AI × BIOLOGY
Explore AI in Biology
AI can predict a protein’s structure, suggest an annotation from its sequence, or help researchers select molecules for testing. Understanding these applications means separating a model’s prediction from experimental evidence.
Learn how researchers interpret these outputs, then compare projects using public structure records and protein annotations.
Select a step to see what happens.
Biological input
Start with the sequence and context
Identify the molecule and the biological question. Structure, function, and treatment effectiveness are different questions.
Predicted structure
Inspect the prediction and its confidence
A structure resource includes information about confidence and limitations. An uncertain region deserves a different interpretation from a well-supported one.
Experimental question
Decide which evidence is missing
Use the prediction to frame a question that experiments could investigate. A predicted shape alone does not prove a biological mechanism.
Schematic molecular form. It is not an experimentally determined structure.
THE BASICS
Sequence, structure and function answer different questions
Biological AI systems work with different kinds of material: molecular sequences, structures, experimental measurements and scientific papers. A structure model proposes an arrangement of atoms. A sequence model suggests an annotation about a protein. A classifier can rank compounds for experimental testing. To interpret any of these outputs, first identify what the model was asked to predict.[1][4][6]
Describing a protein, investigating its function and evaluating a model need different evidence. Choose the biological question before selecting software. A tool-using agent might retrieve records and organize papers, while a structure-prediction model addresses a different computational task.
APPLICATIONS
What can AI tell us about biology?
Follow the information through each task: the input, the model’s contribution, and the person responsible for checking the result.
Published biomolecular structure-prediction research
Predicting how biological molecules fit together
Researchers need a structural explanation for how particular molecules might interact.
- The material
- Specified molecular components, including protein sequences and information about other molecules.
- AI’s role
- AlphaFold 3 predicts joint structures involving proteins, nucleic acids, small molecules, and other components.
- The human decision
- Researchers interpret the arrangement and decide which questions could be tested experimentally.
The limitAlphaFold 3 produces a static prediction. The paper describes errors and limitations, including that these predictions do not capture molecular motion in solution or establish biological function.[1]
An annotation resource with explicitly identified AI predictions
Connecting a protein sequence with a possible role
A protein sequence needs interpretation before a reader can assess claims about its role.
- The material
- Amino-acid sequences and annotations in the UniProt Knowledgebase.
- AI’s role
- UniProt’s ProtNLM models generate protein annotations from sequences. Its ProtNLM2 pilot extends beyond names to information such as predicted function and location.
- The human decision
- A reader follows the annotation’s evidence and checks whether the support is computational, experimental, or a curator’s interpretation.
The limitThe documentation describes model errors and corroboration checks. Follow the annotation’s evidence label to distinguish a computational prediction from an experimental finding.[4][5]
Published antibiotic-discovery research; preclinical findings
Prioritizing molecules for experimental research
Researchers cannot experimentally examine every possible candidate molecule at once.
- The material
- Chemical structures paired with measurements of antibacterial activity and human-cell toxicity.
- AI’s role
- Wong and colleagues used graph neural networks to predict properties and identify chemical substructures associated with promising candidates.
- The human decision
- The team tested selected compounds experimentally, including investigation in mouse infection models.
The limitThe study reports a research pipeline and preclinical evidence. A computational ranking, a laboratory assay, and an effective treatment in people are separate milestones.[6]
CASE STUDY
Reading confidence in a predicted structure
Different regions and interfaces in a molecular model can have different levels of confidence.
In the paper’s example identified as PDB 7T82, different confidence views highlight uncertain tails and differences between molecular interfaces. Evaluation also compares predictions with experimental structures.[1]
EMBL-EBI explains that AlphaFold2’s pLDDT score describes local confidence, while PAE helps assess relative positions. A model can be confident about individual regions and less certain about how those regions sit together.[2][3]
A useful figure caption names the prediction method and explains what the confidence view shows. Identify the uncertain regions and describe which interpretation would need experimental support.
Follow the process Open a step to see what happens.
Identify the object
Name the molecule, organism, record, and prediction method.
Read the uncertainty
Separate local confidence from confidence in relative placement.
Locate the comparison
Ask which experimental observation could test the proposed interpretation.
State the boundary
Describe a hypothesis without claiming demonstrated function or a treatment.
