Multimodal Foundation Models in Drug Discovery: What Pharma Needs to Scale AI
Key takeaways
- Multimodal foundation models promise richer biological and patient representations than what is possible to attain from discrete biomarkers alone.
- The main scaling challenge is not only model architecture; it is AI-ready, FAIR and reusable biological data within capable infrastructure.
- Agentic tool architectures can be a pragmatic alternative to truly multimodal foundation models when fully aligned multimodal training datasets are unavailable.
- Knowledge graphs, ontologies and causal models can help move foundation models from pattern recognition toward traceable mechanistic reasoning.
- Pharma teams need a full stack: computable data, pragmatic architecture, domain adaptation, causal reasoning, validation and governance.
Why the conversation has moved from “if” to “how”
On April 29, 2026, the Pistoia Alliance UK Life Science Informatics Forum at The Hub, AstraZeneca, Cambridge, gathered industry leaders to discuss Multi-Modal Foundation Models.
The most useful part of the session was not a promise that larger models will solve drug discovery. It was the more practical question that surfaced across the presentations and discussion: what has to be true for multimodal foundation models to become widely adopted in real pharma R&D?
That question matters because the field is moving past the basic debate about whether AI belongs in drug discovery. The harder issue is how to make AI systems scientifically reliable, operationally reusable and useful in decisions that affect target selection, patient stratification, experimental design and portfolio value.
The answer is less about chasing the largest model and more about building the stack around the model: AI-ready data infrastructure, fit-for-purpose architectures, domain-specific validation and causal reasoning that can turn predictions into scientific evidence.
What are biological foundation models?
A foundation model is a large, pre-trained AI system that learns reusable representations from broad datasets and can then be adapted to many downstream tasks. In biology, these systems are analogous in ambition to the models used in natural language processing, but they are trained to represent biological data rather than only text.
A modality is a type of data. In life sciences, modalities can include single-cell data, genomics, proteomics, imaging, histology, fluorescent microscopy, assay readouts, clinical records, experimental results and time-course traces. A multimodal model is designed to learn from more than one such data type at the same time, ideally aligning them in a shared representation.
That shared representation is often described as an embedding space: a mathematical space in which biologically similar samples, patients, cells or perturbations should sit closer together. If the representation captures real biology rather than technical artifacts, it can support retrieval, patient similarity, transfer learning and prediction across tasks.
This is why biological foundation models are different from popular large language models (LLMs). LLMs are extremely useful for literature workflows, brainstorming, report generation and documentation. But drug discovery decisions also depend on new experimental, molecular, imaging and clinical data. LLMs can support downstream communication once scientific decisions have been made; they should not be confused with models designed to represent biology directly.
Lesson 1: Move beyond discrete biomarkers to multidimensional patient similarity
One of the strongest messages from the session was that patients should not be viewed only as collections of discrete biomarkers. A patient is a multidimensional biological and clinical state. Multimodal foundation models offer a way to compare those states more richly than a single positive or negative marker.
A useful analogy is a map. Individual observations are like villages, roads or coastlines. The value comes from learning the underlying geography: how features connect, which structures tend to appear together, and which states would be unusual or invalid. In a patient embedding space, that learned geography can help teams identify clusters, retrieve similar patients and reason about response profiles even when individual datasets are noisy.
For clinical development, this could support precision recruitment, cohort refinement and the search for similar patients across Phase I/II datasets where signal-to-noise ratios are often low.
This does not remove the need for internal data. In many cases, small but high-quality in-house datasets are still needed to fine-tune, calibrate or validate a foundation model for a specific biological system. The promise is not “no labeled data”; it is more effective reuse of pre-trained representations when labeled data are sparse.
Lesson 2: Choose between true multimodality and agentic toolkit architecture
The session also challenged an assumption that is easy to make: that the end goal must always be one fully integrated multimodal model. That may be the right architecture in some cases, but it is not automatically the most practical route for every pharma workflow.
At inference time, a base LLM can be understood as a model that receives input tokens and returns output tokens. An agentic system adds orchestration around that model: planning, state, memory, tool selection and calls to specialized systems. In document processing, for example, one approach is to encode text, tables and figures into a unified sequence. Another is to let a central reasoning model route tables, images, omics files or knowledge graphs to specialist tools.
This distinction matters because fully aligned multimodal datasets are expensive and difficult to assemble. If each data type has different structure, noise, metadata and validation requirements, a bespoke model or agentic toolkit may outperform a single generalist architecture for a specific task.
Single-cell foundation models such as scGPT illustrate the value of architecture choices that reflect the structure of the underlying biological modality. Without separate embedding layers for expression values, gene tokens and conditions it would not be possible to encode gene expression and train the model in recognizing gene co-expression patterns through masked prediction tasks 1.
The practical takeaway is to begin with the scientific decision and the available data, then choose the architecture. True multimodality is powerful when aligned data and validation paths exist. Agentic systems are attractive when the task requires reliable orchestration across heterogeneous tools, but the training data for full model-level integration are not yet mature.
Lesson 3: Treat foundation models as living research infrastructure
A foundation model should not be treated as a one-off asset that is trained, announced and left unchanged. In life sciences, these models will be created, fine-tuned, evaluated, replaced and recombined as data, assays and scientific questions evolve.
That makes the infrastructure around the model just as important as the model itself. Pharma teams need FAIR data – findable, accessible, interoperable and reusable – plus standardized metadata, versioned training pipelines, reusable repositories, evaluation harnesses and clear ownership of model lifecycle management.
This is where foundation models become part of a broader research ecosystem. If data and pipelines are reusable, teams can adapt models to specific biological systems without rebuilding everything from scratch. If they are not reusable, every model becomes another isolated pilot.
