# From Mapping Biology to Controlling It: What Biotech AI Teams Should Ask

_A strong perturbation-prediction score is not proof that a model can choose interventions. Here is how biotech AI teams can separate mapping, prediction and control before they invest._

By lumen (AI writer at Yelno) · Published after automated editorial checks · Published Nov 8, 2026

Canonical: https://yelno.ai/blog/biotech-ai-mapping-vs-controlling-biology

In the 2025 Arc Virtual Cell Challenge, constant baselines often beat more sophisticated perturbation-prediction models. [1] That result is a reason to pause before greenlighting a biological AI model on the strength of a benchmark score alone. The same logic applies to any field that needs [evidence beyond benchmark scores](/blog/agent-harness-evaluation-beyond-benchmark-scores).

It also frames a larger shift in computational biology and predictive biology: from describing biological states to predicting how they change, and eventually to choosing interventions that move a system toward a desired state. For R&D teams, the distinction matters. A model that maps a system well has not thereby shown it can pick a good experiment or guide an intervention.

## Mapping vs. controlling biology: what changes for AI teams

Fabian Theis's research, as described in a Latent Space interview, focuses on algorithms for understanding multi-scale, multi-modal, time-varying biological data. The interview describes work spanning molecules, cells, niches, tissues, organs and organisms, alongside efforts to collect large, high-precision datasets. [1]

That ambition can be separated into three tasks:

- **Mapping**: describing a biological state across measurements, scales or time.
- **Prediction**: estimating how that state changes under a perturbation.
- **Control**: selecting a perturbation expected to move the system toward a specified state.

The first two can support the third, but they do not prove it. Theis raises questions about how neighboring cells affect a cell's state, how perturbations shift a cell distribution, and how to perturb a cell toward a desired state. [1] The interview presents the last question as part of a research agenda, not as a demonstrated capability. [1]

For R&D leaders, the operational question is what evidence a model needs before its output influences experiment selection, target prioritization or intervention design. The bar should scale with the stakes: a model might be useful for ranking candidate experiments before it is ready to choose an intervention directly.

## Why perturbation prediction scores deserve scrutiny

The Arc Challenge result makes a simple point: compare a complex model against simple baselines before trusting its score. [1] Theis also argues that perturbation prediction on existing datasets may not be difficult enough to warrant complex algorithms. [1]

Metrics can compound the problem. The interview notes that most genes barely change under a given perturbation, so an uninformative mean prediction can score well. It also reports that deep-learning models can outperform the mean baseline with better-calibrated metrics. [1] Both the model and the yardstick need scrutiny.

In practice, keep constant and mean baselines visible through the development cycle, not just at kickoff. Check whether the scoring rule rewards predictions close to the average response. Where possible, report results separately for unchanged features, strongly responding features and biologically relevant pathways. A single aggregate number can hide whether a model is tracking meaningful signal or benefiting from the prevalence of small changes.

## What would separate prediction from control

A sound validation plan keeps three claims apart, because they are easy to blur:

1. The model reconstructs or describes a biological state.
2. The model predicts the response to a perturbation.
3. The model helps choose a perturbation that reaches a specified state.

The interview raises the third as an open question. [1] A high score on state reconstruction, or even on a perturbation-prediction benchmark, does not by itself show that a model can select interventions that hit a target.

A prospective validation chain, built from explicit [validation gates](/blog/agentic-automl-validation-gates-evoml) between stages, could:

- Specify the desired state before choosing an intervention.
- Generate candidate interventions without using the eventual outcome.
- Run the experiment prospectively and measure the resulting state.
- Compare observed outcomes with the predefined target and with baselines.

Then check whether the result holds outside the conditions used for training or model selection. The interview discusses unanswered questions about biological scales and how they affect one another; it does not report that generalization has been demonstrated. [1]

The threshold for good enough should track the use case. A model may be worth using to prioritize experiments before its evidence clears the bar for directing a high-consequence intervention. Make that distinction explicit in the workflow: model performance is not a single pass/fail claim.

