On November 12, 2026, researchers Tian Cai, Andrew Kleist, Jaimin Patel, Duccio Malinverni, Alexander K. Taylor, and colleagues released a preprint titled AI for Biologists: The Knowledge Boundaries of Foundation Models in the Age of Agentic Systems. The study maps the operational boundaries of biological foundation models across five design elements, establishing where autonomous AI pipelines succeed, where they fail, and why wet-lab experimental ground truth remains indispensable for computational biology teams.

What happened

In a comprehensive review slated for the Annual Review of Biomedical Data Science, investigators from St. Jude Children’s Research Hospital and the University of California, Los Angeles examined the rapid integration of biological foundation models into autonomous agentic workflows. Foundation models trained on hundreds of millions of nucleotide sequences, protein structures, and single-cell transcriptomes—including ESM2, AlphaFold, AlphaGenome, Geneformer, scGPT, and Evo—have transformed biological prediction. Increasingly, researchers interact with these architectures through agentic systems that orchestrate multi-step analyses using large language models and external tool protocols.

However, the authors show that foundation model outputs cannot be universally trusted. They formalize a model’s “knowledge boundary” as the precise range of inputs, tasks, and biological conditions over which its predictions remain valid. This boundary is dictated by five interrelated design elements: input data, model architecture, training scheme, evaluation scheme, and post-analysis. Because foundation models are frequently repurposed far beyond their original pre-training scope, these architectural constraints travel with them into downstream agentic workflows.

Methodology and key findings

The preprint synthesizes extensive empirical evidence comparing biological foundation models against simpler computational baselines across diverse modalities:

Data and sampling constraints: Input corpora determine what can be learned. Protein models capture hydrophobic and electrostatic residue clustering from evolutionary co-variation, but likelihood scores heavily reflect historical phylogenetic sampling bias rather than intrinsic fitness. Tokenization choices also impose hard limits; multi-nucleotide tokens obscure single-nucleotide variants, while rank-based gene encoding discards quantitative expression magnitude.
Architecture and resolution: Attention mechanisms capture long-range interactions at quadratic computational cost, but context windows remain constrained. While models like AlphaGenome process up to one megabase, inter-chromosomal regulation falls outside their scope. Furthermore, AlphaFold2 relies heavily on multiple sequence alignments and frequently fails to predict the structural disruption caused by destabilizing point mutations.
Training objectives versus causality: Self-supervised masked and autoregressive objectives reward statistical correlation rather than causal biological mechanisms. As a result, sequence-to-function models such as Enformer, Borzoi, and AlphaGenome struggle to explain inter-individual expression variance and often predict the wrong direction of effect. In single-cell perturbation experiments, simple linear models and training data averages regularly match or outperform un-fine-tuned foundation models like Geneformer and scGPT.
Evaluation flaws and agentic misalignment: Benchmarks frequently suffer from data leakage and lack simple baseline controls. Examining four integration pathways—foundation models as tools, text representations of measurements, encoder embeddings, and fine-tuning verifiers—the authors identify instances where language model agents generate fluent rationales that directly contradict the foundation model tools they invoke.

Why it matters for research practice

For wet-lab and computational biology groups, the preprint delivers vital operational cautions. As research organizations build automated pipelines, uncritical adoption of foundation model predictions introduces subtle errors into experimental designs. Autonomous agents can propagate failures across multi-step analyses without signaling uncertainty, threatening computational reproducibility across discovery campaigns.

To safeguard scientific rigor, research leaders must audit training data provenance to verify that target organisms, cell states, and molecular assays fall within a model’s validated operational domain. The authors advocate establishing standardized, machine-readable boundary statements—modeled on model cards—that document context lengths, tokenization strategies, pre-training corpora, and baseline benchmark comparisons. Adopting these documentation standards is critical for conducting reproducible AI experiments.

The authors also conclude that computational biology should not wait for an all-encompassing foundation model. Instead, research workflows should orchestrate portfolios of specialized models coupled with automated self-driving laboratories that generate targeted physical ground truth whenever an inquiry crosses into boundary territory.

Sources

Tian Cai, Andrew Kleist, Jaimin Patel, Duccio Malinverni, Alexander K. Taylor, Malvika Sharan, Renato Umeton, and M. Madan Babu, AI for Biologists: The Knowledge Boundaries of Foundation Models in the Age of Agentic Systems (arXiv:2611.17015v1), November 12, 2026
arXiv PDF: 2611.17015v1 (Annual Review of Biomedical Data Science Preprint), November 12, 2026