Final project: a visual study note

Goal

Your main goal is to make an interesting topic easy to understand.

Create an individual study note in the form of a self-contained slide deck. Teach your classmates something interesting about machine learning for molecular science: what problem people are trying to solve, how an approach works, and why it matters. Assume your readers know the course material but have not read your selected paper or studied the startup.

You may explain a paper, study a startup, survey an open research question, or compare a focused set of papers or companies. Choose a question you can explain well within one deck.

Requirements

Submit PDF slides that stand on their own. A reader should be able to follow the explanation without a presentation, narration, or speaker notes.
Aim for 15–25 slides. This is a suggestion, not a strict limit; use fewer or more when it makes the explanation clearer.
Use at most four sentences per slide. Keep headings, figure labels, and citations concise. Develop one main idea at a time, using additional slides when needed.
Include plenty of explanatory figures. Use diagrams, annotated plots, comparisons, and worked examples throughout the deck. Explain what readers should notice, label axes and unfamiliar symbols, and make text readable. Credit the sources of reused or adapted figures.
Ground the study in October 2026. Explain the state of your topic as of October 2026, anchored in papers or substantive startup developments from 2026. Papers include preprints; startup developments may include a technical release, study, or demonstrated milestone. An established startup qualifies; it does not need to have been founded in 2026. Identify and cite the dated work or development. Older sources may provide background.
Focus on scientific questions and actual scientific impact. State the scientific question and explain how the research changes what scientists can understand, test, or discover. Support the impact with concrete evidence, such as a new scientific insight, an experimentally validated discovery, or a capability that enables previously impractical experiments. Distinguish demonstrated impact from potential future benefits; if impact has yet to be demonstrated, explain what evidence is needed.
Explain the evidence and its limits. Distinguish demonstrated results, company claims, and open questions. Cite the sources of factual claims and include references readers can follow.

This is an individual study-note project. Coding, original experiments, and a live presentation are not required. The deadline and submission channel will be announced separately.

Suggested structure

Use the following sequence as a guide, adjusting the number of slides for each part to suit your topic:

Motivating question: What are we trying to understand, and why is it interesting?
Necessary background: What concepts and terms does a classmate need before following the explanation?
Central idea and how it works: What does the method or company do? Show the inputs, the main steps, and the outputs. For an open-question survey, explain what current approaches achieve and where they fall short.
Concrete example and evidence: Walk through an example and explain the result. What does the evidence establish, and how does it change scientific understanding or practice?
Limitations: Where does the approach struggle, what remains untested, and what would convincing progress look like?
Takeaways and references: What should a reader remember, and where can they learn more?

For a comparison, use shared questions to connect the examples. For instance, compare how companies choose candidates, obtain experimental data, and check whether a proposed material works.

Assessment criteria

Ease of understanding is the main criterion. A classmate should finish the deck able to explain the central idea in their own words.

The study note will be assessed qualitatively on:

Clarity: The motivation is clear, unfamiliar terms are introduced, and each step follows naturally from the previous one.
Correctness: Scientific explanations are accurate, sources support the claims, and uncertainty is represented fairly.
Completeness: The deck supplies the background and reasoning needed to follow the chosen question, including the main evidence and limitations.
Scientific impact: The deck identifies a meaningful scientific question and critically explains the evidence for the research’s impact, including what remains to be demonstrated.
Effective figures: Visuals carry part of the explanation, with readable labels and clear takeaways.
Thoughtful synthesis: The deck connects ideas and explains their significance in your own words.

Suggested topics

These are optional starting points. You may choose another topic connected to the course that meets the requirements above. Each example below includes 2026 sources and an idea for a useful visual; develop your explanation to reflect the state of the topic as of October 2026.

The landscape of AI startups for materials discovery

Guiding question: Where do AI startups contribute to materials discovery, and how do their approaches differ?

Compare candidate design, simulation, experimental work, interpretation of results, and scaling production. Explain what each company aims to do and what its public evidence supports.

2026 starting points: CuspAI’s AI Materials Foundry, described by founding member Royce (July 20); Periodic Labs, “Building Labs that Learn” (September 15); and Radical AI’s report on materials testing with Purdue Applied Research Institute (August 7). These announcements provide starting points for examining the companies’ approaches and claims.

Suggested visual: Draw a discovery-workflow map and place the companies at the steps they address. Mark which capabilities are demonstrated and which are stated goals.

Why are people excited about self-driving labs?

Guiding question: What does a self-driving lab automate, who chooses the experiments, and what remains difficult?

Use one discovery campaign to trace how candidate selection, synthesis, measurements, and interpretation fit together. Identify where people make decisions and what level of autonomy the study actually demonstrates.

2026 starting point: Park et al., “Discovery of tunable and soluble organic emitters for solid-state lasers with a self-driving laboratory”, Nature Communications (February 19).

Suggested visual: Draw the experimental workflow, labeling the computational steps, automated equipment, and human decisions. Follow one candidate through the workflow.

What are the open questions in protein science?

Guiding question: What remains open in understanding and designing proteins after recent advances in AI?

Focus on a few connected questions, such as:

How can we predict the different shapes a protein adopts, how it moves between them, and how those motions affect function?
How can we connect sequence and structure to function across mutations and biological conditions?
How can we design proteins with reliable functions, such as enzymes and molecular switches?
What experimental evidence and benchmarks would show that these problems are being solved?

For each question, explain what current methods can do as of October 2026, what remains difficult, and what would count as meaningful progress. Use concrete examples to make the gaps understandable.

2026 starting points: Yang et al., “The past, present and future of de novo protein design”, Nature (April 29); and Wankowicz and Bonomi, “From possibility to precision in macromolecular ensemble prediction”, Nature Methods (May 18).

Suggested visual: Map sequence, structure, dynamics, and function, marking the connections that current methods can predict and those that remain open. Illustrate one gap with a protein that changes shape or performs a specific function.

What must AI learn beyond protein structure to help discover drugs?

Guiding question: How do structure prediction, binding affinity, and experimental validation contribute to drug discovery?

Explain what each task tells us and why a promising computational prediction still needs experimental testing. Use the company’s reported results to discuss what is supported and what remains to be established.

2026 starting point: Isomorphic Labs, “The Isomorphic Labs Drug Design Engine unlocks a new frontier beyond AlphaFold” (February 10), which links to the IsoDDE technical report.

Suggested visual: Trace a candidate from a predicted interaction through affinity assessment and experimental testing, with the question answered at each step.

What can a language model learn from DNA?

Guiding question: How does learning from DNA sequences support biological predictions, and what are the limits?

Explain the input representation and training task, then walk through one prediction or design example. Describe how the example is evaluated and what the result does and does not tell us about biology.

2026 starting point: Brixi et al., “Genome modelling and design across all domains of life with Evo 2”, Nature (March 4).

Suggested visual: Follow a short, labeled DNA sequence through the model to its output, connecting that output to the biological question and evaluation.

What does a “virtual cell” actually predict?

Guiding question: What changes in a cell can a model predict after a perturbation, and how well do those predictions transfer to a new context?

Explain a perturbation, the measured response, and the model’s inputs and outputs. Use one example to distinguish the measurements being predicted from the broader behavior of a living cell, then discuss generalization across contexts.

2026 starting point: Adduri et al., “Predicting cellular responses to perturbation across diverse contexts with State”, Cell (August), also linked from Arc Institute’s tools page.

Suggested visual: Diagram the starting cell population, the perturbation, the predicted response, and the experimental measurements used to evaluate it.