AI for Biologists: The Knowledge Boundaries of Foundation Models in the Age of Agentic Systems

Tian Cai, Andrew Kleist, Jaimin Patel, Duccio Malinverni, Alexander K. Taylor, Malvika Sharan, Renato Umeton, M. Madan BabuarXiv 2026 · arXiv

TL;DR

This paper introduces the concept of a "knowledge boundary"—the specific range of inputs and conditions where biological AI models can be trusted. By breaking down AI development into five core design elements, the authors provide a roadmap for researchers to navigate the strengths and pitfalls of foundation models and agentic systems in modern biology.

What Problem Does This Paper Address?

Biological research is currently being flooded by two types of AI: Foundation Models (FMs), which compress massive datasets (like protein sequences or single-cell transcriptomics) into reusable numeric representations, and Agentic Systems, which use Large Language Models (LLMs) to coordinate these tools into complex workflows.

The problem is that these models are often treated as "black boxes." A researcher might get a high-confidence prediction for a protein structure or a gene perturbation effect without knowing if that specific query actually falls within the model's reliable range. Because the convergence of AI and biology is so new, we lack a clear picture of what these models have truly "learned" versus what they are simply mimicking from biased training data.

Methods

As a review and perspective piece, the authors propose a conceptual framework to define a model's knowledge boundary based on five design elements:

  1. Data: The training corpus and tokenization strategy (e.g., amino acids vs. k-mers).
  2. Architecture: The underlying math (e.g., Transformers vs. CNNs) that determines which biological relationships can be captured.
  3. Training Scheme: The objective the model was rewarded to learn (e.g., predicting a masked amino acid vs. generating new sequences).
  4. Evaluation Scheme: How the model is tested, which determines which weaknesses remain hidden.
  5. Post-analysis: How the model is ultimately used (e.g., fine-tuning vs. zero-shot prediction).

The authors use the analogy of deciphering an extinct language: we have the "texts" (biological measurements), but no native speakers to confirm the "grammar" (biological mechanisms).

Figure 1. Complementary roles of biological foundation models and agentic systems.

Datasets

The paper reviews models trained on several massive public repositories:

  • Proteins: UniRef (hundreds of millions of sequences) and the Protein Data Bank (PDB).
  • Genomics: Diverse genomes spanning species (used by models like Nucleotide Transformer and Evo 2).
  • Single-cell: The CZ CELLxGENE Discover database (over 150 million cells).
  • Benchmarks: The authors highlight the importance of independent benchmarks like ProteinGym for fitness prediction and CASP for structure.

Key Insights and Ideas

  • Correlation $\neq$ Causation: Foundation models are excellent at pattern recognition (e.g., predicting residue contacts) but often lack a representation of the physical and chemical forces driving those patterns. They learn what occurs together, not what causes what.
  • The Agentic Paradox: Agentic systems can reach the "right" answer through reasoning that contradicts their own tools. The paper cites an instance where an agent correctly identified a causal gene but reported the wrong direction of effect, ignoring the output of the model it supposedly consulted.
  • Evolutionary Niche: The authors argue that we are currently in a "large generalist" phase of AI development. They predict a shift toward "adaptive radiation," where smaller, specialized models fill the boundary regions where general models currently fail.
Figure 3. Design elements of agentic systems and developmental maturity of AI-based tools.

Key Results

  • Successes: Protein language models (like ESM2) can recover biological properties like hydrophobicity and charge without being explicitly told them. They can also predict mutational effects that correlate moderately with experimental fitness.
  • Failures: Many foundation models struggle to outperform simple linear baselines when predicting responses to genetic perturbations. Furthermore, models like AlphaFold2, while revolutionary for structure, often fail to predict the structural disruption caused by single-point mutations.
  • Scale Limits: Increasing the number of parameters doesn't always help; larger models do not consistently improve zero-shot fitness prediction, suggesting that data quality and diversity are the current bottlenecks.

Limitations and Future Work

The authors identify several critical gaps:

  • The "History of Sampling" Bias: Models often learn the history of what researchers chose to study rather than pure biology.
  • Structured Boundary Statements: The authors propose that every model should be released with a machine-readable "boundary statement" (similar to Model Cards) that explicitly defines its training limits.
  • Closing the Loop: The future lies in integrating foundation models with self-driving laboratories. These robotic platforms can target the "knowledge boundaries" by running experiments specifically in the regions where models are most uncertain, creating a feedback loop that expands the boundary of what is known.
Figure 2. Design elements of biological foundation models.

Summary updated

Abstract

Artificial intelligence (AI) is changing how biological research is done. Biological foundation models trained on millions of nucleotides, protein, and single-cell measurements predict protein structures, mutation effects, and cellular responses to genetic perturbations. Agentic systems, AI assistants built on large language models, can run multistep analyses from plain-language requests and increasingly call foundation models as tools. Yet it is often unclear when these models' outputs can be trusted. We frame this as a model's knowledge boundary: the range of inputs, tasks, and conditions over which its outputs are reliable. This review explains how foundation models are built and proposes that five design elements (data, architecture, training scheme, evaluation scheme, and post-analysis) define this boundary. We examine evidence on where current models succeed or fail and present standardized descriptions of model limits so that researchers, and the agents acting for them, can use these tools reliably.
BibTeX
@misc{cai2026ai,
  author = {Tian Cai and Andrew Kleist and Jaimin Patel and Duccio Malinverni and Alexander K. Taylor and Malvika Sharan and Renato Umeton and M. Madan Babu},
  title = {{AI for Biologists: The Knowledge Boundaries of Foundation Models in the Age of Agentic Systems}},
  year = {2026},
  eprint = {2611.17015},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.17015},
}

Loading related content...

Figure