AI / ML in Chemistry & Biochemistry

This seminar will explore some topics of how artificial intelligence / machine learning are changing and impacting modern chemistry. It is your opportunity to develop your skills at leading a scientific discussion with your peers. What makes for a good scientific conversation? How do you engage your audience? How do you encourage participation?

Below are 14 sample topics that we will discuss (and we can open it up to other topics if you find something interesting). Each of you will be responsible for leading a ~45 minute discussion with the rest of the class on one of these topics. The order of assigned discussions has been chosen at random. At the first class, you may request a particular topic if it aligns with your research or professional interests. The Wednesday before your assigned presentation, you’ll be expected to email a brief (1 page) abstract to the rest of the class along with links to any articles or materials you’d like us to read in advance.

In each of these discussions, we’ll discuss both the AI/ML topic, and how that topic bears on the work of lab chemists and biochemists, and what this implies for the future of our field.

One important point: This is a chemistry class, so we’re interested in how AI/ML works in chemistry, why it works (or doesn’t work), and what preconditions were necessary for these developments. This is not a philosophy or theology class, so we’ll be less interested in the ethics or morality of using AI / ML. Those are fascinating topics, and we can certainly talk about them, but chemistry should be the focus here.

Attendance at all class sessions is required unless you have an officially excused absence. Your grades for this course will be based on the quality of the discussion you lead (50%) as well as your contributions to the discussions on other topics (50%).

The course Canvas page will include reading assignments from your weekly discussion leaders, which is a growing set of links to interesting essays, papers, videos, and online resources to get you started. These links should not be the only reading you do in preparing to lead a discussion; they are just there to get you started on your research.

To get us started, let's look at two interesting discussions:

Talking With ChatGPT about Cyclohexane (this was your instructor trying to figure out if ChatGPT had a conceptual understanding of chemistry)
Jordan Ellenberg's 2022 Christmas Lecture (@ND!) - A really interesting Markov model explanation of LLMs starts at around 22 minutes, and the crowd experiment starts at around 29 minutes. (Jordan Ellenberg is a mathematician who works at UW Madison who recently wrote a book called "Shape".)
Introduction to Artificial Neural Networks
Introduction to Diffusion Models

Syllabus & Fundamental concepts (Dr. Gezelter will lead)
What are the main types of AI/ML technologies? What are Markov chains, and how do these become large language models (LLMs)? What are neural networks? What are diffusion models? What do we mean when we say a person ‘understands’ chemistry? Does ChatGPT ‘understand’ cyclohexane?
AlphaFold, Protein Foundation Models, and Generative Protein Design
DeepMind’s AlphaFold revolutionized protein structure prediction, folding millions of known proteins with high accuracy and solving a decades-old problem. Complementing this, David Baker’s lab used AI-based design (Rosetta) to create entirely novel proteins with applications in vaccines, sensors, and nanomaterials. What exactly did AlphaFold solve, and what did it not solve? What were the preconditions for AlphaFold to develop? What does it mean to “generate” a protein? Is AI discovering the rules of protein evolution, or exploiting statistical regularities? How much experimental validation is still required?
Molecular Foundation Models: What Does a “Language Model for Chemistry” Learn?
Rather than training a separate model for every property, scientists are asking whether large models pre-trained on diverse chemical data can be adapted to other problems. This is now a major theme in both atomistic simulation and materials discovery. What exactly is a "foundation model" in chemistry? What should the equivalent of text, tokens, and pre-training be for molecules? Can one model learn representations useful across reactions, spectra, properties, and simulations? Some chemical representation topics to discuss: SMILES, molecular graphs, 3D coordinate generation, reaction sequences, spectra.
AI for Reaction Prediction and Retrosynthesis: Has It Become Useful?
Can a neural network predict a reaction? When does reaction prediction actually beat a competent chemist? How should retrosynthesis systems handle missing reactions or bad literature? Can LLMs improve synthesis planning? What happens when AI planning is coupled to databases, literature search, or laboratory execution? How should chemists benchmark these systems?
Chemical AI Agents: From ChemCrow to Autonomous Research Assistants
The frontier scientific agentic AIs can: search literature and databases, formulate plans, run computational chemistry programs, perform calculations, inspect and critique their results, revise their strategy, potentially interact with laboratory hardware. Recent work on multi-agent, self-correcting systems for autonomous chemical experimentation illustrates how quickly this area is moving.
Self-Driving Laboratories
The core idea is: AI proposes an experiment → robot performs the experiment → instrument collects data → AI interprets results → next experiment is selected. The field has progressed from isolated demonstrations toward multipurpose autonomous laboratories, although scalability, interoperability, and complete experimental provenance remain major unresolved problems.
Generative Molecular Design: Designing Drugs, Catalysts, and Functional Molecules
The forward problem is straightforward: “What properties will this molecule have?” The reverse problem is more interesting: “What molecule should I make to have these properties?” What generative methods are out there for drugs, ligands, catalysts, etc. Can AI generate molecules or materials that satisfy a desired set of properties?
AI for Spectroscopy and Structure Elucidation
 Can AI solve a structure from the data chemists actually collect (e.g. NMR, IR, MS, Raman, X-ray diffraction, or combinations of these). You can start with SpectraML, but this topic has grown significantly beyond that one tool.
Universal Machine-Learned Potentials (MLIPs) for Molecular Simulation
MLIPs replace traditional empirical force fields or expensive quantum mechanical calculations (or DFT) with neural networks, Gaussian processes, or kernel methods trained on high-fidelity quantum data. Can we build a general-purpose model of atomic interactions that works across many molecules and materials?
AI for Materials Discovery
 Google DeepMind’s GNoME system discovered over 2 million new stable inorganic crystal structures, validated via autonomous robotic synthesis. This is the start of an interesting story that could include: large materials databases (e.g. the materials genome project), crystal structure prediction, phase-diagram prediction, robotic synthesis, and experimental validation.
Data as Infrastructure: OMol25, Open Datasets, Synthetic Data, and Data Quality
 Last year, Meta and some of the U.S. national labs released Open Molecules 2025, which is a dataset of over 100 million quantum mechanical density‑functional theory calculations, enabling ML models with quantum‑level accuracy to train on vast amount of molecular data. Let’s have a conversation about some of the issues around large data sets: open vs. proprietary, experiments vs. computational data, negative results, metadata, bias and duplication, licensing & intellectual property, data contamination.
Can AI Discover New Chemistry?
What evidence would convince us that an AI had actually made a scientific discovery? In mathematics, there have already been several AI-assisted discoveries and proofs. In chemistry, is AI at a level where it can be a tool for scientific reasoning? Or is it still at the level of prediction and automation?
AI for teaching chemistry
This topic could include AI-powered virtual labs, LLMs for question-answering partners, or any of the myriad ways that AI is changing the face of chemical education (including by increasing the rate of cheating on papers and exams). What are the advantages and disadvantages of AI in chemical education? What does it mean to learn and understand a new chemical concept?
Trusting AI in Chemistry: Uncertainty, Reproducibility, Safety, Judgment
How do we know when AI is wrong? Let’s talk about: hallucinated chemistry and references, uncertainty quantification, out-of-distribution predictions, benchmark contamination, biased or incomplete chemical datasets, irreproducible ML papers, explainability, provenance and data governance, human oversight, how experimental validation should be incorporated into AI claims

Quality of Discussion you lead:		50%
Contribution to other discussions:		50%