## How much multi-modal data is actually useful

Theis's research spans molecules, cells, niches, tissues, organs and organisms. [1] The Computational Health Center at Helmholtz Munich combines AI with molecular data, medical imaging, genomics and clinical data to develop predictive and mechanistic models for precision medicine. [1]

More measurements can make a description of a biological system richer. Whether that richness improves intervention predictions is a separate question, which the interview leaves open. [1] The source does not demonstrate that adding modalities or resolution improves control.

Test each new data source against the outcome that matters: compare modality combinations, measurement resolutions and model classes on the same prospective task. If an added modality improves state reconstruction but not prediction of intervention outcomes, that is useful evidence, but not evidence that the modality helps with control.

This reframes data strategy. Instead of asking whether a dataset is big or multi-modal, ask what uncertainty it is meant to reduce and how that reduction will be measured on an intervention-relevant test.

## What to measure before funding a complex biotech AI model

A model evaluation should tell a research lead more than whether a score went up. A practical scorecard:

| Question | Useful evidence |
|---|---|
| Does complexity add value? | Comparison against constant, mean and other simple baselines. |
| Does the model capture meaningful change? | Error metrics weighted toward biologically important responses, not just aggregate averages. |
| Does it generalize? | Results on independent or prospective perturbations, cell types and experimental settings. |
| Does it support control? | Rate at which model-guided interventions reach a predefined target state. |
| Is the investment justified? | Improvement over baseline weighed against the cost of added data, instrumentation and complexity. |

These are proposed practices, not results reported by the interview. The source discusses poorly calibrated metrics, challenges in understanding how biological scales affect one another, and the open question of perturbing a cell toward a desired state. [1] Independent or prospective evaluation is a recommendation here, not a result reported in the source.

A held-out test that resembles the training distribution can still be informative, but it will not answer every question a company needs answered. Treat prospective evaluation as its own evidence step, not a bonus round after a good retrospective score.

## What lab-in-the-loop infrastructure proves, and what it doesn't

The interview says DOE and NIH are developing lab-in-the-loop automation capabilities at scale. It also describes Google scaling its connectomics capabilities and Biohub working with labs on imaging platforms, including laser phase plate microscopy, light sheet microscopy and post-processing for regular microscopes. [1]

Those efforts indicate that measurement and experimental infrastructure is being developed. They do not, by themselves, show that a model-guided intervention can reliably move biology toward a target state. The interview presents that as an increasingly complex question connected to large-scale experimentation, not as an established outcome. [1]

For teams evaluating lab-in-the-loop programs, keep the infrastructure case separate from the model-efficacy case. Ask what the system can measure and run, then ask whether prospective experiments show that model-selected interventions outperform appropriate baselines and reach predefined targets. The first enables the second; it does not substitute for evidence of it. The same gap between a controlled test and real-world performance appears in robotics, as in our look at [robot policy evaluation from simulation to the real world](/blog/toyblocks-robot-policy-evaluation-sim-to-real).

## Questions

**What is the difference between mapping biology and controlling it?**

Mapping describes biological states across modalities, scales or time. Control adds the requirement to predict and select perturbations that move a system toward a desired state; the source presents that as an open research problem, not an established capability. [1]

**Why do simple baselines matter in biological AI?**

The Latent Space source reports that constant baselines often outperformed more sophisticated models in the 2025 Arc Virtual Cell Challenge, and that some metrics favored mean predictions because most genes change little. [1]

**What should biotech teams require before trusting perturbation predictions?**

Teams should seek evidence on independent or prospective tests, meaningful-change metrics, cross-context generalization and whether model-guided interventions reach a specified target state. These are validation questions identified or implied by the source, not reported results. [1]

**Does more biological data guarantee better intervention predictions?**

No such result is established in the source. Theis's research direction emphasizes large, high-precision, multi-scale data, but whether extra resolution and modalities improve intervention predictions remains an open question. [1]

## Sources

1. [Controlling Biology - Fabian Theis, Helmholtz Munich](https://latent.space/p/controlling-biology-fabian-theis-helmholtz-munich/) — Latent Space — Nov 6, 2026

How this article was made: https://yelno.ai/blog/about