📄 plant biology

BOTANIC-1: a series of long-context plant genomic foundation models in the agentic era

This paper introduces Botanic1, a family of long-context, agent-integrated genomic language models trained on unannotated plant data that outperform existing models in predicting trait-associated regions and provide biological insights through mechanistic interpretability, all while being made available to the research community to accelerate climate-resilient crop development.

Original authors: Barozet, A., Cabeli, V., Ogier du Terrail, J., Rukhovich, A., Janssoone, T., Klajer, G., Sheikhitarghi, Z., Andrews, G., Veran, C., Strouk, L.

Published 2026-09-07

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Barozet, A., Cabeli, V., Ogier du Terrail, J., Rukhovich, A., Janssoone, T., Klajer, G., Sheikhitarghi, Z., Andrews, G., Veran, C., Strouk, L.

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

Imagine a world where the code of life is written in a language of four letters: A, C, G, and T. These letters form the DNA sequences that dictate how a plant grows, how it resists drought, and how it fights off disease. For decades, scientists have tried to read this language, but it is a massive, complex text with billions of words, much of it written in a dialect that changes from one species to another. When a plant faces a harsh climate or a new pest, the solution often lies in a tiny change to this code—a single letter swap that might make a crop stronger. Finding that specific change is like searching for a needle in a haystack, but the haystack is the size of a library, and the needle is invisible to the naked eye. Traditional methods rely on comparing many plants to find patterns, but this is slow, expensive, and often leaves researchers with thousands of possibilities instead of one clear answer.

A team of researchers in Paris has now introduced a new way to read this biological text. They have built a family of artificial intelligence models, which they call Botanic1, designed specifically to understand the grammar of plant DNA. Unlike previous tools that needed to be taught the rules of biology by humans, these models learned the language on their own. They were fed the genetic sequences of 320 different plant species, ranging from tiny mosses to towering trees, and asked to predict missing letters in the sequence. By doing this millions of times, the models learned the hidden rules of how DNA is structured, how genes are turned on and off, and how tiny changes in the code affect the plant. The result is a system that can look at a long stretch of DNA and instantly sense which parts are critical for the plant's survival and which changes might be harmful or beneficial.

The researchers did not just build one model; they created a "factory" that uses intelligent software agents to manage the entire process of training and testing these models. These agents handle the heavy lifting, organizing data, running experiments, and fixing errors, allowing human scientists to focus on the big ideas. This approach allowed them to train models that can read DNA sequences up to 128,000 letters long, a length that captures the long-range interactions between different parts of the genome that were previously impossible to analyze together. In tests, these models outperformed every other existing tool for understanding plant genetics, including those that were much larger and trained on far more data. They were able to identify the most important parts of the DNA, such as the boundaries where genes start and stop, and even spot the specific signals that tell a cell how to cut and paste its genetic instructions.

What makes this discovery particularly powerful is that the models learned these rules without being told what they were. When the researchers looked inside the models to see what they had learned, they found that the AI had independently discovered biological concepts like "splice sites," which are the instructions for how a gene is edited before it becomes a protein. In one striking example, the model identified a specific location in a gene that the standard scientific databases had marked incorrectly for over twenty years. The model's internal logic pointed to a different spot, and when the researchers checked the history of that gene, they found that the original 1999 description was actually correct, and the model had rediscovered the truth that had been lost in later updates. This suggests that these models can act as a new kind of microscope, revealing biological truths that are hidden in the data but missed by human annotation.

The team also demonstrated how these models can work alongside human researchers in a new way. They set up a scenario where a general-purpose artificial intelligence agent was asked to find the cause of a specific trait in melons: why some plants produce female flowers and others produce both male and female flowers. The agent was given a list of thousands of genetic differences and asked to find the one responsible. When the agent tried to solve this using only standard tools, it could narrow the list down but could not pick the right answer. However, when the researchers gave the agent access to the Botanic1 model as a tool, the agent immediately ranked the correct genetic change as the number one candidate. The model provided a continuous measure of how surprising a genetic change was, which helped the agent distinguish the true cause from the noise.