Report the prediction and its uncertainty together. Local confidence and confidence in relative placement answer different questions.[1][2][3]
SKILLS & STUDY
What to learn first
- Biological foundations
- Learn what genes, proteins, and molecular interactions mean. Keep the organism and biological context attached to each question.
- Evidence and provenance
- Record the accession, version, method, and source of a claim. Explain whether you are looking at a prediction, an observation, or an interpretation.
- Data and evaluation
- Ask what a model was trained to predict and what an evaluation measures. A useful comparison needs a defined baseline and data that do not simply repeat training examples.
- Technical progression
- An annotated research brief can begin without coding. Comparing downloaded tables adds spreadsheet skills or Python. Training and validating a molecular model requires more mathematics, computing and molecular-science preparation.
PROJECT IDEAS
Questions to investigate with public biological evidence
Choose a question, inspect the requirements, and keep a brief for later. Each proposal includes a way to check the work.
Save an idea in this browser or copy its brief to discuss with a mentor.
No account neededExplain one public structure prediction
Which parts of a selected protein model support a cautious interpretation, and which remain uncertain?
- Intended output
- An annotated figure and plain-language evidence note with a record identifier, model version, and clearly separated observations and questions.
Skills, sources and checks
- Starting knowledge
- Protein basics, sequence positions, and the difference between local confidence and relative-position confidence. No model training is necessary.
- Material to consider
- A mentor-selected public AlphaFold Database entry and the EMBL-EBI explanation of its confidence measures; cite the original record and check figure reuse terms.
- How to judge the result
- Check the labels and confidence interpretation against the source record. Ask a knowledgeable reader to identify unsupported inferences or missing information.
Explain the selected prediction and what it leaves unresolved about the protein’s function.
Trace the evidence behind a protein annotation
What supports a narrow statement about a protein’s role?
- Intended output
- A claim-to-evidence map comparing the database statement, its evidence label, and selected supporting publications.
Skills, sources and checks
- Starting knowledge
- Basic molecular biology and the ability to distinguish a study’s methods from its conclusion. Automated extraction adds programming and source-checking work.
- Material to consider
- A small, defined set of public UniProt records and openly accessible papers. Keep species and sequence identifiers explicit; use no personal genetic information.
- How to judge the result
- Trace each statement to its evidence. Separate experimental findings from similarity-based or model-generated annotations and identify questions the selected material cannot answer.
Limit the comparison to the selected records, statements and supporting papers.
These projects use public scientific records and published research. They exclude personal genetic data, treatment recommendations and laboratory experiments.
MENTOR SUPPORT
Scope a biological evidence project with a mentor
STEMtiq’s Agentic AI Scholars Program teaches agent-building and evaluation. A literature-and-evidence question may fit its biomedical pathway. For an interest in protein structures or sequences, a mentor needs to confirm suitable sources, prerequisites and scope.
Discuss what you want to explain and how the output could be checked. The shared curriculum supports agent-building; specialist molecular-model training and experimental drug discovery are outside the offering described here.
5 weeks · 25 live hours · Grades 9-12
COMMON QUESTIONS
What students ask
Does a predicted protein structure prove what the protein does?
No. A structure can suggest questions about function, but the proposed role needs its own supporting evidence. Check both the model’s uncertainty and the experimental basis for the functional claim.[1][5]
Is every UniProt annotation generated by AI?
No. UniProt includes reviewed and unreviewed records with different evidence sources. Follow the annotation’s origin and supporting references. Even a manually entered statement may rely on computational evidence.[5]
What should I learn before reading an AlphaFold result?
Start with amino-acid sequences, protein regions, and the meaning of a predicted structure. Then learn the confidence measures for the particular method. Explaining one selected record carefully is a useful first milestone.
Could a short student project discover a new medicine?
A first project can analyze existing evidence or document a computational workflow. The antibiotic study also involved experimental testing and preclinical research. Developing an effective medicine requires evidence well beyond a short student analysis.[6]
How much coding is needed for a biology evidence project?
A carefully cited comparison can use a document and spreadsheet. Processing many records or building a source-retrieval workflow adds coding. Training a model is a separate, more demanding project and should be scoped accordingly.
Prepare a biological project proposal
Identify the biological question, the public records you would examine and the checks you would use. Discuss that proposal with a mentor to assess its fit with the program.
Sources & further reading
- [1]
- [2]
- [3]
- [4]
- [5]
- [6]
Sources inform this guide; their organizations do not endorse STEMtiq. Project ideas and diagrams are original editorial explanations.