Appendix G: Course Syllabus Template

Published

October 22, 2026

This syllabus is a modular educational template, not an accredited course offering or a substitute for local curriculum, research-integrity, data-governance, and institutional review. Faculty may adapt the structure, readings, and assignments to their specific context and audience.

Learning Objectives

Faculty using this template should be able to:

  1. Build a graduate course on AI in biology and drug discovery from the handbook’s foundations, research-domain, translation, and governance chapters.
  2. Teach model outputs, in-silico predictions, and benchmark scores as hypotheses that need experimental evidence before they change a program decision.
  3. Assess learners on evaluation design, evidence appraisal, and validation planning rather than on familiarity with particular tools.
  4. Adapt the sequence, workload, and research-domain coverage to the learners and local program requirements.
Using This Template
  • 12-week format: Standard graduate semester structure
  • 6-week intensive format: Combine weeks as indicated for industry training or accelerated programs
  • Modular use: Individual weeks can be extracted for journal clubs or focused workshops
  • Research-domain tracks: Week 7 lets each program choose the tissue, organismal, or environmental chapters that fit its learners

What this provides:

  • A 12-week graduate seminar syllabus using the Life Sciences AI Handbook as the primary text
  • Weekly required readings, learning objectives, and discussion questions, with handbook case studies attached to most weeks
  • Three written assignments built on the handbook’s evaluation worksheets, with a grading rubric
  • A policy for learners’ use of generative AI, a course evaluation plan, and a 6-week intensive format

Key pedagogical approach: Teach evaluation before capability: a model output is a hypothesis until an experiment, a benchmark, or a deployment context gives it evidentiary weight.

Illustrative assessment weights: Participation 15%, evaluation critique 15%, midterm 20%, validation plan 20%, and final proposal 30%. Faculty should revise these weights to match the learners, credit requirements, and institutional policies.

Course Overview

Course Title: AI for Biological Discovery: Evidence, Evaluation, and Translation

Level: Graduate seminar (PhD, MS, MD-PhD), postdoctoral training, or professional development for biotechnology and pharmaceutical research teams

Prerequisites: Graduate coursework in biology, biochemistry, or chemistry, and introductory statistics. No prior AI knowledge required.

Format: One 2-hour seminar per week (24 contact hours over 12 weeks), with about 4 hours of reading and assignment work outside class. Set the credit value under the institution’s credit-hour policy.

Primary Text: Tegomoh, B. (2026). The Life Sciences AI Handbook: Evidence for AI in Biology and Drug Discovery. DOI: 10.5281/zenodo.22073136

Learning Objectives:

By the end of this course, students will be able to:

  1. Explain how data infrastructure, pretraining, and task definition limit what a biological AI model can learn
  2. Match a model claim to the split, baseline, metric, and experimental evidence that could support it
  3. Read confidence outputs for structure, sequence, variant, and cell models without treating predictions as measurements
  4. Distinguish demonstrated, theoretical, and beyond-current capabilities across molecular, cellular, organismal, and therapeutic applications
  5. Design the experiments that would move a model prediction to a research or program decision
  6. Evaluate claims about self-driving laboratories and agentic research systems for loop completeness, provenance, and human control
  7. Apply reproducibility and information-hazard standards to AI-biology work before it is shared or published

12-Week Syllabus

Week 1: What AI Has Changed in Biology

Theme: Current models in the longer history of computational biology

Required Reading:

Learning Objectives:

  • Define life sciences AI as a discovery layer whose outputs remain hypotheses until evidence supports them
  • Explain why shared data, benchmarks, and well-posed biological tasks changed which methods mattered
  • Use history to calibrate new model claims without treating novelty as proof

Discussion Questions:

  1. What made protein structure prediction tractable for machine learning, and which other biological problems share those conditions?
  2. What would have to be true for a model output to change a decision in your own research program?
  3. Which AI-biology claims from the past five years would you now treat as demonstrated, and which remain projections?

