In-silico Evaluation of Aging-Related Interventions Using Omics Data and Predictive Modeling

1. Introduction

Aging research increasingly focuses on identifying and testing interventions — pharmacological, genetic, dietary or behavioral — that might slow, stop or reverse aging processes and improve health in later life (Lopez-Otin et al., 2013; Lopez-Otin and Kroemer, 2021). Yet, demonstrating long-term health and lifespan benefits and low toxicity in humans is inherently challenging, and clinical studies of long-term health would require to follow participants for decades. Surrogate biomarkers, including blood values, gene expression and methylation data, have risen to prominence, enabling the training of “phenotypic”, transcriptomic and epigenetic aging clocks to predict intervention effects (Fuellen et al., 2019; Hartmann et al., 2021; Hartmann et al., 2023; Moqri et al., 2023; Moqri et al., 2024). However, almost all of these surrogate-based predictors are established exclusively on observational data, not on interventional outcome data. The resulting domain shift can lead to misguided predictions of intervention efficacy, as already noted for reprogramming interventions (Kriukov et al., 2024). Only very recently, (Belikov et al., 2024) published intervention-based predictors to identify compounds that may extend the lifespan of mice (see Box 1).

In this Review, we examine the rapidly evolving field of in-silico intervention analytics, with a focus on learning from intervention data rather than observational data whenever possible, see Figure 1. We also discuss why safety/toxicity considerations are critical in preventive interventions. In depth, we highlight intervention-based benchmarks that enable the comparison of prediction methodologies, see Table 1. We then delineate two steps important for robust predictions — feature selection/extraction and predictor learning — and survey a variety of established and emerging approaches, see Figure 2. Feature extraction methods such as principal component analysis (PCA) reduce complexity and can highlight biologically interpretable signatures; predictive modeling approaches range from linear regressors and classifiers to non-linear machine learning methods such as random forests, gradient-boosted trees, and neural networks (Eckhart et al., 2024; Pantazis et al., 2020; Piccolo et al., 2022). Finally, we discuss the emerging use of generative AI and large language models (LLMs) in this context. Although preliminary and often challenging, LLMs could eventually provide excellent assistance in extracting features, building models, and generating mechanistic hypotheses.

Box 1. Predicting Lifespan-Extending Compounds in Mice.

Belikov and colleagues (Belikov et al., 2024) used machine learning to identify compounds that can extend the lifespan of mice, leveraging murine lifespan data sourced from the DrugAge database. The authors used three different kinds of features: (i) direct protein target annotations, capturing gene ontology and pathway descriptors of a compound’s protein interactors; (ii) gene expression signatures from the LINCS repository, using consolidated expression values for each compound; and (iii) PubChem-based chemical substructure representations. Random Forest models were trained and tested via cross-validation. Both area under the ROC curve (AUC) and geometric mean (GMean) metrics were reported; the latter was usually more appropriate because of imbalances in the dataset. The best performance arose from the target-based annotations. By contrast, LINCS gene expression data and chemical substructure representations underperformed, perhaps because no feature extraction/selection was done. Finally, the study used selected top models to identify potentially novel lifespan-extending compounds from DrugBank.

2. The Challenge of Predicting Intervention Outcomes

Predictive modeling of intervention outcomes (e.g., lifespan, health, toxicity) is central to aging research. Yet, most surrogate-based predictors, known as “aging clocks”, are derived from observational data, such as blood cell counts or other laboratory values, transcriptomics or methylation data, from individuals of varying age (Fuellen et al., 2019; Hartmann et al., 2021; Hartmann et al., 2023; Moqri et al., 2023; Moqri et al., 2024). Instead, training predictors directly on intervention data, when interventions with known outcomes and corresponding omics profiles are available, should provide more reliable inference. Why is that so? For a detailed description, let us consider the specific example of training on transcriptomic (gene expression) data after compound intervention, to predict health and lifespan outcomes, as compared to the standard use of aging clocks.

