G: Unnatural Selection

Episode Summary

Executive Summary: This episode examines the promise and peril of using polygenic scores to predict human traits, especially intelligence, from DNA. It contrasts criticism of eugenic thinking with Steve Hsu’s view that genomics can guide embryo selection in IVF. The story highlights real predictive gains, but also major limits, population bias, and ethical concerns about turning probabilistic genetic scores into reproductive decisions.

Main Topics: Variation, Darwin, and the eugenics debate (Priority: 5/5): The episode opens by revisiting Lulu Miller’s argument that variation is central to evolution and warning how eugenics twisted Darwinian ideas into harmful human optimization projects. Steve Hsu’s path from physics to genomics (Priority: 4/5): Hsu describes how childhood fascination with Star Trek, IQ tests, and the idea of genetic supermen pushed him toward studying genomics and predicting human traits. From single-gene hopes to polygenic prediction (Priority: 5/5): The episode explains how genetic research moved from searching for one gene per trait to using huge datasets and genome-wide association studies to model complex traits influenced by thousands of loci. Predicting height and other traits with large datasets (Priority: 4/5): Hsu’s team uses machine learning on massive genomic datasets to predict height and some disease risks with notable accuracy, illustrating the power of modern GWAS-based tools. Embryo selection and consumer genetic testing (Priority: 5/5): Hsu’s company, Genomic Prediction, applies these predictors to IVF embryos, aiming to estimate risks for diseases and potentially cognitive traits before implantation. Ethical limits, uncertainty, and population bias (Priority: 5/5): Critics argue the predictions are weak at the individual level, may have unintended tradeoffs, and are based largely on European-descent data, making the tool less reliable and more inequitable for non-white populations. Parental choice and the future of reproductive genetics (Priority: 4/5): The episode ends by surveying IVF parents, showing that many are torn between wanting more information and fearing a future where wealthier people can optimize children genetically.

Key Arguments: Variation matters to evolution, but variation plus selection—not variation alone—drives biological success. Human traits like intelligence are highly polygenic, so modern genomics needs massive datasets rather than one-gene explanations. Polygenic prediction can be useful at the group level even when it is only modest at the individual level. Hsu argues embryo screening for cognitive risk is a reasonable extension of existing IVF genetic testing. Critics say a predictive model that is only slightly better than chance for individuals should not be treated as decisive in embryo selection. The datasets underlying these scores are heavily European, so the tools may not transfer well to non-European populations. Selecting against one genetic signal may unintentionally select for other harmful traits because the biology is not fully understood. The availability of these tools may widen inequality by giving wealthier families more reproductive options and information.

Data Points: Human genome completion date: June 26, 2000 - Bill Clinton announced the completion of the first survey of the human genome. Cost of first human genome sequence: $1 billion - Hsu contrasts the initial sequencing cost with later drops in price. Current genome sequencing cost: about $10,000 - Used to show the dramatic decline in sequencing costs. Genotyping cost: about $100 - Mentioned as evidence that DNA analysis has become cheap enough for wider use. Height predictor accuracy: plus or minus 1 inch - Hsu describes a genomic predictor for height built from GWAS data. Height-associated genomic locations: about 20,000 - The model for height is trained on many loci across the genome. Cognitive ability correlation: 0.3 to 0.4 - Correlation between genetic predictors and actual IQ or educational attainment-related outcomes. Individual prediction accuracy example: about 60% correct - Dan Benjamin explains that when comparing two people, the higher polygenic score corresponds to more schooling about 60% of the time. Incorrect ordering rate: about 40% - Benjamin’s example shows the model gets the higher-scorer wrong roughly 40% of the time. IVF births worldwide annually: about 1 million - Used to show the scale of embryo testing demand. Genetic influence on intelligence: 20% to 50% - The episode cites twin and related studies for heritability estimates of intelligence-related traits. Human genome scale: nearly 3 billion letters - Clinton’s announcement emphasizes the size of the genetic code being studied. Dataset size used in large GWAS: 1.1 million people - By summer 2018, multiple biobanks and companies pooled data for analyses. UK Biobank contribution: 500,000 genomes - A major public dataset contributing to the large-scale analyses.

Pivotal Quotes: "Variation. Variation. That, like, what makes a species really? Resilient is difference." — Pat Walters: Opening reflection on the earlier episode’s argument about Darwin and eugenics. "What we're doing is identifying outcomes that I think most people agree maybe are not good." — Steve Hsu: Hsu defends embryo screening as risk reduction rather than human optimization. "The evidence that we know today suggests the genes are responsible for somewhere between 20 to 50% of how smart people are." — Pat Walters: A summary of the heritability evidence used to justify the possibility of prediction.

Implications: Genomic embryo screening is becoming technically feasible, but its accuracy, fairness, and ethics remain contested. The industry may expand quickly, yet policy, counseling, and limits on use will shape whether it becomes routine or a new form of genetic inequality.