这个页面有中文版 →
Blog / Research

How can an undergrad do a real research project with public data?

How an undergraduate can do a small, real research project with public data in computational biology or AI, with a six-step plan.

Short answer

An undergrad can do a real research project with public data by keeping it small: pick one published data set, read its metadata before downloading anything, reproduce one result from the original paper, then ask one question the paper didn't. Record every step so someone else can rerun it, and get one experienced person to check the result. Published course reports show undergraduates doing exactly this with public microbiome data.

How can an undergrad do a real research project with public data?

You don't have a lab bench, a grant or a sequencing machine. You do have a laptop, and thousands of research data sets are open to anyone. In computational biology and in machine learning, that is enough to do a small study that is real: a question, a method, a result you can defend.

We read five pieces that bear on this. Two are course reports where undergraduates did research with public data: a University of Minnesota study (Kirkpatrick and colleagues, 2019) and a University of British Columbia course report (Sun and colleagues, 2022). Two are rule lists written for working researchers: ten rules for using public biological data (Oza and colleagues, 2023) and ten rules for reproducible computational research (Sandve and colleagues, 2013). The fifth is Andrej Karpathy's 2019 essay, A Recipe for Training Neural Networks, for the machine learning side.

What the sources agree on

New findings don't need new data. The public-data rules paper opens with the point that reuse serves validation, new questions and hypothesis generation, and says "novel studies do not always require new data." The two course reports show it at undergraduate level. At Minnesota, students in an introductory biology lab analyzed published human gut microbiome data in R; one example project asked whether caffeine intake tracks with certain groups of bacteria. At British Columbia, teams of three or four spent 16 weeks on public microbiome data sets, asking questions the original studies had not addressed.

The data set limits the question, so read the metadata first. The rules paper tells researchers to examine the metadata (what the samples are, how they were collected, how many, which groups) before downloading. The Minnesota authors note that their bioinformatics students' questions were "limited by available data and metadata." The British Columbia instructors ranked candidate data sets on three criteria: variables the original paper left unexplored, completeness, and data quality. They add that "more variables support more diverse research questions over time."

Look at the data and build something simple before anything clever. Karpathy's first step is to inspect the data by hand before writing any model code, looking for duplicates, bad labels and imbalance. His second is a simple baseline you can't easily get wrong. For choosing a model, his advice is "Don't be a hero": find the most related paper and copy its simplest working setup. The biology rules paper says the same in its own terms: check distributions and confounders with exploratory analysis first.

Keep a record someone else could rerun. The reproducibility rules ask you to track how every result was produced, avoid manual edits to data, put scripts under version control, note random seeds, and keep the data behind every plot. The public-data rules add: download with a script instead of by clicking, and write down the accession number, database version and download date.

Check the terms, and cite the data. Open repositories let you download freely, but the rules paper says acknowledgment is still expected, some data sets carry licenses or embargoes, and controlled-access human data requires a formal request. For a first project, stay with open, de-identified data. Cite the original paper and the accession.

Where they differ, and what they don't say

The course reports are not about students working alone. Both courses had instructors, scaffolding and feedback. The British Columbia team writes that they could source only one usable data set from local researchers and had to hire a teaching assistant to curate published ones. Finding a workable data set is real work, and a solo student should budget time for it.

The rule lists were written for researchers, and the public-data paper never mentions students. It does say to make sure the team has the methodological expertise to interpret the data. That is the honest gap in the "no lab needed" story: you can do the analysis alone, but you can't be the only person who checks it.

The evidence is also modest. The Minnesota study measured attitudes with end-of-course surveys, not learning or research quality. In the British Columbia report, eight teams chose to submit their manuscripts for formal review at an undergraduate journal, which shows the work can reach that standard, not that it usually does. None of the five says how graduate admissions readers or labs value a self-run public-data project. And Karpathy's essay is about training neural networks, not about choosing a research question.

A six-step plan for one student

This is our own synthesis of the sources above, scoped for one student over six to eight weeks.

  1. Start from a paper, not a repository. Pick a recent paper you can mostly understand whose data is open. In biology that usually means an accession number in the methods; in machine learning, a public benchmark.
  2. Read the metadata before downloading. List the variables, the sample count per group, and anything the paper measured but barely discussed.
  3. Reproduce one figure or one number. If you can't get close, find out why. That gap is often the most educational part of the project.
  4. Ask one new question the data can answer. Use the table below.
  5. Write down the outcome before you look. One sentence: what result would mean yes, what would mean no. Then run it, with a simple baseline or control.
  6. Write four pages and get them checked. Question, data, method, result, limits. Put the code in a public repository. Ask a graduate student, instructor or researcher to read it.
Question typeExampleWhy it fits a first project
Does the result hold in a second data set?Rerun the paper's comparison on another public cohortClear yes or no; method already defined
Does a simpler method do as well?A linear model against the published deep modelForces a baseline; a "no" is still a finding
Does an unexplored variable matter?A metadata field the paper collected but didn't analyzeUses what the data set uniquely offers
Does the result survive a different choice?Another normalization, split or random seedTeaches how fragile or robust results are

A null result is a legitimate ending. "The simpler model did not match the published one, and here is the evidence" is a real finding if every step can be rerun.

How STARC's research track relates

A small public-data study is something you can do on your own, and it gives you something concrete to show when you ask for more. Our research track is the other kind of experience: you join a project an active researcher is actually working on. A spot opens only when a mentor genuinely has room. It begins with a short trial task, then a milestone agreed up front, and publication is never promised. The FAQ covers the rest.

The short version

One paper, one data set, one reproduced result, one new question, one written record, one person who checks it. Small is fine. It has to be yours, and it has to be rerunnable.

Related questions

Do I need a lab or a professor to do research with public data?

You don't need a lab to start. Open repositories let anyone download data. But a paper on reusing public biological data advises making sure someone on the team has the methods expertise to interpret it, so plan to have one experienced person check your work.

Which public data set should I pick?

A University of British Columbia course ranked data sets on three things: variables the original paper didn't fully explore, completeness (every sample the paper mentions is really there), and sufficient data quality. Start from a paper you can read and understand, not from the biggest data set.

Does a public-data project count as real research?

It can. A University of Minnesota study found students in a computer-based research course reported interest and a sense of achievement similar to or higher than students in bench-based ones. Whether it counts for you depends on whether you asked a real question and can defend the answer. None of the sources say how admissions readers weigh it.

How long should a first public-data project take?

The course we read ran 16 weeks with teams and instructors. For one student working alone, we suggest a smaller scope over six to eight weeks: one data set, one reproduced figure, one new question. That is our suggestion, not a finding.

Looking for real research experience?

Our research track opens rarely and never promises a paper, but it is real research.

See the research track Apply

Follow STARC on X. New posts, what we're building, and the questions students send us: @Starc_Institute

Follow on X

Last updated: 2026-11-19. Written by the STARC team. Rules quoted here come from our how it works page; if they change, this post is updated.

More questions