Clara Wong-Fannjiang

I'm a Senior Machine Learning Scientist at Genentech in South San Francisco, where I am part of the Frontier Research team within Prescient Design. My research works toward trustworthy scientific inquiry and decision-making using machine learning, with a current focus on uncertainty quantification methods.

I received my Ph.D. in Electrical Engineering and Computer Sciences from UC Berkeley in August 2023, where I was advised by Michael I. Jordan and Jennifer Listgarten as an NSF Graduate Research Fellow. I also spent a summer working with Nikhil Naik and Ali Madani at Salesforce Research. Prior to Berkeley, I conducted marine biology fieldwork with Kakani Katija at the Monterey Bay Aquarium Research Insitute and worked with David Sussillo as a Google AI Resident. I received my B.S. in Computer Science in 2016 from Stanford University, where I benefited from the guidance of Stephen Boyd, Coraline Rinn Iordan, and Fei Fei Li.

Beyond research, I am an avid hiker, fan of women's artistic gymnastics and aroids, and proud mom to an extraordinary son.

Email  ·  CV  ·  Scholar   

Research

I am broadly interested in the question of how to make trustworthy machine learning-guided scientific conclusions and decisions. My current work tackles this question through the lens of uncertainty quantification methods, and is particularly motivated by problems in drug discovery and clinical decision-making. Some highlighted work is below, where (α-β) denotes alphabetical ordering and * denotes equal contribution.

A diverse set of designed objects enables a practitioner to hedge against unknown future desiderata. We develop a machine learning-guided design method that navigates a trade-off between design diversity and achieving desirable predicted values, and apply it to the design of diverse AAV capsids with high packaging efficiency.

A discussion of considerations when using machine learning-based predictions to guide the design of novel objects, which by definition differ from the training data and are therefore particularly prone to prediction error.

We develop a framework for performing valid statistical inference—constructing confidence sets and p-values—when gold-standard data is supplemented with predictions from any predictor, machine learning-based or otherwise.

We characterize feedback covariate shift, a type of distribution shift in which the test and training data are dependent: the former is determined by the latter, such as in machine learning-guided design. We generalize conformal prediction to the feedback covariate shift setting.

Simple baselines outperform fine-tuned protein language models on protein fitness prediction.

Given fixed training data, algorithms for machine learning-guided design often consult a fixed predictive model. We propose instead retraining the predictive model as such algorithms proceed, on the same training data but carefully reweighted to reflect where the algorithm is currently "focused" in input space.

A spunky first step toward using machine learning to make sense of wild jellyfish telemetry collected in the Monterey Bay. A meta-lesson is that for some, perhaps most, unsolved scientific problems, insights and methods for gathering the right kinds of data matter more than machine learning ever will.

Talks

Simons Institute Workshop on AI≡Science: Strengthening the Bond Between the Sciences and AI, June 14, 2024 (recorded).
Machine Learning for Health Care and Life Sciences Workshop, ETH AI Center, November 30, 2023.
Microbiology Society Annual Conference, April 17, 2023.
EPFL Physics of Living Systems Seminar Series, November 18, 2022.
Machine Learning for Protein Engineering Seminar Series, October 18, 2022 (recorded).
AIDD Summer School on Advanced Machine Learning for Drug Disovery, May 13, 2022.

UC Berkeley EECS Ph.D. Dissertation Talk, May 10, 2023 (recorded).

Writing

Pop science piece on work by Karageorgi et al. (2019). Artwork by Alison Nguyen.

Pop science piece on Sara A. Stoudt's and Kelly Iknayan's work in statistical modeling in ecology. Artwork by Alison Nguyen.

Updated February 2025. Grateful user of Jon Barron's website code.