Closing The Loop Between AI And Experimental Biology

A conversation with Rosie Rodriguez, Ph.D., President of Growth, Relation

Relation recently announced its launch of MORGAN, a multimodal foundation model for biology, as well as its newly expanded collaboration with GSK to generate large-scale human cellular perturbation data sets and deploy foundation models to accelerate target discovery.

Relation combines one of the world's largest human tissue data sets with automated experimental biology to train multimodal foundation models capable of predicting how human cells respond to genetic and drug perturbations.

In this Q&A, Life Science Connect’s Morgan Kohler caught up with Rosie Rodriguez, Ph.D., president of growth at Relation, to discuss AI’s place in accelerating discovery.

Why does the launch of MORGAN represent an important step toward foundation models that can learn and predict human biology?

MORGAN represents an important step because biological foundation models need to move beyond simply recognizing patterns in existing biological data toward understanding how biological systems respond when they are perturbed. By training on large-scale, time-resolved cellular perturbation data, models can begin to learn relationships between interventions and downstream biological responses. That is critical if we want AI not only to describe human biology but ultimately to predict how it may respond to a therapeutic intervention. The broader opportunity is to build models with a more causal and predictive understanding of disease biology that can inform drug discovery and translational research.

How are large-scale cellular perturbation data sets advancing the next generation of AI models for target discovery and translational research?

Scale is important, but the type of data matters just as much. Cellular perturbation experiments allow us to systematically intervene in biology and observe what happens across multiple molecular dimensions and over time. When you generate these experiments at sufficient scale, you create data sets that can help models learn which biological relationships may be causal rather than simply correlated.

For target discovery, that can be particularly powerful. Instead of identifying a target because it is associated with disease, we can ask what happens when we perturb that target, which pathways are affected, how different cell states respond, and whether those effects are consistent with the therapeutic hypothesis. That has the potential to improve both target identification and validation.

For example, our collaboration with Novartis in atopic diseases demonstrates how we are applying this approach in R&D. Relation’s lab-in-the-loop platform combines large-scale data generation from human tissue, single-cell omics, cellular perturbation experiments, and AI models of gene and protein networks to identify and validate potential therapeutic targets. The collaboration is focused on using these capabilities to discover and advance new targets for atopic diseases. For us, it’s a good example of how generating the right experimental data and feeding those results back into our models can help move from predictions about human biology toward therapeutic hypotheses that can ultimately be tested and translated into new medicines.

Why is pharma doubling down on AI partnerships now, and what are companies like GSK looking for in next-generation AI platforms?

Pharma has spent several years evaluating where AI can meaningfully change drug discovery, and we're now moving from experimentation toward deployment against real R&D problems. The industry increasingly recognizes that the greatest value isn't necessarily going to come from algorithms alone. It's going to come from combining sophisticated models with differentiated biological data and experimental capabilities that allow predictions to be tested and improved.

Companies like GSK are looking at how AI and large-scale biological data sets can deepen our understanding of disease, identify new therapeutic opportunities, and ultimately increase the probability of success in drug development. The most valuable partnerships bring together complementary capabilities: pharma's deep drug development expertise with technology companies' ability to generate new types of data and build models capable of learning from them.

And I think that last point is important. There is something structurally powerful about smaller technology-native companies: our speed of learning can be incredibly fast. At Relation, we think a lot about excellence in execution, not simply speed for speed's sake but being both efficient and effective in how we make decisions, run experiments, learn, and iterate. The competitive advantage is how quickly you can turn an insight into an experiment, an experiment into data, and that data into the next decision. By tightly integrating biology, computation, and experimentation, we can continuously shorten that cycle.

When you combine that speed of learning and execution with the scale, expertise, and development capabilities of a company like GSK, you create something neither organization could achieve as effectively alone. The best partnerships aren't about outsourcing a capability; they're about bringing together genuinely complementary strengths to move faster, improve the quality of the science, and ultimately make better decisions. That's why I think we're seeing the best AI partnerships evolve from technology transactions into genuinely integrated scientific collaborations.

Why is high-quality human biology data becoming the key differentiator in building predictive AI for drug discovery?

AI models are fundamentally constrained by the information they're trained on. Drug discovery has historically relied heavily on animal models, immortalized cell lines, and relatively static biological data sets. Those approaches remain useful, but they don't always capture the complexity of human disease.

If the goal is to predict what will happen in patients, we believe models need much richer information grounded in human biology. That means combining high-quality human-derived data with perturbational experiments that show how biological systems respond to intervention. As models become increasingly sophisticated, access to differentiated, biologically meaningful training data will become one of the biggest determinants of their usefulness.

How can integrating experimental biology with machine learning improve confidence in therapeutic targets before they reach the clinic?

One of the fundamental challenges in drug development is that a target can look compelling computationally but fail when tested experimentally — or ultimately in patients. Integrating machine learning directly with experimental biology creates a feedback loop.

Models can generate hypotheses about targets and biological mechanisms, experiments can test those hypotheses, and the resulting data can then inform subsequent modeling and experimentation. Rather than treating computational prediction and laboratory validation as separate stages, you can continuously iterate between them.

But making that work isn't only a technology challenge; it's also an organizational one. You need drug developers, biologists, computational scientists, and technologists working as one integrated team, rather than handing work from one discipline to another. That closeness allows you to interpret results, make decisions, and determine the next experiment much faster.

The goal is to build a much stronger body of evidence around a target before significant resources are committed to clinical development. It won't eliminate biological risk, but it can help us make better-informed decisions earlier.

Where is AI-native drug discovery headed over the next five years?

I think we'll see the field move from AI primarily helping scientists analyze existing data sets toward systems that can increasingly predict biology and help determine which experiments should be performed next.

We'll also see much tighter integration between machine learning and experimental science. Automated laboratories, large-scale perturbation experiments, multimodal biological data, and increasingly capable foundation models can operate as a connected system in which predictions are experimentally tested and the resulting data continuously improve the models.

Ultimately, success won't be measured by how sophisticated an AI model is. It will be measured by whether we can translate these advances into better therapeutic hypotheses, better targets, and, ultimately, better medicines for patients. And that requires more than technology. It requires truly interdisciplinary companies where biologists, drug developers, computational scientists, and technologists work together around the same problem with a shared sense of urgency. Because, in the end, patients are waiting.

About The Expert

Rosie Rodriguez, Ph.D., is Relation’s president of growth. She brings over 18 years of successful research and development experience in pharmaceuticals and life sciences. She is a senior executive accomplished in clinical development, alliance and portfolio management, and business strategy. Rosie holds a Ph.D. from The Institute of Cancer Research (ICR CRUK).