This Nature News article gives a nice overview of the current status of CRISPR technology and its potential application in human reproduction. As we discussed in this bioethics conversation (Manifold Podcast #9 with philosopher Sam Kerstein of the University of Maryland), it is somewhat challenging to come up with examples where gene editing is favored over embryo selection (a well-established technology) for avoidance of a disease-linked mutation.
Nature: ... He found out about a process called preimplantation genetic diagnosis or PGD. By conceiving through in vitro fertilization (IVF) and screening the embryos, Carroll and his wife could all but eliminate the chance of passing on the mutation. They decided to give it a shot, and had twins free of the Huntington’s mutation in 2006.
Now Carroll is a researcher at Western Washington University in Bellingham, where he uses another technique that might help couples in his position: CRISPR gene editing. He has been using the powerful tool to tweak expression of the gene responsible for Huntington’s disease in mouse cells. Because it is caused by a single gene and is so devastating, Huntington’s is sometimes held up as an example of a condition in which gene editing a human embryo — controversial because it would cause changes that would be inherited by future generations — could be really powerful. But the prospect of using CRISPR to alter the gene in human embryos still worries Carroll. “That’s a big red line,” he says. “I get that people want to go over it — I do, too. But we have to be super humble about this stuff.” There could be many unintended consequences, both for the health of individuals and for society. It would take decades of research, he says, before the technology could be used safely.
James Cham is a partner at Bloomberg Beta, a venture capital firm focused on the future of work. James invests in companies applying machine intelligence to businesses and society. Prior to Bloomberg Beta, James was a Principal at Trinity Ventures and a VP at Bessemer Venture Partners. He was educated in computer science at Harvard and at the MIT Sloan School of Business.
In mathematics, a manifold is a topological space that locally resembles Euclidean space near each point.
Steve Hsu and Corey Washington have been friends for almost 30 years, and between them hold PhDs in Neuroscience, Philosophy, and Theoretical Physics. Join them for wide ranging and unfiltered conversations with leading writers, scientists, technologists, academics, entrepreneurs, investors, and more.
Steve Hsu is VP for Research and Professor of Theoretical Physics at Michigan State University. He is also a researcher in computational genomics and founder of several Silicon Valley startups, ranging from information security to biotech. Educated at Caltech and Berkeley, he was a Harvard Junior Fellow and held faculty positions at Yale and the University of Oregon before joining MSU.
Corey Washington is Director of Analytics in the Office of Research and Innovation at Michigan State University. He was educated at Amherst College and MIT before receiving a PhD in Philosophy from Stanford and a PhD in a Neuroscience from Columbia. He held faculty positions at the University Washington and the University of Maryland. Prior to MSU, Corey worked as a biotech consultant and is founder of a medical diagnostics startup.
This study reports a validation of Polygenic Risk Scores for Coronary Artery Disease in a French Canadian population. Outliers in PRS are much more likely to have CAD than typical individuals.
In our replication tests of a variety of traits (both disease risks and quantitative traits) using European ancestry validation datasets, there is strong consistency in performance of the predictors. (See AUC consistency below.) This suggests that the genomic predictors are robust to differences in environmental conditions and also moderate differences in ethnicity (i.e., within the European population). The results are not brittle, and I believe that widespread clinical applications are coming very soon.
Florian Wünnemann , Ken Sin Lo , Alexandra Langford-Alevar , David Busseuil , Marie-Pierre Dubé , Jean-Claude Tardif , and Guillaume Lettre
Genomic and Precision Medicine
Abstract
Background: Coronary artery disease (CAD) represents one of the leading causes of morbidity and mortality worldwide. Given the healthcare risks and societal impacts associated with CAD, their clinical management would benefit from improved prevention and prediction tools. Polygenic risk scores (PRS) based on an individual's genome sequence are emerging as potentially powerful biomarkers to predict the risk to develop CAD. Two recently derived genome-wide PRS have shown high specificity and sensitivity to identify CAD cases in European-ancestry participants from the UK Biobank. However, validation of the PRS predictive power and transferability in other populations is now required to support their clinical utility.
Methods: We calculated both PRS (GPSCAD and metaGRSCAD) in French-Canadian individuals from three cohorts totaling 3639 prevalent CAD cases and 7382 controls, and tested their power to predict prevalent, incident and recurrent CAD. We also estimated the impact of the founder French-Canadian familial hypercholesterolemia deletion (LDLR delta > 15kb deletion) on CAD risk in one of these cohorts and used this estimate to calibrate the impact of the PRS.
