This post documents the first stage of evaluating an AI assistant agent for bioinformatics workflows, specifically targeting a molecular dynamics simulation pipeline for antibody-antigen complexes. While the agent was able to generate a reasonable plan in under 30 minutes, it encountered multiple issues related to dependency setup and reproducibility when the protocol was repeated.
This is Part 1 of a multi-part series. The current focus is on initial usability, reproducibility, and DevOps-related challenges.
A common scenario in bioinformatics teams is being tasked with a data analysis workflow that hasn’t been run in-house before. In such cases, significant time is often spent on:
This makes such tasks a strong candidate for AI agent support—particularly in rapidly generating scaffold workflows that can be later customized to institutional standards.
Objective:
To assess whether a predicted antibody-antigen structure is stable in a simulated dynamic environment using molecular dynamics (MD) tools.
Target Timeline:
Approx. 1 working day
Clear Explanation and Checking of Input
The agent provided a helpful summary of input formats after checking.
Great Interactive User Experience: Resilience to Execution Errors
When Jupyter notebooks failed to execute, the agent pivoted to generate standalone Python scripts instead.
Productivity Boost for Initial Planning
The agent’s scaffold saved several hours of setup and reading time. While not error-free, it offered a fast starting point.
Jupyter Integration Failed
Attempted three times, but notebooks failed to load or run. Ideally, the agent should a link like Google Colab or provide .ipynb files with download links.
Intepretation of Output Improvement
No output files were linked for download, making it harder to test or intepret outputs without direct agent assistance.
Further request on interpretation of output files encounter Conversation Error.
Tool Usage Lacked Context
The pipeline would benefit from a concise summary of each tool’s:
The first step is setup the enviroment to run/repeat locally since the Jupyter netwrok doesn’t work.
I provided the error messaged and agent generated newer Dockerfiles, but encountered different errors for 3 loops.
Error1_agent
failed to solve: nvidia/cuda:11.8-devel-ubuntu22.04: not found
Error2_agent: Miniconda Install Failure
wget https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-x86_64.sh
exit code: 133
Error3_agent with Miniforge
rosetta error: failed to open elf at /lib64/ld-linux-x86-64.so.2
I checked Dockerfile Error #1, and noticed should be corrected to nvidia/cuda:11.8.0-devel-ubuntu22.04. But later encountered another error:
ERROR: failed to build: failed to solve: process "/bin/sh -c wget https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-x86_64.sh -O miniconda.sh && bash miniconda.sh -b -p /opt/conda && rm miniconda.sh" did not complete successfully: exit code: 133
Another LLM-based agent (Claude) was tested as a fallback but encountered other new errors.
In this early test, the AI agent demonstrated promising capabilities in quickly drafting a new analysis workflow, particularly for less familiar tasks. However, repeatability and deployment remain major blockers, for example the environment setup and Docker builds.