From Mapping to Controlling: Fabian Theis on the Next Stop for Biology

Systems Biology Is Shifting Gears: From "Drawing Maps" to "Issuing Commands"

Over the past decade or so, the core task of systems biology has been to see life in finer detail, across broader scope, and through more modalities. Single-cell sequencing, spatial transcriptomics, multi-omics integration—these tools have enabled researchers to draw biology's "map" at ever higher resolution. Now Fabian Theis of Helmholtz Munich is pushing this curve one step further: from mapping biology, toward predicting biology, and ultimately toward controlling biology.

In a recent talk, he laid out this roadmap systematically. The core judgment is clear: once systems biology's data accumulation reaches a certain density, the real value lies not in drawing yet another finer map, but in using these maps to predict system behavior and, further, to apply interventions and control outcomes.

The significance of this shift is that the objective function of biological research has changed. The endpoint of traditional omics research is a descriptive paper and a public dataset; the endpoint of "control"-oriented research is an actionable intervention strategy—which gene to modify, at what time point, at what dosage. This imposes entirely different requirements on both computational methods and experimental design.

Three Steps: Mapping, Predicting, Controlling

The framework Theis proposes can be understood as three progressive stages, each building on the data foundation of the previous one.

Mapping: Pushing Resolution, Breadth, and Modality Higher Simultaneously

This is the main battleground of current systems biology. Higher resolution means seeing heterogeneity at the single-cell level; greater breadth means covering more samples, more tissues, more species; more modalities means layering information from transcriptome, epigenome, proteome, and other levels. One of Theis's central efforts is to advance all three axes at once.

The key challenge here is not generating data but integrating it. How to align data from different modalities, batches, and platforms into a shared representation space is a long-standing hard problem in computational biology.

Multi-omics integration refers to computationally aligning and jointly modeling multiple molecular layers—such as transcriptome (RNA expression), epigenome (DNA methylation, chromatin accessibility), and proteome—generated from the same biological sample. Common strategies include matrix concatenation based on shared features (e.g., genes), joint embedding via latent variables (e.g., multimodal VAEs), and, more recently, cross-modal alignment based on contrastive learning. The core goal is to allow information from different modalities to complement each other within a shared representation space, thereby characterizing cell states more comprehensively. The main obstacles to integration are scale differences between modalities, batch effects, and missing modalities—which is why "batch removal" and "biological signal preservation" are the two metrics commonly used when evaluating integration methods.

Predicting: Teaching Models to Learn "What Happens Next"

Once the atlas has accumulated to sufficient scale, the next question naturally arises: can we predict? Given a cell's current state, where will it go next? Given a perturbation, how will the system respond? This is essentially a shift from static description to dynamic modeling.

In his talk, Theis emphasized that predictive capability is the bridge connecting "seeing" and "doing." Without reliable predictions, any intervention can only proceed by trial and error.

Controlling: From Understanding Systems to Manipulating Them

This is the ultimate goal and the hardest step. Controlling biology means not just predicting what consequences a given perturbation will bring, but the reverse: given a desired outcome, solve for what perturbation should be applied. This is an inverse problem, far more difficult than forward prediction.

From an engineering perspective, this amounts to transforming biology from an "observational science" into a "design science." But the nonlinearity, feedback loops, and noise characteristics of biological systems make direct application of classical control theory very difficult.

Inverse problems in mathematics refer to problems of inferring causes or inputs from observed outcomes. In controlling biology, the forward problem is "given a perturbation, predict the system response," while the inverse problem is "given a desired response, solve for the required perturbation." Inverse problems are typically ill-posed: there may be multiple perturbation combinations achieving the same effect, there may be no solution, and solutions are highly sensitive to noise. Classical control theory handles such problems through feedback mechanisms and system identification, but biological systems exhibit strong nonlinearity, time-varying behavior, and unobserved confounders, so directly applying linear control frameworks often fails. As a result, practical approaches often adopt iterative strategies such as model predictive control and Bayesian optimization, gradually approaching the target within a closed loop of "predictive model + experimental feedback."

What This Means for AI Methods

Theis's roadmap implies a methodological judgment: the ability to predict and control depends on whether models can learn transferable representations.

Single-cell data is notoriously high-dimensional, sparse, and noisy. Making predictions on such data, models easily overfit to a specific dataset and fail when transferred to another tissue or platform. Therefore, one of the long-term directions Theis's team has invested in is building general representations that span datasets and modalities—which is also the backdrop for the recent rise of biological foundation model approaches.

But we need to view this soberly: biological foundation models are still clearly far from "reliable prediction," let alone "control." Data scale, annotation quality, and causal inference capability are all bottlenecks. Theis's statement is more of a directional declaration than a checklist of already-delivered results.

Why "Control" Is a Harder Proposition

From prediction to control, there lies a causal gap in between.

Correlation models can tell you "the expression of these two genes always changes in sync," but cannot tell you "if you knock out one, what happens to the other." Control requires exactly the latter. The vast majority of current mainstream multi-omics data is observational; perturbation data is far smaller in scale.

This brings a practical constraint: the pace of development in controlling biology may not depend on advances in computational methods, but on the rate at which interventional experimental data accumulates. For Theis's roadmap to be realized, closer collaboration between the experimental and computational sides is needed—which is also why he repeatedly emphasizes interdisciplinary cooperation.

Another less-discussed difficulty is evaluation. Prediction tasks have clear evaluation metrics, but "whether a control strategy is good" often needs to be validated in living systems, with long feedback cycles and high costs. This makes the iteration speed of methods far slower than in purely computational fields.

Perturbation data refers to omics readouts collected after actively intervening in a system through experimental means. Typical forms include CRISPR knockout/knockdown/activation screens, drug treatments, and small-molecule perturbations, paired with single-cell transcriptomic readouts (e.g., Perturb-seq). Compared with observational data, perturbation data can reveal the causal relationship between "intervention and response" and is a key resource for training predictive models and solving control strategies. Its limitations are: throughput far lower than observational data, high cost of building separate libraries for each perturbation, exponential growth of the combinatorial space for multiple perturbations, and limited transferability across cell types and experimental platforms. These factors together determine that the pace of development in controlling biology is largely constrained by the rate of accumulation and the degree of standardization of perturbation data.

Conclusion

The mapping → predicting → controlling trilogy proposed by Theis is less a technical roadmap than a resetting of goals. It pushes systems biology from "accumulating knowledge" toward "generating intervention capability," and gives a clearer positioning of AI's role in the life sciences: not to replace experiments, but to make experiments more directed.

The speed at which this path is realized depends on how tightly data, methods, and experimental validation are coupled. Looking at current progress, the first stage is still accelerating, the second stage has just begun, and the third stage is more of a long-term goal worth taking seriously.

One-line Summary

Theis proposes a three-stage roadmap for systems biology: from mapping to predicting to controlling.

Paragraph Summary

The article outlines Fabian Theis's three-stage roadmap for systems biology: from the current mapping stage, dominated by high-resolution, multi-modal mapping of biological atlases, to a predicting stage that forecasts dynamic system responses, and ultimately to a controlling stage that works backward from a desired outcome to derive an intervention strategy. The author points out that this shift is essentially a change in the research objective function—from producing descriptive papers and datasets to producing actionable intervention strategies. The article further analyzes the role and limitations of AI methods in this process: biological foundation models are not yet sufficient to support reliable prediction, a causal gap exists between prediction and control, observational data cannot replace perturbation experiments, and evaluation of control strategies relies on long-cycle, high-cost in vivo validation.