Responsible AI Use, Validation, and Reproducibility in Bioinformatics

Goals for this talk

• Provide background information on AI for Bioinformatics.
• Cover the NIH Responsible AI Policy.
• Equip you with practical tips for minimizing the risks of using LLMs & agents for bioinformatics.

Bioinformatics is central to biomedical research

Audience roles: bioinformatician, computational biologist, data analyst, research software engineer – anyone who writes code to analyze biological data.

Academic Definition of Generative AI

Agentic LLMs can accelerate bioinformatics

NIH Responsible AI Policy

NIH AI Use Checklist

Human in the loop

Human in driving the loop

Practical tips for responsible agent use

Reproducibility is essential for science

Reproducible practices for responsible AI use

file system tree

/data/$USER/projects
├── my-rnaseq-project
│   ├── AGENTS.md
│   ├── data/
│   ├── docs/
│   ├── README.md
│   ├── results/
│   ├── scripts/
│   └── tests/
└── r-dataviz-explorer
    ├── references/
    ├── scripts/
    ├── SKILL.md
    └── tests/

Reproducible practices for responsible AI use

file system tree

/data/$USER/projects
├── my-rnaseq-project
│   ├── AGENTS.md
│   ├── data/
│   ├── docs/
│   ├── README.md
│   ├── results/
│   ├── scripts/
│   └── tests/
└── r-dataviz-explorer
    ├── references/
    ├── scripts/
    ├── SKILL.md
    └── tests/

Reproducible practices for responsible AI use

Review, test, and validate

From the NIH Responsible AI policy:

Do not rely on the technology to be a software developer by proxy:

• All well-written code must adhere to security design and ethical principles.
• All code output needs to be reviewed for completeness, quality, efficiency, and, most of all, security.
• Leverage manual and automated validation tools and testing technologies to help ensure these factors.
• If you cannot identify or understand what a piece of AI generated code does, you should not use it.

AI outputs may be incomplete, inaccurate, biased, or fabricated.

Review

• Read the entire output of each prompt and ensure you understand it.
• Be skeptical. Do not blindly trust that the output is complete or correct.
• Treat agents as untrustworthy, in terms of both inputs and outputs.
  • Do not expose passwords, access tokens, or sensitive information.
  • Refer to the NIH Responsible AI page - approved tools and data types for which types of data can be used with the tools we have access to.

Review

Testing & validation

Write tests to make sure the code behaves as expected.

• Unit tests for individual functions on trivial inputs/data.
• Integration tests run multiple functions or even an entire analysis on small & simulated datasets.
• R: testthat. Python: pytest, unittest

file system tree

my-python-project
├── data/
├── src/mypackage
│   └── main.py
└── tests
    └── test_main.py

src/mypackage/main.py

def gc_content(seq):
    """Fraction of G/C bases in a DNA sequence."""
    seq = seq.upper()
    return (seq.count("G") + seq.count("C")) / len(seq)

tests/test_main.py

from mypackage.main import gc_content

def test_gc_content():
    assert gc_content("GGCC") == 1.0
    assert gc_content("ATAT") == 0.0
    assert gc_content("atgc") == 0.5

Testing & validation

Testing & validation

Document everything

Inline comments, function docstrings, vignettes, tutorials, getting started guide

Rscript

#' Add together two numbers
#'
#' @param x A number
#' @param y A number
#' @returns The sum of `x` and `y`
#' @examples
#' add(1, 1)
#' add(10, 1)
add <- function(x, y) {
  x + y
}

A checklist for agent-assisted Bioinformatics

Final reminders

AI Use Statement

• Human roles: all text content was written by the author and reviewed by colleagues.
• Generative AI use: Copilot was used for reformatting and syntax lookup while drafting this presentation.
• Images: image sources are noted on the slides where they appear; images created by others may or may not have used GenAI for assistance. Images listed without a source were created by the author, without GenAI.