This work represents a significant step forward in how we study plants. By teaching machines to read the language of DNA, the researchers have created tools that can accelerate the search for crops that can withstand a changing climate. These models are not just faster; they are more accurate and can see patterns that were previously invisible. They offer a way to move from guessing which genes might be important to knowing with high confidence which ones are. As the researchers release these tools to the scientific community, they open the door for a new era of plant biology, where the complex code of life can be understood and improved with a speed and precision that was previously out of reach. The models are now available for other scientists to use, promising to help breed better crops and understand the fundamental mechanisms of life in the plant kingdom.

Technical Summary: BOTANIC-1

Problem Statement

The development of climate-resilient crops is hindered by the slow breeding cycle and the difficulty of translating trait-associated loci into mechanistic, causal understanding. Current bioinformatics pipelines for identifying causal variants from Genome-Wide Association Studies (GWAS) and Bulk Segregant Analysis (BSA) face two primary limitations: a heavy reliance on expensive, slow-to-generate experimental data, and the inability to reliably separate causal variants from statistically correlated ones due to linkage disequilibrium (LD), a problem known as fine-mapping. This is particularly acute in plant genomics, where annotation data is often limited to reference genomes and regulatory mechanisms are less understood than in humans. While machine learning models like AlphaFold have revolutionized protein structure prediction, and protein language models have shown success in variant effect prediction, there is a lack of strong, general-purpose genomic language models (gLMs) specifically designed for plants that can reason over long genomic sequences and learn regulatory grammar beyond current annotations.

Methodology

The authors introduce Botanic1, a family of plant-specific genomic language models developed within an agentic Model Factory. This framework utilizes AI agents to automate and industrialize the model development lifecycle, including data curation, hyperparameter tuning, and experimental iteration, allowing researchers to focus on experimental design.

Model Architecture and Training

Architecture: Botanic1 models utilize a bidirectional Mamba2 (BiMamba2) backbone, an encoder-only architecture trained via Masked Language Modeling (MLM) on single-nucleotide positions. The models are released in four sizes: Botanic1-S (318M), Botanic1-M (688M), Botanic1-L (2.1B), and Botanic1-XL (3.2B) parameters.
Data Curation: The pre-training corpus spans 320 plant species (102 families, 48 orders), totaling 59 Gbp of data. The data pipeline employs rigorous controls:
- Intra-species: Functional window sampling is biased to include functional regions (1:1 ratio of functional to non-functional windows) rather than uniform sampling.
- Inter-window: Redundancy reduction is achieved using a MinHash filter (similarity threshold 0.3) to remove ~20% of redundant windows.
- Inter-species: Genome length correction is applied to balance representation across species of varying sizes.
Context Extension: To capture long-range regulatory interactions (up to 128 kbp), the authors employ a four-stage curriculum learning approach. Starting from an 8 kbp pre-training context, the model is continued pre-trained on progressively longer windows (16k, 32k, 64k, and 131k bp) without catastrophic forgetting of short-context capabilities.

Evaluation Framework

The models are evaluated across a comprehensive suite of tasks:

Frozen Representations: Zero-shot performance on 22 tasks across 9 families (e.g., genomic region classification, causal variant discovery, splicing, chromatin accessibility) using the Plant Genomic Benchmark (PGB) and PlantCaduceus datasets.
Adaptation: Fine-tuning via full fine-tuning, LoRA, and IA3 on regulatory tasks (promoter/terminator strength, poly(A) site identification, lncRNA classification) and base-resolution ATAC-seq prediction.
Long-Context Capabilities: Evaluation via pseudo-perplexity (PPPL) on nested contexts and "Needle-in-a-Haystack" (NIAH) retrieval tasks up to 131 kbp.
Mechanistic Interpretability: Analysis using Sparse Autoencoders (SAEs) and in silico mutagenesis (ISM) to identify features corresponding to biological concepts (splice sites, coding regions, motifs).
Agentic Workflow: Integration of Botanic1 as a callable tool for a generalist LLM agent (Gemma 4) to rank causal variants in a retrospective melon sex-determination case study.

Key Contributions and Results

1. State-of-the-Art Performance with Efficiency

Botanic1 models achieve the top four positions on the aggregate benchmark ( $S_{test}^{bal}$ ) across 22 frozen-model evaluation tasks, outperforming all generalist and plant-specific competitors, including PlantCAD2-L, Evo 2, and Carbon models.

