Crop yield prediction is a key challenge in plant science, requiring the integration of heterogeneous experimental data to extract predictive biological knowledge. In this study, we present a machine learning framework for crop yield prediction, combining phenotypic, environmental, and molecular gene expression data with yield outcomes. Using multi-season, multi-location field trials, we construct a unified dataset that captures plant development through time-resolved measurements of canopy dynamics, environmental conditions, and final performance.
We apply structured feature engineering tailored to each data type, including transformation, standardization, and aggregation of time-resolved measurements across developmental stages. These representations capture developmental and environmental variation in a form suitable for prediction, including biologically meaningful derived variables. Machine learning models trained on these features achieve strong predictive performance, with canopy development and environmental signals contributing most to accuracy, while molecular data provides complementary insights.
Beyond predictive performance, this study demonstrates how heterogeneous plant data can be integrated into reusable analytical frameworks that support biological interpretation and hypothesis generation. The results highlight the value of preserving experimental context and enabling consistent comparisons across experiments, which is essential for comparative and meta-analyses in plant science. These representations also support the development of models that can be transferred beyond a single study, facilitating hypothesis generation and exploration of plant responses under varying environmental conditions.
This approach aligns with the ELIXIR Plant Sciences Community, which promotes the integration of heterogeneous plant data for interoperable analysis and modelling in plant science.

Large language model agents are increasingly proposed as assistants for routine bioinformatics, yet existing benchmarks either score notebook reasoning (BixBench), need cloud infrastructure (BioAgent Bench), or depend on drifting LLM judges. None lets a small group track open-model progress reproducibly under limited resources.
We present BioTerm-Bench V1, a Harbor-compatible benchmark of 25 deterministically-graded terminal tasks spanning file-format manipulation, region operations, variant calling and filtering, RNA-seq normalisation, sequence search, and quality control. The benchmark is specifically designed to evaluate AI agents acting as orchestrators within federated data infrastructure — agents that assist with analysis directly on the node where sensitive genomic data resides, without data ever leaving the institution. This federated deployment requirement drives three core design choices: tasks target smaller, locally-runnable open models; evaluation is fully deterministic, enabling reliable automated quality checks within each node; and all reference data and graders are containerised for reproducible local execution.
We evaluated ten models, three times each, ranked by strict all-three-pass rule: GLM-5, Kimi-K2.5 24/25; DeepSeek-V3.2, Gemma-4-26b-A4B 23/25; GPT-5.4-mini, Qwen3.5-35b-a3b 18/25; Mistral-Small-2603 17/25; GPT-OSS-120B 13/25; GPT-5.4-nano 12/25; GPT-OSS-20B 6/25. Main discriminators were database search, multi-step ORF selection, and trio-aware variant logic. A full sweep runs ~1h per model. Harder successor packs are planned as V1 saturates, mirroring ARC-AGI. BioTerm-Bench V1 is released under Apache 2.0 at https://github.com/ML-Bioinfo-CEITEC/BioTerm-Bench.