Results: Our results confirm the ability of both PRS to predict prevalent CAD comparable to the original reports (area under the curve (AUC)=0.72-0.89). Furthermore, the PRS identified about 6-7% of individuals at CAD risk similar to carriers of the LDLR delta > 15kb mutation, consistent with previous estimates. However, the PRS did not perform as well in predicting incident or recurrent CAD (AUC=0.56-0.60), maybe due to confounding because 76% of the participants were on statin treatment. This result suggests that additional work is warranted to better understand how ascertainment biases and study design impact PRS for CAD.
Conclusions: Collectively, our results confirm that novel, genome-wide PRS are able to predict CAD in French-Canadians; with further improvements, this is likely to pave the way towards more targeted strategies to predict and prevent CAD-related adverse events.
"PRSs, built using very large data sets of people with and without heart disease, look for genetic changes in the DNA that influence disease risk, whereas individual genes might have only a small effect on disease predisposition," said Guillaume Lettre, Ph.D., lead author of the study and an associate professor at the Montreal Heart Institute and Université de Montréal in Montreal, Quebec, Canada. "The PRS is like having a snapshot of the whole genetic variation found in one's DNA and can more powerfully predict one's disease risk. Using the score, we can better understand whether someone is at higher or lower risk to develop a heart problem."
Early prediction would benefit prevention, optimal management and treatment strategies for heart disease. Because PRSs are simple and relatively inexpensive, their implementation in the clinical setting holds great promises. For heart disease, early detection could lead to simple yet effective therapeutic interventions such as the use of statins, aspirin or other medications.
... The American Heart Association named the use of polygenic risk scores as one of the biggest advances in heart disease and stroke research in 2018.
Sadly, reaction to these breakthroughs in human genomics will follow the usual pattern:
It's Wrong! Genomes are too complex to decipher, GWAS is a failure, precision medicine is all hype, biology is so ineffably beautiful and incomprehensible, Hey, whaddaya, you're a physicist! ...
It's Trivial! I knew it all along. Of course, everything is heritable to some degree. Well, if you just get enough data...
The paper below by some of my MSU colleagues examines the performance of a number of ML algorithms, both linear and nonlinear, including deep neural nets, in genomic prediction across several different species.
When I give talks about prediction of disease risks and complex traits in humans, I am often asked why we are not using fancy (trendy?) methods such as Deep Learning (DL). Instead, we focus on L1 penalization methods ("sparse learning") because 1. the theoretical framework (including theorems providing performance guarantees) is well-developed, and (relatedly) 2. the L1 methods perform as well or better than other methods in our own testing.
The term theoretical framework may seem unusual in ML, which is at the moment largely an empirical subject. Experience in theoretical physics shows that when powerful mathematical results are available, they can be very useful to guide investigation. In the case of sparse learning we can make specific estimates for how much data is required to "solve" a trait -- i.e., capture most of the estimated heritability in the predictor. Five years ago we predicted a threshold of a few hundred thousand genomes for height, and this turned out to be correct. Currently, this kind of performance characterization is not possible for DL or other methods.
What is especially powerful about deep neural nets is that they yield a quasi-convex (or at least reasonably efficient) optimization procedure which can learn high dimensional functions. The class of models is both tractable from a learning/optimization perspective, but also highly expressive. As I wrote here in my ICML notes (see also Elad's work which relates DL to Sparse Learning):
It may turn out that the problems on which DL works well are precisely those in which the training data (and underlying generative processes) have a hierarchical structure which is sparse, level by level. Layered networks perform a kind of coarse graining (renormalization group flow): first layers filter by feature, subsequent layers by combinations of features, etc. But the whole thing can be understood as products of sparse filters, and the performance under training is described by sparse performance guarantees (ReLU = thresholded penalization?).
However, currently in genomic prediction one typically finds that nonlinear interactions are small, which means features more complicated than single SNPs are unnecessary. (In a recent post I discussed a new T1D predictor that makes use of nonlinear haplotype interaction effects, but even there the effects are not large.) Eventually I expect this situation to change -- when we have enough whole genomes to work with, a DL approach which can (automatically) identify important features (motifs?) may allow us to go beyond SNPs and simple linear models.
Note, though, that from an information theoretic perspective (see, e.g., any performance theorems in compressed sensing) it is obvious that we will need much more data than we currently have to advance this program. Also, note that Visscher et al.'s recent GCTA work suggests that additive SNP models using rare variants (i.e., extracted from whole genome data), can account for nearly all the expected heritability for height. This implies that the power of nonlinear methods like DL may not yield qualitatively better results than simpler L1 approaches, even in the limit of very large whole genome datasets.