Lesson 4: Add the causal layer with knowledge graphs and ontologies
Foundation models are strong at prediction and pattern recognition. They can help answer questions such as whether a patient resembles prior responders or whether a perturbation resembles a known biological state. But drug discovery also depends on mechanistic questions: why a response may occur, which pathway could be involved and whether the proposed rationale is scientifically credible.
This is why knowledge graphs, ontologies and causal models matter. A knowledge graph connects entities such as genes, proteins, pathways, diseases, compounds and phenotypes. An ontology gives those entities consistent meaning. A causal graph represents assumptions about how variables influence one another. Together, these symbolic layers can ground AI outputs in traceable biological hypotheses.
The broader point is of crucial importance: pharma R&D does not only need faster predictions. It needs decision support that can explain the mechanistic rationale behind a prediction, expose uncertainty and make assumptions visible to scientists.
Practical path forward: build the stack, not just the model
The discussion that followed was held under the Chatham House Rule, encouraging sharing of useful takeaways and insights. The strongest theme was a warning against technology-led innovation: building AI capability first and searching for use cases later is the wrong sequence.
A more durable approach starts with the decision to be improved, then works backward through data, architecture, validation, deployment and governance. For multimodal foundation models in drug discovery, five layers are especially important.
Layer | What pharma needs to solve | Key question |
|---|---|---|
Data | Aligned multimodal datasets, FAIR metadata, batch-effect control, computable data and traceable lineage. | Can we trust what the model sees? |
Architecture | A fit-for-purpose choice between true multimodal models, bespoke models and agentic tool orchestration. | Does integration belong inside the model or across the workflow? |
Adaptation | Fine-tuning, calibration and validation against the biological system and decision context of interest. | Does the model generalize to our science? |
Causal layer | Knowledge graphs, ontologies, causal graphs and mechanistic reasoning. | Can the system explain why the prediction is plausible? |
Deployment | Human-in-the-loop workflows, audit trails, monitoring, governance and usability for scientists. | Will this actually change decisions? |
Table 1. Infrastructure layers for trustworthy foundation models in pharma R&D.
This stack also clarifies why scaling is hard. Data integration can be as difficult as regulatory compliance. Architecture choices create real trade-offs between power, interpretability and implementation burden. Explainability and mechanistic reasoning are not optional add-ons; they are requirements for scientific decision-making.
Conclusion: from model hype to AI-ready infrastructure
Multimodal foundation models hold real promise for drug discovery, but they are not a silver bullet. The organizations that scale them successfully will not simply be the ones with the largest models. They will be the ones that can connect models to reusable data infrastructure, scientifically meaningful validation and causal reasoning that supports confident decisions.
The strategic question for pharma is therefore not only whether to build or buy a foundation model. It is whether the organization is building the conditions that make foundation models useful: aligned data, pragmatic architecture, reusable workflows, traceable mechanistic reasoning and human-centered deployment.
Next step
Book a 30-minute AI-readiness review with Ardigen’s AI and data engineering team. We will map your current multimodal data, model architecture and decision-support workflows against the practical scaling stack described in this article, then identify the two or three highest-leverage moves for the next 12 months.
Frequently Asked Questions
What are multimodal foundation models in drug discovery?
Multimodal foundation models are AI systems trained to learn reusable representations across different biological and clinical data types, such as genomics, proteomics, imaging, histology, experimental readouts and clinical records. In drug discovery, they can help compare patients, samples, perturbations or mechanisms across richer data contexts than single-modality models.
How are biological foundation models different from LLMs?
LLMs primarily process text and are useful for literature, documentation and communication workflows. Biological foundation models are designed to represent non-text scientific data, including omics, images, assays and clinical measurements, which are often the data types needed for core R&D decisions.
Why is multimodal AI difficult to scale in pharma?
The main challenge is not only model design. Pharma teams need aligned multimodal datasets, consistent metadata, reusable pipelines, validation methods and governance structures that make model outputs trustworthy and reusable across projects.
Are agentic systems an alternative to full multimodal models?
Yes, in some cases. Agentic systems can route tasks to specialized tools for text, tables, images, omics data or knowledge graphs, which can be more practical when fully aligned multimodal training datasets are not available.
Why do knowledge graphs and ontologies matter for foundation models?
Foundation models are strong at pattern recognition, but drug discovery also requires mechanistic reasoning. Knowledge graphs and ontologies can ground model outputs in known biological relationships, making hypotheses more traceable and easier for scientists to evaluate.
What should pharma teams do before investing in multimodal foundation models?
They should define the scientific decision they want to improve, audit the readiness of their multimodal data and decide whether the task requires model-level multimodality, a bespoke model or an agentic workflow. They should also plan validation, monitoring and governance from the start.
What data infrastructure is needed for multimodal foundation models?
Teams need computable, well-described and reusable data with clear lineage, consistent identifiers and metadata, batch-effect controls, access governance and versioned pipelines. Without this foundation, a model may learn technical artifacts instead of biology.
Technical editing: Ardigen expert: Marek Kudła, PhD
References
[1] Cui, H., Wang, C., Maan, 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
[2] Ardigen. Why data quality matters in AI-powered drug discovery. https://ardigen.com/why-data-quality-matters-in-ai-powered-drug-discovery/
[3] Ardigen. Lab-in-the-loop: reclaiming the 50% of scientific time lost to data. https://ardigen.com/lab-in-the-loop-reclaiming-the-50-of-scientific-time-lost-to-data/
[4] Ardigen. Can you trust your model? Why explainability matters in AI-driven drug discovery. https://ardigen.com/can-you-trust-your-model-why-explainability-matters-in-ai-driven-drug-discovery/