In the conceptual example of Figure 1 left, the decision boundary of a predictor is visualized, which was learned to decide between interventions that extend lifespan (+) and interventions that do not (–), where the feature space is defined by the similarity of the gene expression effects of the interventions. In Figure 1 right, predictor training was instead based on samples of young (+) versus old (–) tissues or cells, again in gene expression similarity space. If the predictors are then asked to classify a new intervention (marked by ?), the predictor on the right suffers from the domain shift from observational to intervention data. Everything else being equal, that is, assuming the same input data quality, etc., the domain shift is expected to trigger inferior results, because the similarity of interventions can only be directly measured in the space of intervention effects. This domain shift was very recently described for the special case of using aging clocks to assess reprogramming interventions (Kriukov et al., 2024), including ample empirical evidence that the use of observational data for the interpretation of intervention data may be problematic. Moreover, unlike cancer drug development, this difference is not simply about the control imposed by an interventional trial (optimally a randomised controlled clinical trial), to minimise confounding factors and to better establish causality. Judging longevity interventions by standard aging clocks is a unique situation in that the clocks themselves rely solely on observational data; the only “intervention” is the passage of time. Of note, moving from the classification task considered in Figure 1 to the analogous regression task, the scenario on the right corresponds to the standard way of using observational data for learning aging clocks that are then used to assess intervention effects.

Use of intervention training data is conceptually straightforward as visualized in Figure 1, but practically challenging: high-quality interventional datasets with paired phenotypic outcome data are scarce. Yet, resources are improving. For long-term health, studies that associate lifespan outcomes in model organisms with intervention-induced omics changes have begun to appear (Belikov et al., 2024), see Box 1. It thus becomes easier to train predictors directly on intervention data, producing models that are expected to generalize better to new interventions and reduce the uncertainty arising from the conceptual domain shift from observational to interventional data. For toxicity, where the use of intervention data is standard, these are now consolidated in repositories like TOXRIC, which provide large-scale transcriptomic data and well-defined toxicity outcomes (Wu et al., 2023), see Box 2. Toxicology is not just a role model for predictor learning, though; it is an important and sometimes overlooked aspect of assessing any aging-related intervention.

Box 2. The TOXRIC Database.

TOXRIC (TOxicology Resources for Intelligent Computation) (Wu et al., 2023) is a large-scale database aimed at supporting the development and benchmarking of toxicity prediction models. At time of publication, it contained more than 113,000 compounds, spanning 13 toxicity categories (e.g., acute toxicity, ecotoxicity, hepatotoxicity, endocrine disruption) and 1,474 in vivo or in vitro endpoints.  A key strength of TOXRIC is its “ML-ready” focus: the database provides curated, standardized toxicity labels and up to 39 feature types (e.g., molecular fingerprints, transcriptomic profiles, target annotations), so that these can be used directly as input (features) or output (labels) for machine learning. For instance, the database offers transcriptomic readouts from LINCS, Open TG-GATEs, and DrugMatrix, providing high-dimensional gene expression information after compound exposure. Structural descriptors and target protein annotations are also included, broadening the scope of potential modeling strategies. Benchmarking plays a central role in TOXRIC. Multiple classification and regression tasks are established (e.g., predicting toxic vs. non-toxic status or estimating LD50), and performance metrics are reported for four “baseline” algorithms frequently used in toxicity modeling—eXtreme Gradient Boosting, Random Forest, Support Vector Machine, and Deep Neural Network. Researchers can quickly compare how different feature types or model classes perform for each endpoint, and then download the corresponding data subsets for further experiments; the website also offers a range of results, visualized appropriately. In the context of in-silico intervention analytics, TOXRIC’s standardized datasets and benchmark results offer a comprehensive source of toxicity information that can be leveraged to train or validate predictive models focused on toxic effects.