Efficiency: Botanic1-S (318M parameters) outperforms PlantCAD2-L (694M) and Evo 2-7B (7B) despite having significantly fewer trainable parameters and training on ~13× fewer tokens (314.6B vs. ~4T for PlantCAD2-L).
Scaling: Performance scales monotonically with model size. The models demonstrate that careful data curation and architecture choice (BiMamba2) are more critical than sheer model size or autoregressive training objectives for plant-specific tasks.

2. Long-Context Capabilities

Botanic1 models successfully extend their context from 8 kbp to 128 kbp (131,072 bp) without degrading short-context performance.

Retrieval: In NIAH tasks, context-extended models show improved retrieval accuracy, with Botanic1-S4 (extended to 128k) achieving a normalized score of 0.83, significantly outperforming non-extended models and competitors like PlantCAD2.
Biological Application: On chromatin accessibility prediction, extending context to 32 kbp improves performance when combined with smart pooling strategies (focusing on the central peak region) to avoid signal dilution.

3. Adaptation and Fine-Tuning

Botanic1 models adapt effectively to downstream tasks:

Regulatory Tasks: Botanic1-S outperforms specialized baselines (e.g., ChromBPNet) and other gLMs (AgroNT, PlantCAD2-L) on promoter/terminator strength prediction and lncRNA classification.
ATAC-seq Prediction: Pre-trained Botanic1-S surpasses the specialized ChromBPNet model in predicting base-resolution ATAC-seq profiles, with gains attributed to pre-training rather than architecture alone.
TF Binding: Fine-tuned Botanic1-S achieves a macro average precision of 0.72 on genome-wide TF-family binding tasks, outperforming bespoke architectures trained from scratch.

4. Mechanistic Interpretability

Sparse Autoencoders (SAEs) trained on Botanic1-M reveal interpretable features:

Biological Concepts: Features correspond to splice donor/acceptor sites, intron regions, and codon positions with high Matthews Correlation Coefficients (MCC > 0.87 for splice sites).
Annotation Correction: The model identified a discrepancy in the AT1G65170 gene annotation, predicting a splice donor site 54 bp downstream of the annotated position. This prediction aligned with the original 1999 BAC submission and the 2026 TAIR12 release, suggesting the model can detect errors in reference annotations.
Alternative Splicing: Features were found to encode quantitative splice-site usage (Percent Splice Site Usage), distinguishing between constitutive and intermediate-usage donors.

5. Zero-Shot Causal Variant Prioritization

Botanic1 demonstrates superior zero-shot performance in ranking experimentally validated causal variants:

Benchmark: On a new benchmark of 545 loci across 14 species, Botanic1-XL ranks the causal variant in the top 0.1% of candidates for 16.5% of studies and in the top 1% for 49.9% of studies, outperforming all other gLMs and bioinformatics baselines (PhyloP, PhastCons, Ensembl VEP).
Agentic Integration: In a case study involving a melon sex-determination locus, a generalist LLM agent (Gemma 4) failed to identify the causal variant using only text or standard bioinformatics tools. When equipped with Botanic1-L as a tool to score variants via log-likelihood ratios, the agent ranked the causal variant first in all 20 repetitions.

Significance

The paper claims that Botanic1 represents a significant step forward in plant genomics by providing a scalable, efficient, and interpretable foundation for plant research. The authors emphasize that:

Data-Centric Discipline: The success of Botanic1 validates the application of rigorous, data-centric pre-training methodologies (previously successful in LLMs) to plant genomics, showing that careful data curation and species coverage (320 species) are critical.
Agentic Workflows: The integration of gLMs into generalist agent workflows demonstrates a practical path to accelerating biological discovery, where agents can leverage specialized models to solve complex problems like causal variant discovery that stymie standard pipelines.
Biological Insight: Beyond benchmark performance, the models learn biologically meaningful representations (e.g., regulatory grammar, splice sites) that can be used to interpret genomic sequences and potentially correct existing annotations.
Long-Range Interactions: The ability to process 128 kbp contexts allows these models to capture long-range regulatory interactions previously inaccessible to standard genomic models, opening new avenues for studying gene regulation.

The authors position Botanic1 not just as a model, but as a foundational layer for a new era of agentic plant biology, where AI agents can reason directly over genomic sequences to accelerate the development of climate-resilient crops.