908. https://doi.org/10.1093/nar/gkad927
17. [17]Avelar, R. A. et al. (2020). A multidimensional systems biology analysis of cellular senescence in aging and disease. Genome Biology 21, 91. https://doi.org/10.1186/s13059-020-01990-9
18. [18]Bunu, G. et al. (2020). SynergyAge, a curated database for synergistic and antagonistic interactions of longevity-associated genes. Scientific Data 7, 366. https://doi.org/10.1038/s41597-020-00710-z