Christina B. Azodi, Andrew McCarren, Mark Roantree, Gustavo de los Campos, Shin-Han Shiu
The usefulness of Genomic Prediction (GP) in crop and livestock breeding programs has led to efforts to develop new and improved GP approaches including non-linear algorithm, such as artificial neural networks (ANN) (i.e. deep learning) and gradient tree boosting. However, the performance of these algorithms has not been compared in a systematic manner using a wide range of GP datasets and models. Using data of 18 traits across six plant species with different marker densities and training population sizes, we compared the performance of six linear and five non-linear algorithms, including ANNs. First, we found that hyperparameter selection was critical for all non-linear algorithms and that feature selection prior to model training was necessary for ANNs when the markers greatly outnumbered the number of training lines. Across all species and trait combinations, no one algorithm performed best, however predictions based on a combination of results from multiple GP algorithms (i.e. ensemble predictions) performed consistently well. While linear and non-linear algorithms performed best for a similar number of traits, the performance of non-linear algorithms vary more between traits than that of linear algorithms. Although ANNs did not perform best for any trait, we identified strategies (i.e. feature selection, seeded starting weights) that boosted their performance near the level of other algorithms. These results, together with the fact that even small improvements in GP performance could accumulate into large genetic gains over the course of a breeding program, highlights the importance of algorithm selection for the prediction of trait values.
In an earlier post I collected links related to recent progress in Polygenic Risk Scores (PRS) and health care applications. The paper below describes a new (published in 2019) predictor for Type 1 Diabetes (T1D) that achieves impressive accuracy (AUC > 0.9) using 67 SNPs. It incorporates model features such as nonlinear interactions between haplotypes.
T1D is highly heritable and tends to manifest at an early age. One application of this predictor is to differentiate between T1D and the more common (in later life) T2D. Another application is to embryo screening. Genomic Prediction has independently validated this predictor on sibling data and may implement it in their embryo biopsy pipeline, which includes tests for aneuploidy, single gene mutations, and polygenic risk.
Sharp, et al.
Diabetes Care 2019;42:200–207 | https://doi.org/10.2337/dc18-1785
OBJECTIVE
Previously generated genetic risk scores (GRSs) for type 1 diabetes (T1D) have not captured all known information at non-HLA loci or, particularly, at HLA risk loci. We aimed to more completely incorporate HLA alleles, their interactions, and recently discovered non-HLA loci into an improved T1D GRS (termed the “T1D GRS2”) to better discriminate diabetes subtypes and to predict T1D in newborn screening studies.
RESEARCH DESIGN AND METHODS
In 6,481 case and 9,247 control subjects from the Type 1 Diabetes Genetics Consortium, we analyzed variants associated with T1D both in the HLA region and across the genome. We modeled interactions between variants marking strongly associated HLA haplotypes and generated odds ratios to create the improved GRS, the T1D GRS2. We validated our findings in UK Biobank. We assessed the impact of the T1D GRS2 in newborn screening and diabetes classification and sought to provide a framework for comparison with previous scores.
RESULTS
The T1D GRS2 used 67 single nucleotide polymorphisms (SNPs) and accounted for interactions between 18 HLA DR-DQ haplotype combinations. The T1D GRS2 was highly discriminative for all T1D (area under the curve [AUC] 0.92; P < 0.0001 vs. older scores) and even more discriminative for early-onset T1D (AUC 0.96). In simulated newborn screening, the T1D GRS2 was nearly twice as efficient as HLA genotyping alone and 50% better than current genetic scores in general population T1D prediction.
CONCLUSIONS
An improved T1D GRS, the T1D GRS2, is highly useful for classifying adult incident diabetes type and improving newborn screening. Given the cost-effectiveness of SNP genotyping, this approach has great clinical and research potential in T1D.
The figure below gives some idea as to the ability of the new predictor GRS2 (panels B and D) to differentiate cases vs controls, and T1D vs T2D.
Corey and Steve talk with Joe Cesario about his recent work which argues that, contrary to activist claims and media reports, there is no widespread racial bias in police shootings. Joe discusses his analysis of national criminal justice data and his experimental studies with police officers in a specially designed realistic simulator. He maintains that racial bias does exist in other uses of force such as tasering but that the decision to shoot is fundamentally different: it is driven by specific events and context, rather than race.
Cesario is associate professor of Psychology at Michigan State University. He studies social cognition and decision-making. His recent topics of study include police use of deadly force and computational modeling of fast decisions. Cesario is dedicated to reform in the practice, reporting, and publication of psychological science.
Is There Evidence of Racial Disparity in Police Use of Deadly Force? Analyses of Officer-Involved Fatal Shootings in 2015–2016 https://journals.sagepub.com/doi/abs/...