Week 2: Data, Foundation Models, and Literature AI

Theme: Why data infrastructure sets the ceiling on model quality

Required Reading:

Learning Objectives:

  • Explain how provenance, metadata, negative data, assay context, and versioning determine what a model can learn
  • Describe how pretraining on sequences, structures, cells, and genomes changes representation learning, and why pretraining alone does not make a model a reasoning system
  • Assess literature and knowledge-graph tools by provenance, entity resolution, and citation boundaries

Discussion Questions:

  1. Your laboratory’s assay data will train a model. Which metadata would another group need to reuse it?
  2. When should a foundation model’s representation outperform a task-specific model, and how would you test it?
  3. A literature tool returns a confident answer with citations. What do you check before relying on it?

Week 3: Evaluating Biological AI

Theme: Matching the evidence to the claim

Required Reading:

Learning Objectives:

  • Match a biological or translational claim to the split, comparator, validity checks, and uncertainty analysis that can support it
  • Identify leakage, weak baselines, and metric choices that inflate reported performance
  • Convert an evaluation result into a bounded adopt, revise, reject, or monitor decision

Discussion Questions:

  1. A paper reports the best score yet on a standard benchmark. What would you check before concluding that the model generalizes to your system?
  2. When should a simple baseline be required, and what should count as beating it?
  3. Which evaluation results justify an experiment, and which justify a program decision?

Assignment: Evaluation critique (1,500 words): Critically evaluate a published biological AI model paper, using the universal decision record to structure the critique

Week 4: Protein Structure and Design

Theme: The best-established capability, and what it does and does not solve

Required Reading:

Learning Objectives:

  • Judge whether a predicted structure is usable from its confidence scores, conformational state, disorder, and assay context
  • Explain when a protein design counts as a result: expression, folding, binding, activity, stability, and specificity measured in the intended context
  • Distinguish in-silico confidence filters from experimental measurements

Discussion Questions:

  1. A predicted structure shows high confidence across one domain and low confidence elsewhere. How should that change the experiments you plan?
  2. What does a confident structure prediction not tell you about function, dynamics, or ligand binding?
  3. Which experimental results should a design paper report before its designs count as working proteins?

Week 5: Genomes and Variant Effects

Theme: Sequence-to-function models and the line between prediction and interpretation

Required Reading:

Learning Objectives:

  • Judge whether a sequence-based functional claim is supported by tissue, cell type, chromatin, perturbation, and measurement evidence
  • Separate pathogenicity, splicing, regulatory effect, protein function, ancestry, and penetrance as distinct evidence layers
  • Explain why a variant effect score enters clinical variant interpretation as one line of evidence and does not replace it

Discussion Questions:

  1. How should a variant effect score be weighed against functional assay data for the same variant?
  2. Which populations are underrepresented in the data behind the variant predictors you use, and how could that affect your results?
  3. What evidence would show that a regulatory sequence model generalizes to a cell type it was not trained on?

Week 6: Cells and Perturbations

Theme: Single-cell foundation models and the counterfactual problem

Required Reading:

Learning Objectives:

  • Distinguish demonstrated representation, integration, and annotation uses from less established perturbation and mechanistic claims
  • Design a task-specific evaluation with biological hold-outs, strong classical baselines, and uncertainty
  • Explain why held-out genes, combinations, cell types, donors, and laboratories test different claims

Discussion Questions:

  1. A perturbation model matches a linear baseline on held-out genes. What, if anything, has it shown?
  2. Which hold-out design would convince you that a model predicts responses in a cell type it has not seen?
  3. What would a virtual cell model need to demonstrate before its prediction could stand in for an experiment?

Midterm: Take-home exam covering Weeks 1–6

Week 7: Tissues, Organisms, and Environments (Choose Your Track)

Theme: Applying the same evaluation discipline beyond molecules and cells

Faculty Note: Select one or two chapters that fit the program and the learners’ research.

Suggested Options:

Learning Objectives:

  • Identify how platform, assay, species, and environmental context shape every downstream claim in the chosen domain
  • Apply the handbook’s demonstrated, theoretical, and beyond-current tiers to a domain outside molecular biology
  • Compare the validation burden across measurement systems

Discussion Questions:

  1. Which measurement artifacts in your chosen domain (batch, scanner, staining, recording modality, field conditions) could a model learn in place of biology?
  2. What does a strong result in one species, tissue, or field site establish about another?
  3. Which claim in your chosen chapter is closest to changing practice, and what evidence is still missing?