3. The Importance of Safety and Toxicity Assessments

In the context of promoting aging-related interventions, especially in healthy individuals, safety is paramount. Unlike interventions in severely ill patients where some toxicity might be acceptable if the therapeutic benefit is high, preventive aging-related interventions must meet a much higher safety threshold. Predictive modeling efforts must therefore emphasize toxicity and side-effect prediction (Janssens and Houtkooper, 2020; Uner et al., 2023; Wang et al., 2016; Wu et al., 2023). Moreover, safety/toxicity considerations can specifically influence the selection of the features on which to focus in the intervention analysis pipeline. For example, when using gene expression data to establish predictors, features may be specifically selected or extracted with toxicity in mind, based on genes known to be involved in toxicity-related processes and pathways (Saarimaki et al., 2023a; Saarimaki et al., 2023b), see also below.

4. Establishing Benchmarks for Model Comparison

Rigorous benchmarking is essential for progress. Without standardized benchmarks, it is difficult to compare methods or assess generalizability. Benchmark datasets should represent various intervention types (e.g., drug treatments, dietary changes) and outcomes (e.g., lifespan, toxicity), and include both successful and neutral or harmful interventions. Table 1 outlines our suggestions for such intervention-based benchmarks. For toxicity, TOXRIC provides large-scale transcriptomic data linked to various toxicity endpoints (Wu et al., 2023), see also Box 2. For lifespan extension, data from DrugAge and the Interventions Testing Program (ITP) (Nadon et al., 2008) have recently been curated and connected to LINCS gene expression profiles (Belikov et al., 2024), see also Box 1. Side-effect data (Kuhn et al., 2016) were also connected to LINCS (Uner et al., 2023; Wang et al., 2016). Integrative datasets, featuring aging-related interventions with functional outcomes and gene expression data, are becoming available (Tyshkovskiy et al., 2024), also, for example, with a focus on partial reprogramming (Browder et al., 2022; Hishida et al., 2022; Sarkar et al., 2020). Additional benchmarks can be constructed for senotherapeutics (Smer-Barreto et al., 2023), nutritional interventions (Ford et al., 2023), and for in-vivo rat toxicity (Gwinn et al., 2020). Using these benchmarks, researchers can systematically compare feature extraction and predictor learning pipelines and monitor performance improvements.

5. Feature Selection/Extraction and Predictor Learning

A pipeline for the interpretable machine learning of intervention effects is described in Figure 2, consolidating existing approaches already described in (Wu et al., 2023) and (Belikov et al., 2024), and taking inspiration from machine learning of cancer drug sensitivity and cancer type (Eckhart et al., 2024; Pantazis et al., 2020; Piccolo et al., 2022; Smith et al., 2020), in a single unified scheme. In the next sections, the various steps of the pipeline will be discussed in more detail. In our case, the input is high-dimensional high-throughput molecular data, which may, e.g., be the gene expression (transcriptomics), protein abundance (proteomics) or methylation data (epigenomics) measured in control versus intervention samples. These data are frequently undergoing feature processing, by selection (keeping features as they are) or by extraction (calculating new features based on formulas). The consequent reduction of the feature space can avoid overfitting (see below). Training, testing and validation of predictors is then performed based on these features, using the labels associated with the samples, such as lifespan extension or toxicity for classification. Predictors can also employ regression, learning numerical labels. Training/testing refers to the (automated) machine learning of predictors (see below), and validation to the use of the learned predictors on new but similar data, testing their generalizability. (In the literature, the term “testing” sometimes refers to what we call “validation”, and vice versa.) Method choice can strongly influence the ability to assign biological/biomedical meaning to the features underlying the prediction results, enabling various grades of interpretability, which may be intrinsic to the method, but can also be offered by post-hoc analyses, see Figure 2.

6. Feature Selection/Extraction

Omics data, particularly transcriptomics, often measure thousands of genes. Directly applying machine learning on these high-dimensional, noisy data can lead to overfitting (Eckhart et al., 2024; Smith et al., 2020). Feature selection/extraction simplifies the feature space, generating a smaller set of features.