In mathematics, a manifold is a topological space that locally resembles Euclidean space near each point.
Steve Hsu and Corey Washington have been friends for almost 30 years, and between them hold PhDs in Neuroscience, Philosophy, and Theoretical Physics. Join them for wide ranging and unfiltered conversations with leading writers, scientists, technologists, academics, entrepreneurs, investors, and more.
Steve Hsu is VP for Research and Professor of Theoretical Physics at Michigan State University. He is also a researcher in computational genomics and founder of several Silicon Valley startups, ranging from information security to biotech. Educated at Caltech and Berkeley, he was a Harvard Junior Fellow and held faculty positions at Yale and the University of Oregon before joining MSU.
Corey Washington is Director of Analytics in the Office of Research and Innovation at Michigan State University. He was educated at Amherst College and MIT before receiving a PhD in Philosophy from Stanford and a PhD in a Neuroscience from Columbia. He held faculty positions at the University Washington and the University of Maryland. Prior to MSU, Corey worked as a biotech consultant and is founder of a medical diagnostics startup.
In this scenario Ted Chiang forecasts that recipients of government-funded genetic enhancement will not catch up to children of elites who receive similar enhancements. The latter are born to rich, highly educated parents and have access to elite social networks, better schools, etc. The system is still not entirely fair (i.e., invariant to accidents of birth), because many non-genetic advantages still exist. But can we ever achieve equality of outcome? At what cost?
Nevertheless, perhaps the beneficiaries of the Gene Equality Project are at least better off than their siblings who were not in the program?
It is interesting that the Times is already flirting with the idea of redistribution of genetic endowments. See also The Neanderthal Problem.
It’s 2059, and the Rich Kids Are Still Winning
DNA tweaks won’t fix our problems.
Ted Chiang is an award-winning science fiction writer.
Editors’ note: This is the first installment in a new series, “Op-Eds From the Future,” in which science fiction authors, futurists, philosophers and scientists write op-eds that they imagine we might read 10, 20 or even 100 years in the future.
Last week, The Times published an article about the long-term results of the Gene Equality Project, the philanthropic effort to bring genetic cognitive enhancements to low-income communities. The results were largely disappointing: While most of the children born of the project have now graduated from a four-year college, few attended elite universities and even fewer have found jobs with good salaries or opportunities for advancement. With the results in hand, it is time for us to re-examine the efficacy and desirability of genetic engineering.
The intentions behind the Gene Equality Project were good. Therapeutic genetic interventions, such as correcting the genes that cause cystic fibrosis and Huntington’s disease, have been covered by Medicare ever since their approval by the Food and Drug Administration, making them available to the children of low-income parents. However, augmentations like cognitive enhancements have never been covered — not even by private insurance — and were available only to affluent parents. Amid fears that we were witnessing the creation of a caste system based on genetic differences, the Gene Equality Project was begun 25 years ago, enabling 500 pairs of low-income parents to increase the intelligence of their children.
The project offered a common cognitive-enhancement protocol involving modifications to 80 genes associated with intelligence. Each individual modification had only a small effect on intelligence, but in combination they typically gave a child an I.Q. of 130, putting the child in the top 5 percent of the population. This protocol has become one of the most popular enhancements purchased by affluent parents, and it is often referenced in media profiles of the “New Elite,” the genetically engineered young people who are increasingly prevalent in management positions of corporate America today. Yet the 500 subjects of the Gene Equality Project are not enjoying career success that is remotely comparable to the success of the New Elite, despite having received the same protocol.
A range of explanations has been offered for the project’s results. White supremacist groups have claimed that its failure shows that certain races are incapable of being improved, given that many — although by no means all — of the beneficiaries of the project were people of color. Conspiracy theorists have accused the participating geneticists of malfeasance, claiming that they pursued a secret agenda to withhold genetic enhancements from the lower classes. But these explanations are unnecessary when one realizes the fundamental mistake underlying the Gene Equality Project: Cognitive enhancements are useful only when you live in a society that rewards ability, and the United States isn’t one.
It has long been known that a person’s ZIP code is an excellent predictor of lifetime income, educational success and health. Yet we continue to ignore this because it runs counter to one of the founding myths of this nation: that anyone who is smart and hardworking can get ahead. Our lack of hereditary titles has made it easy for people to dismiss the importance of family wealth and claim that everyone who is successful has earned it. The fact that affluent parents believe that genetic enhancements will improve their children’s prospects is a sign of this: They believe that ability will lead to success because they assume that their own success was a result of their ability.
