Learning path 05 · Beginner to advanced
Use AI thoughtfully in research
AI literacy has several entry points. Practical LLM use can begin early; training biological models and building tool-using agents require deeper foundations. Keep the scientific question and evidence visible at every stage.
IntermediateMachine learning and biological models
Learn what a model represents, optimizes, and can reasonably generalize to.
Before you begin: Basic Python, statistics, and introductory linear algebra.
What to understand
- Training, validation, test sets, leakage, baselines, and distribution shift.
- Embeddings and transformers; variational autoencoders and latent variables.
- Diffusion and conditional generation, with links to molecular design.
- Using a pretrained model versus fine-tuning or training a new one.
What you’ll be able to do
Describe a model's objective, input representation, baseline, and evaluation task.
Try this
Read one scvi-tools model guide and one protein-design model description. Compare their inputs, objectives, outputs, and evidence of usefulness.
Keep in mind
A benchmark score only speaks to its evaluation setup. Ask how data were split and whether the test resembles the intended scientific application.
A few good resources
- Dive into Deep Learning ↗
Build conceptual and practical foundations in deep learning; select chapters relevant to the model you are studying.
Book · Intermediate · Free online · English - Google Machine Learning Crash Course ↗
An applied introduction to model fitting, generalization, data handling, and evaluation.
Course · Intermediate · Free · English - scvi-tools tutorials and model guides ↗
Choose probabilistic models by task and modality, and read their assumptions alongside the worked tutorials.
Documentation · Advanced · Free; compute may cost · English - RFdiffusion ↗
Explore diffusion-based protein backbone generation and conditional design through the authors' code and examples.
Tool · Advanced · Code available; GPU compute may cost · English
BeginnerWorking with LLMs for science
Use an LLM for bounded tasks with explicit context and inspectable outputs.
Before you begin: A clear task and the ability to assess the result or consult a qualified reviewer.
What to understand
- Provide the question, source material, constraints, and desired output format.
- Ground literature summaries in supplied papers and check claims against the original text.
- For code assistance, provide schemas and small examples, inspect changes, and run meaningful checks.
- Record the model, date, inputs, important instructions, and human verification.
What you’ll be able to do
Create a useful research aid with a record of its sources and checks.
Try this
Give an LLM a paper's methods section and ask for a table of reported parameters with supporting quotations and locations. Mark absent information as not reported, then verify every row yourself.
Keep in mind
Fluent output can contain incorrect claims, citations, or code. Use suitable public or authorized data and check the service's data-handling terms before supplying unpublished or sensitive material.
A few good resources
- Ten simple rules for using large language models in science ↗
A science-specific starting point for transparent, critical LLM use, with discussion of verification and appropriate applications.
Paper · Beginner · Open access · English - Developing Software with Generative AI ↗
Practice AI-assisted software development through a research-computing lesson. The material is under development; check its current status and verify generated code.
Tutorial · Intermediate · Free · English - Hugging Face LLM course ↗
Understand language models and their practical ecosystem; pair technical learning with source verification and task-level checks.
Course · Intermediate · Free; optional compute may cost · English
AdvancedAgentic AI and scientific tools
Connect a model to tools while keeping execution and evaluation observable.
Before you begin: Programming, APIs, practical LLM use, and a reproducible baseline task.
What to understand
- Tool calling, structured inputs and outputs, retrieval, state, and execution traces.
- Bounded tool permissions, explicit stopping conditions, and human review where needed.
- Task-level evaluation, failure cases, tool-result verification, and cost tracking.
What you’ll be able to do
Specify and evaluate a small agent against a simpler scripted baseline.
Try this
Build a read-only assistant over a few public tutorials. Require source locations for answers and test it on a fixed set of answerable and unanswerable questions.
Keep in mind
Start with a narrow task and inspect the trace. Greater autonomy is not itself evidence of better scientific results.
A few good resources
- Hugging Face Agents Course ↗
An introduction to agent concepts, tools, and implementation. Evaluate any example on a bounded scientific task before expanding it.
Course · Advanced · Course free; external services may cost · English - Developing Software with Generative AI ↗
Practice AI-assisted software development through a research-computing lesson. The material is under development; check its current status and verify generated code.
Tutorial · Intermediate · Free · English
IntermediateEvaluate and document AI-assisted work
Make it clear which results come from evidence, computation, and model suggestions.
Before you begin: Practical LLM experience or an introductory machine-learning project.
What to understand
- Reference examples, baselines, held-out tasks, and explicit success criteria.
- Citation verification, numerical checks, reproducible tool outputs, and failure analysis.
- Versioned prompts or instructions, model settings, tool versions, and limitations.
What you’ll be able to do
Write a short evaluation report with successes, failures, and reproducibility details.
Try this
Compare an LLM-assisted solution with a manual or scripted baseline on five fixed tasks. Report factual accuracy and completion, not just writing quality.
Keep in mind
The model's own confidence is not an independent check. Verify important claims and calculations against external evidence or deterministic computation.
A few good resources
- Ten simple rules for using large language models in science ↗
A science-specific starting point for transparent, critical LLM use, with discussion of verification and appropriate applications.
Paper · Beginner · Open access · English - Developing Software with Generative AI ↗
Practice AI-assisted software development through a research-computing lesson. The material is under development; check its current status and verify generated code.
Tutorial · Intermediate · Free · English - Hugging Face Agents Course ↗
An introduction to agent concepts, tools, and implementation. Evaluate any example on a bounded scientific task before expanding it.
Course · Advanced · Course free; external services may cost · English - Quarto: getting started ↗
Combine analysis, narrative, and outputs in a reproducible report rather than copying results into an unrelated document.
Tutorial · Beginner · Free · English