BixBench: a Comprehensive Benchmark for LLM-based Agents in Computational Biology

Ludovico Mitchener; Jon M Laurent; Alex Andonian; Benjamin Tenmann; Siddharth Narayanan; Geemi P Wellawatte; Andrew White; Lorenzo Sani; Samuel G Rodriques

arXiv · 2503.00096

BixBench: a Comprehensive Benchmark for LLM-based Agents in Computational Biology

Ludovico Mitchener·Jon M Laurent·Alex Andonian·Benjamin Tenmann·Siddharth Narayanan·Geemi P Wellawatte·Andrew White·Lorenzo Sani·Samuel G Rodriques

View original source ↗Read source document ↗Export metadata ↓

Abstract

Large Language Models (LLMs) and LLM-based agents show great promise in accelerating scientific research. Existing benchmarks for measuring this potential and guiding future development continue to evolve from pure recall and rote knowledge tasks, towards more practical work such as literature review and experimental planning. Bioinformatics is a domain where fully autonomous AI-driven discovery may be near, but no extensive benchmarks for measuring progress have been introduced to date. We therefore present the Bioinformatics Benchmark (BixBench), a dataset comprising over 50 real-world scenarios of practical biological data analysis with nearly 300 associated open-answer questions designed to measure the ability of LLM-based agents to explore biological datasets, perform long, multi-step analytical trajectories, and interpret the nuanced results of those analyses. We evaluate the performance of two frontier LLMs (GPT-4o and Claude 3.5 Sonnet) using a custom agent framework we open source. We find that even the latest frontier models only achieve 17% accuracy in the open-answer regime, and no better than random in a multiple-choice setting. By exposing the current limitations of frontier models, we hope BixBench can spur the development of agents capable of conducting rigorous bioinformatic analysis and accelerate scientific discovery.

Explore related subjects

q-bio.QM cs.AI

Keep this discovery

BibTeX RIS

Ludovico Mitchener, Jon M Laurent, Alex Andonian, Benjamin Tenmann, Siddharth Narayanan, Geemi P Wellawatte, Andrew White, Lorenzo Sani, Samuel G Rodriques. 2025-02-28. BixBench: a Comprehensive Benchmark for LLM-based Agents in Computational Biology. https://arxiv.org/abs/2503.00096

Cite the original work for its findings. Save a collection to share your selection of sources.

KEEP EXPLORING

Related papers

arXiv2026

Biology-in-the-loop: Amortized Adaptive Hit Discovery in CRISPR Screens

Many biological discovery problems require experiments to be selected sequentially under constrained budgets. CRISPR screening is a prominent example, as exhaustive perturbation testing is often infeasible and candidate perturbations must instead be prioritized over multiple experimental rounds. Despite the importance of this problem, existing benchmarks for adaptive hit discovery remain limited in scale and diversity. Here, we introduce AssayBench-Loop, a large-scale benchmark for adaptive hit discovery comprising 1,389 CRISPR screens across five phenotype categories. Beyond enabling systematic evaluation, its scale makes it possible to learn acquisition strategies across historical experiments. Building on this resource, we introduce AssayLoop, a sequential experimental design framework combining AssayFormer, a transformer-based amortized acquisition policy trained across historical screens to adapt from experimental feedback, with LLM-derived biological priors through an adaptive handoff. In this view, completed experiments become training data for learning how accumulated evidence should guide what to test next, while LLMs provide prior biological knowledge to seed the search. We further introduce AssayLLM, showing that the same principle can be extended directly to an LLM through task-specific post-training. On temporally held-out screens, AssayLoop achieves a 5.67-fold enrichment over random selection and recovers 27.7% of hits after assaying approximately 5% of the candidate library, outperforming existing adaptive-design methods and standalone LLMs, and AssayFormer alone. Performance improves with increasing historical training data and transfers to phenotype categories excluded from training. These results demonstrate the value of learning acquisition policies across historical experiments and combining them with broad biological priors for efficient adaptive hit discovery.

q-bio.QM↗

arXiv2026

Multi-Task Bacterial Colony Detection and Classification Using YOLOv8 with Edge Optimization for Resource-Constrained Deployment

Manual counting and classification of bacterial colonies are critical yet labor-intensive tasks in microbiology, prone to human error particularly on densely populated plates. This work proposes a multi-task deep learning framework trained on the Annotated Germs for Automated Recognition (AGAR) dataset (18,000 images; 9,202 training / 3,067 testing) to automate Colony Forming Unit (CFU) enumeration and species classification. A custom multi-task CNN employing global regression served as the baseline, but demonstrated limited performance in clustered colony environments due to the absence of spatial localization. To address this, a YOLOv8 object detection architecture was adopted with high-resolution 1024x1024 inputs, enabling instance-level colony detection and label assignment. The model achieved a classification accuracy of 98.13% and a counting accuracy of 98.27% (within a 10-colony margin), demonstrating strong predictive capability. To bridge the gap between model performance and practical deployability, the trained model was optimized through unstructured and structured pruning, ONNX conversion, and reduced-precision inference (FP32, FP16, INT8). On a Raspberry Pi 4B, ONNX FP32 and FP16 variants offered the best balance between inference speed (~6.4s) and accuracy (MAE ~2.20). Unstructured pruning preserved predictive accuracy (MAE ~2.01) without runtime gains, while structured pruning resulted in significant accuracy degradation (MAE ~6.3), revealing the sensitivity of instance-level colony detection to architectural compression. These findings provide practical guidance for selecting optimization strategies in resource-constrained laboratory deployments.

q-bio.QM↗

arXiv2026

ADMET-EvO: a self-evolving scientific agent for sustained research across heterogeneous tasks

Scientific agents can move beyond automated model building by using accumulated evidence to revise both their questions and experimental strategies. The challenge is sustaining this adaptation across heterogeneous tasks without overfitting decisions to internal validation. Absorption, distribution, metabolism, excretion and toxicity (ADMET) prediction provides a demanding setting across diverse assays, datasets and chemical domains. We therefore developed ADMET-EvO, an evidence-gated agent that formalizes endpoints, generates falsifiable hypotheses and tests interventions across data, feature and model axes. It carries supported, rejected and inconclusive outcomes forward to guide each new cycle. Across the 22-task Therapeutics Data Commons (TDC) ADMET benchmark, ADMET-EvO achieved the highest task-normalized score of 96.77. Evidence-guided selection reduced cumulative fitting time by 72.2% within a predefined non-inferiority margin. It also formalized 43 toxicity-related tasks and constructed endpoint-specific predictors. Together, these results show how ADMET-EvO can accumulate evidence, revise its strategy and expand its research scope over time.

q-bio.QM↗