Week 8: Targets and Molecules

Theme: From target scores to compound series

Required Reading:

Learning Objectives:

  • Distinguish a target score from a target hypothesis that states disease context, direction of effect, mechanism, safety rationale, and modality
  • Evaluate generated molecules against route feasibility, binding, selectivity, developability, and biological evidence
  • Explain how an aggregate metric can hide chemically invalid outputs

Discussion Questions:

  1. A model ranks a target highly. What genetic, biological, and safety evidence would you need before committing a program to it?
  2. Why is a compound series, rather than a single generated structure, the unit that matters?
  3. Which ADMET predictions, if any, would you trust enough to skip an assay?

Week 9: Translation and Clinical Evidence

Theme: Why AI-supported discovery still fails downstream

Required Reading:

Learning Objectives:

  • Explain why benchmark success can fail to translate at the target, assay, molecule, toxicology, trial, or patient-subgroup decision
  • Weigh context of use, analytical validity, clinical validity, and clinical utility for a candidate biomarker
  • Define a context of use and validation record for a clinical-trial AI tool

Discussion Questions:

  1. What does a Phase IIa result for an AI-discovered drug show about the contribution of AI, and what does it leave unanswered?
  2. Which translational failure in the readings could better evaluation have caught earlier, and at what step?
  3. At what point does a trial-support model acquire regulatory consequence?

Assignment: Validation plan (2,000 words): Design the experiments that would move one published model prediction to a research or program decision, with the decision rule stated in advance

Week 10: Self-Driving Laboratories and Agentic Science

Theme: Automation, autonomy, and human control

Required Reading:

Learning Objectives:

  • Distinguish laboratory automation, lab-in-the-loop research, and a closed autonomous experimental loop
  • Evaluate a self-driving-laboratory claim for loop completeness, endpoint validity, exception handling, provenance, and independent validation
  • Specify the gates that literature retrieval, code execution, laboratory actions, and biological design each require in an agentic workflow

Discussion Questions:

  1. How many completed loop iterations should a paper report before it describes a system as autonomous?
  2. Which decisions in an agentic research workflow should always require a person’s sign-off?
  3. How would you verify a claim that an automated platform discovered something new?

Week 11: Reproducibility, Information Hazards, and Institutional Readiness

Theme: What a credible AI-biology program has to get right before it publishes

Required Reading:

Learning Objectives:

  • Separate computational reproducibility from experimental reproducibility, and specify what each requires
  • Apply the disclosure test: whether methods, details, or artifacts meaningfully change access to harmful capability
  • Identify the people, compute, data engineering, laboratory access, and governance a credible AI-biology program needs

Discussion Questions:

  1. What would another laboratory need to reproduce a result in your field: code, weights, data, protocols, or all of them?
  2. When should a capability result be published in full, in part, or not at all, and who should decide?
  3. What is the smallest team that could carry out the validation plan you wrote in Week 9?

Week 12: Practice and Emerging Frontiers

Theme: Choosing tools and reading frontier claims

Required Reading:

Learning Objectives:

  • Choose tools by modality fit, evidence, license, cost, reproducibility, and data privacy
  • Separate near-term frontiers from forecast-driven claims
  • Name the evidence that would move a frontier capability from theoretical to demonstrated

Discussion Questions:

  1. Which frontier claim in the readings has the strongest evidence today, and which has the weakest?
  2. What result would change your view of virtual cells or autonomous laboratories?
  3. Which tool would you adopt in your own work this year, and what would you require before relying on it?

Final Assignment: Research program proposal or evaluation report (3,000 words)

Assessment Structure

The weights below are an example, not a universal grading standard. Adapt them to course credit, learner level, accessibility, and institutional policy.

Component Weight Due
Class participation 15% Ongoing
Evaluation critique (Week 3) 15% Week 3
Midterm exam 20% Week 6
Validation plan (Week 9) 20% Week 9
Final proposal or report 30% Week 12

Grading Rubric

One rubric covers the three written assignments. Weight its rows to the assignment: the evaluation critique rests on evidence appraisal, the validation plan on experimental design, and the final proposal on all five.