Linear Methods: MRMR, PCA and Contrastive PCA. A wide variety of feature selection methods are available, exemplified here by Maximum Relevance Minimum Redundancy (MRMR), which is often highlighted for its strong empirical performance and its theoretical appeal, as it aims for features strongly dependent on the outcome but weakly dependent on one another, see (Eckhart et al., 2024). This principle can be implemented in a linear way based on correlation as well as in a non-linear way based on mutual information. A popular linear method for feature extraction is principal component analysis (PCA), which produces features as linear combinations of genes (or other molecular variables) that explain the greatest variance in the data (Ringner, 2008). On that basis, PCA can separate samples by intervention type or outcome and reveal underlying biological processes. Because PCA features are linear, they are often easier to interpret biologically. For instance, certain principal components may correspond to activation of proliferative pathways or stress responses. However, PCA may also capture confounders such as batch effects or differences in experimental conditions. Contrastive PCA (cPCA) is a variant that uses “background” or control datasets to highlight features specific to the intervention condition (Abid et al., 2018; Boileau et al., 2020; de Oliveira et al., 2024). This may be particularly valuable in aging-related intervention studies, where differences in age, species, or tissue can obscure the signal of interest. cPCA may help isolate the molecular signatures of the interventions beyond baseline variability (Iturria-Medina et al., 2022), but this conjecture lacks confirmation for aging-related interventions.

Linear Methods: Gene Set Approaches. Interpretable features can also be derived from annotated gene sets, including biological pathways or processes, and specifically hallmarks of aging, or adverse outcome pathways (Basili et al., 2022; Pun et al., 2022; Saarimaki et al., 2023a; Saarimaki et al., 2023b; Subramanian et al., 2005). Such gene sets may simply define the features to be selected. More frequently, by summarizing expression levels for sets of genes with shared biological meaning, highlighted by their significant up- or downregulation, one obtains “enrichment features” that reflect the activity of entire pathways or processes. Tools like gene set enrichment analysis (GSEA), single-sample GSEA, and gene set variation analysis (Hanzelmann et al., 2013; Subramanian et al., 2005) can therefore translate gene expression into biologically interpretable pathway or process enrichment scores, enabling more meaningful comparisons between interventions. Enrichment-based features can also reduce dimensionality and mitigate the risk of overfitting. These gene-set–based features can be visualized in low-dimensional embeddings and network maps, helping researchers identify clusters of interventions that share similar mechanistic signatures (Merico et al., 2010; Nguyen et al., 2021). Such visualization can guide the generation of mechanistic hypotheses and highlight potential safety concerns if certain pathways are associated with toxicity.

Non-linear Dimensionality Reduction. Non-linear methods for feature processing can be based, e.g., on Neural Networks (incl. Autoencoders) (Eckhart et al., 2024). Moreover, non-linear alternatives to PCA, such as UMAP or t-SNE can be used to represent samples in lower-dimensional spaces (Yang et al., 2021), and these representations can be used to define features. While powerful for visualizing complex relationships and subgroups, the results of these methods are harder to interpret biologically. Since interpretability is crucial in preventive health interventions, linear or gene set–based approaches may be preferable.

7. Predictor Learning

Once features are defined, the next step is to train and test predictive models. Such predictors then estimate intervention outcomes (e.g., lifespan extension, toxicity) from the extracted (or selected) features. However, feature processing is an optional step, and generally, predictors can be given all features as-is, even if these are high-dimensional, potentially at the risk of overfitting; some methods can also combine feature selection and predictor definition into a single formula.

Linear Predictors. Linear predictors, including logistic regression or elastic nets, can be combined with PCA-based or gene set–based features to produce interpretable models (Eckhart et al., 2024). The weights assigned to each feature can then highlight biologically meaningful signals and simplify the interpretation of predictions. However, purely linear models may not fully capture the non-linear relationships between gene expression changes and the outcome phenotype.

