A statistical geneticist who learned to read GWAS noise, then learned to read pitch decks. Now writing both kinds of papers - the published ones and the term sheets - out of GV's San Francisco office.
Brendan Bulik-Sullivan funds biotech companies for a living. The interesting part is the diligence. He once sat at a desk in Cambridge, Massachusetts, deriving why a regression line through chi-squared statistics could tell a geneticist whether their result was confounded by population structure or just genuinely polygenic. That paper, "LD Score Regression Distinguishes Confounding from Polygenicity in Genome-Wide Association Studies," landed in Nature Genetics in 2015. It is now a default citation in human genetics. The code lives on GitHub under a four-letter handle: bulik.
At GV, the venture firm formerly known as Google Ventures, he is one of the partners who can read the underlying methods of a Nature paper and then decide if the founders pitching off of it have actually built something. Most days the second part is the bigger problem.
He works out of the San Francisco office. He covers therapeutics and diagnostics. He was promoted from Principal to General Partner in June 2023, alongside three colleagues, in a wave that the firm described as people with "vastly different skill sets and experiences." His skill set is the rare one: a working researcher who became a working investor and did not lose the habit of opening the supplementary materials first.
The arc is short and unusually linear. UC Berkeley for a bachelor's in mathematics. Vrije Universiteit Amsterdam for a Ph.D. in statistical genetics. The Broad Institute's Stanley Center for Psychiatric Research, where the LD Score Regression work happened. Three years at Gritstone Oncology in the Bay Area, designing the neural networks that try to predict which mutated peptides a tumor will actually display on its surface - the central problem of personalized cancer immunotherapy. Then GV, in 2019. Then partner, four years later.
Most venture partners have a thesis. Bulik-Sullivan's thesis is closer to a method.
In his free time, Brendan enjoys reviewing data and reading scientific journals.- GV team bio, gv.com
Nature Genetics / Broad Institute
The problem: a genome-wide association study can find a signal because a trait really is polygenic, or because the population you sampled is structured in a way that confounds everything. The two look identical at a glance. Bulik-Sullivan and his collaborators showed that if you regress chi-squared association statistics on linkage disequilibrium scores, the slope and the intercept tell you which is which.
It sounds like a small statistical trick. It quietly reshaped how every large GWAS gets read. Almost every paper that estimates SNP-heritability from summary statistics now passes through ldsc, the open-source package he maintained as a graduate student.
Nature Biotechnology / Gritstone Oncology
At Gritstone, he led work on a neural network trained on tumor HLA peptide mass spectrometry data. The job: predict which of a patient's tumor mutations actually produce peptides that the immune system can see. Get it right and you can build a personalized cancer vaccine. Get it wrong and you ship a beautiful immunotherapy that targets nothing.
The paper is still cited in every immuno-oncology deck that talks about neoantigen prediction. The patents he is listed on for "neoantigen identification using hotspots" are assigned to Gritstone.
A founder walks in claiming their model predicts immunogenic neoantigens. They show ROC curves. Most investors nod. Bulik-Sullivan is one of maybe a dozen people in the asset class who has trained that exact model, on that exact kind of data, and knows precisely which validation sets are honest and which are not.
FIGURE 1 - LD Score Regression, in one line
The intercept (the "+1") is the part that tells you whether the inflation in your statistics is real signal or population structure. The slope tells you how heritable the trait is. Cited thousands of times.
Bulik-Sullivan's research has appeared in:
The shortlist most academic biotech founders are still chasing.
At GV he represents the firm on a handful of biotech boards and has a board observer seat at Areteia, a respiratory therapeutics developer. The mix is therapeutics-heavy and platform-friendly - the kind of bets that benefit from a partner who can argue about validation data without a translator.
Four letters on GitHub, the home of the LD Score Regression repository. The repo is still a standard tool in human genetics almost a decade after publication.
He did his doctorate at Vrije Universiteit Amsterdam - unusual for an American researcher who worked at the Broad in Cambridge.
GV's own bio note flags it without irony: he spends his free time reading scientific journals and reviewing data.
No public Twitter feed listed. No Substack. No newsletter. The work is the brand.
Lives in Berkeley, where he did his undergrad. Works out of GV's San Francisco office.
Listed as inventor on neoantigen patents assigned to Gritstone - rare among investors who joined biotech firms only briefly.
GV describes his focus as "innovative therapeutic biotechnology and diagnostics across all stages and therapeutic areas." That is the surface. The underneath is harder to copy: a willingness to spend an afternoon arguing with founders about how their training set was constructed, what their negative controls look like, and whether the readout would survive a tougher reviewer than the one they actually got.
It is also, quietly, an answer to a question that has hung over computational biotech for a decade. Can the people who built the methods become the people who fund the next generation of them? In Bulik-Sullivan's case, the answer happened to be yes - and the proof showed up on a partner page in 2023.