Criterion Exceeds expectations Meets expectations Below expectations
Evidence appraisal Names the data, split, baseline, metric, and endpoint; identifies leakage or confounding; states what the evidence cannot show Names the evaluation design and its main limitations Treats a benchmark score or model confidence as proof of a biological effect
Source verification Traces every claim to the primary paper, dataset documentation, or model card; labels developer-reported performance as reported and preprints as preprints Cites primary sources for most claims Relies on press releases, social posts, or AI-generated summaries without verification
Experimental design Specifies experiments, controls, replication, and a decision rule set before the results are known Proposes relevant experiments and controls Proposes no experiment, or experiments that cannot test the claim
Reproducibility and disclosure Specifies the code, data, weights, and protocols to share, and applies the information-hazard test to anything withheld Addresses reproducibility and disclosure Omits reproducibility or disclosure
Communication Clear to the stated audience; states uncertainty plainly; the recommendation follows from the evidence Clear, with minor gaps in stating limits Unclear or overstated, or the recommendation does not follow from the evidence

Participation is assessed on preparation and on contributions that use evidence, not on volume.

Learner Use of Generative AI

Adopt a written policy before the first session and align it with the institution’s academic-integrity, research-integrity, and data-governance rules. A default policy for this course:

  • Learners may use generative AI tools to search, brainstorm, write code, and edit. They disclose each use at the end of the assignment: the tool, the task, and how the output was checked.
  • Learners verify every citation, statistic, and model claim against the primary paper, dataset documentation, or model card. An unverified or fabricated citation is graded under source verification, whatever produced it.
  • Unpublished data, proprietary sequences or compound structures, and data from human participants stay out of any tool that the institution or the data owner has not approved for that use.
  • No assignment asks for, and no submission should contain, methods for designing, enhancing, or synthesizing hazardous biological agents.
  • The instructor states which assessments exclude AI assistance and how that is checked. A take-home exam cannot rely on a prohibition alone.
  • Course requirements use only tools the institution provides to every learner.
  • Learners remain responsible for everything they submit.

The course itself teaches the skill this policy depends on. Week 2 covers literature tools that return confident answers with weak provenance, so one exercise can ask learners to verify and correct an AI-generated summary of a published model paper.

Evaluating the Course

Plan the evaluation before the course starts:

  • Learner feedback: a short mid-course check at Week 6 and an end-of-course survey on relevance, workload, and readings.
  • Learning: two comparable published model papers, one appraised in Week 1 and one in Week 12, scored with the evidence-appraisal and source-verification rows of the rubric. The change between them measures what the course taught better than a final exam score alone.
  • Application: a follow-up three to six months later asking whether learners have changed how they evaluate, adopt, or report AI tools in their own research.
  • Currency: an annual review of the readings against the handbook’s Changelog, because models, benchmarks, and evidence change within an academic year.

6-Week Intensive Format

For industry training or accelerated courses, combine weeks as follows:

Intensive Week Standard Weeks Focus
1 1–2 History, data infrastructure, foundation models
2 3–4 Evaluation principles, benchmarks, protein structure and design
3 5–6 Genome and variant models, single-cell and perturbation models
4 7–8 Research-domain track, targets and molecules
5 9–10 Translation, self-driving laboratories, agentic workflows
6 11–12 Reproducibility, information hazards, tools, frontiers

Supplementary Resources

Adaptation Notes for Instructors

Co-teaching helps: a machine learning or data science faculty member for Weeks 2 and 3, and a translational or regulatory scientist for Week 9.

For PhD and MS programs in the biological sciences: Emphasize Weeks 3–7 and require the evaluation critique on a paper from the learner’s own field.

For computational biology and bioinformatics programs: Expand Weeks 2, 3, and 6 with hands-on evaluation exercises using public models and datasets.

For translational and MD-PhD programs: Emphasize Weeks 8 and 9, and add a session on how clinical AI tools are validated against patient outcomes.

For industry research and development training: Use the 6-week format and replace the final proposal with an evaluation report on a tool the team is considering.

License

This syllabus template is released under the same CC BY 4.0 license as the Life Sciences AI Handbook. Faculty may adapt freely with attribution.