Non-linear Machine Learning Models. Non-linear models like random forests, gradient boosting (e.g., XGBoost), and neural networks can often but not always yield better predictive performance (Eckhart et al., 2024; Pantazis et al., 2020; Piccolo et al., 2022). In particular, gradient boosting methods have shown strong performance on toxicity benchmarks (Wu et al., 2023) Nevertheless, non-linear models tend to be less interpretable. Techniques like LIME or Shapley values can offer post-hoc explanations (Ortigossa et al., 2024). Random forests also provide variable importance scores that indicate which features are most influential, fostering interpretability. However, balancing accuracy with interpretability remains a challenge. Since safety and transparency are vital, especially for interventions in healthy individuals, simple but interpretable models may sometimes be preferred over complex black-box models.

Automated Machine Learning (AutoML) and Model Ensembles. Given the complexity of feature extraction and the multitude of available predictors, automated machine learning (AutoML) frameworks can expedite the search for optimal intervention analysis pipelines (Tsamardinos et al., 2022). These frameworks systematically test different feature extraction methods, prediction model types, and hyperparameters. Furthermore, ensemble strategies that combine multiple prediction models can yield more robust predictions (Campager et al., 2023). Incorporating uncertainty estimates as suggested by (Kriukov et al., 2024) can further increase the reliability and safety of intervention recommendations.

8. Early Experiences With Generative AI and LLMs

Large language models and related AI techniques hold promise for integrating disparate knowledge sources and generating new hypotheses (Joachimiak et al., 2024; Liu et al., 2024; Pickard et al., 2024; Simm et al., 2024; Tang et al., 2024; Wang et al., 2024; Xin et al., 2024; Zhou et al., 2024). They could assist with data preprocessing, identifying confounders, or proposing candidate gene sets. However, early attempts have shown limitations — such as difficulties in handling complex, messy biological datasets and improper statistical analyses (Joachimiak et al., 2024). In the near future, LLMs might help design improved intervention analysis pipelines, facilitate interpretability, or assist in evaluating interventions against curated benchmarks. At present, generative AI is best viewed as a coding assistant or brainstorming partner rather than a fully autonomous analyst. Ensuring correctness and interpretability remains a major challenge (Fuellen et al., 2024). As such, LLMs should be integrated cautiously, with human oversight and rigorous validation.

9. Perspectives and Future Directions

As the field advances, several key priorities emerge. First, expanding and refining high-quality intervention datasets, both in model organisms and in humans, will improve training and testing of predictive models. Notably, large-scale multi-intervention trials for mice are being run with increasing size and sophistication, e.g. the Robust Mouse Rejuvenation trials (Lewis and de Grey, 2024), and the same holds for human intervention trials, such as DO-Health (Kistler-Fischbacher et al., 2024), CALERIE (Ryan et al., 2024) and Cosmos (Vyas et al., 2024). Second, developing methods that integrate safety and efficacy predictions simultaneously is paramount; aging interventions must be safe if they are to be recommended for healthy individuals. Third, consistent benchmarking and publication of standardized protocols will foster reproducibility and progress. Fourth, the tension between model complexity and interpretability will continue to shape methodological choices. Achieving explainability without sacrificing accuracy, and providing robust uncertainty estimates to guard against overconfident predictions, will be crucial. Finally, incorporating generative AI tools in a transparent and carefully monitored manner may lead to innovative modeling strategies and new insights.

10. Conclusions

Validated and explainable predictive models for intervention outcomes would profoundly enhance our ability to identify strategies that extend lifespan and reduce disease and dysfunction. By prioritizing training on intervention data, consideration of safety/toxicity, and embracing robust benchmarks, we can better predict real-world outcomes, optimally with strong interpretability. As the available datasets and computational tools improve, in-silico evaluation can become an indispensable asset for aging-related intervention research, complementing experimental studies and ultimately informing clinical and public health decisions.