For those who assume that the New Elite are ascending the corporate ladder purely on the basis of merit, consider that many of them are in leadership positions, but I.Q. has historically had only a weak correlation with effectiveness as a leader. Also consider that genetic height enhancement is frequently purchased by affluent parents, and the tendency to view taller individuals as more capable leaders is well documented. In a society increasingly obsessed with credentials, being genetically engineered is like having an Ivy-League M.B.A.: It is a marker of status that makes a candidate a safe bet for hiring, rather than an indicator of actual competence.
This is not to say that the genes associated with intelligence play no role in creating successful individuals — they absolutely do. They are an essential part of a positive feedback loop: When children demonstrate an aptitude at any activity, we reward them with more resources — equipment, private tutors, encouragement — to develop that aptitude; their genes enable them to translate those resources into improved performance, which we reward with even better resources, and the cycle continues until as adults they achieve exceptional career success. But low-income families living in neighborhoods with underfunded public schools often cannot sustain this feedback loop; the Gene Equality Project didn’t offer any resources besides better genes, and without these additional resources, the full potential of those genes was never realized.
We are indeed witnessing the creation of a caste system, not one based on biological differences in ability, but one that uses biology as a justification to solidify existing class distinctions. It is imperative that we put an end to this, but doing so will take more than free genetic enhancements supplied by a philanthropic foundation. It will require us to address structural inequalities in every aspect of our society, from housing to education to jobs. We won’t solve this by trying to improve people; we’ll only solve it by trying to improve the way we treat people.
This doesn’t necessarily mean that the Gene Equality Project is something that never needs to be repeated. Instead of thinking of it as a cure to an illness, we could think of it as a diagnostic test — something we would conduct at regular intervals to gauge how close we are to reaching our goal. When the beneficiaries of free genetic cognitive enhancements become as successful as the ones whose parents bought the enhancements for them, only then will we have reason to believe that we live in an equitable society.
Finally, let’s recall one of the arguments made during the original debate about legalizing genetic cognitive enhancements. Some proponents claimed that we had an ethical obligation to pursue cognitive enhancements because of the benefits to humanity that would accrue as a result. But there have surely been many geniuses whose world-changing contributions were lost because their potential was crushed by their impoverished surroundings.
Our goal should be to ensure that every individual has the opportunity to reach his or her full potential, no matter the circumstances of birth. That course of action would be just as beneficial to humanity as pursuing genetic cognitive enhancements, and it would do a much better job of fulfilling our ethical obligations.
This is one of the Reader Picks comments:
Mark
Philadelphia May 27
I have mixed feelings about the concept of this article. Surely, private schools confer numerous advantages to their students, who are from wealthy backgrounds and connections to higher education and corporate America.
But, look at Stuyvesant. The super intelligent and successful students are very often from middle class, lower-middle class, and even poor backgrounds. They are often first generation immigrants. They are just smart and hard working and their families care desperately about education.
Some kids are just smart, while others, are just average, or below average. You really think if you went into a school in the South Bronx and donated $1 billion the students would start cranking out perfect SATs?
Ask Zuckerberg how is $50 million donation to Newark public schools went. Darwinism is cruel, but some people aren't just cut out to be good students or white collar professionals.
Much of this has little to do with class and everything to do with drive and innate ability.
I've collected some recent links related to polygenic risk scores below.
Many experts anticipate large scale clinical use of these scores within the next few years. Research progress has been very rapid -- it will be interesting to see how long it takes for these breakthroughs to be applied in health care. Graph below shows number of papers per year.
Guardian article on UK Health Minister's proposal for widespread genetic testing:
"The latest predictive tests for a range of common diseases take a different approach: they aggregate the tiny contributions to risk made by hundreds or even thousands of genes to give a personalised score. Because the risk is spread out over many genes, people can end up at the very high-risk end of the spectrum by chance, without having a family history of a particular illness."
"Prof John Bell, a professor of medicine at Oxford university who led a recent government-commissioned review of the life sciences industry, said the approach could have a “quite profound” effect on the ability to manage disease. ... David Spiegelhalter, professor for the public understanding of risk at the University of Cambridge, agrees that genetic tests could allow the NHS to rapidly identify those who may need closer monitoring."
Editorial in New England Journal of Medicine anticipates broad clinical use of polygenic scores:
"It is likely that tailoring decisions about prescribing preventive medicines or screening practices will be the main future use of genetic risk scores. If a PRS adds to existing clinical predictors of risk such as the Framingham Risk Score or the Q index for heart disease, it could be incorporated into preventive care as readily as any other biomarker."
"There seems little doubt that interpretation of these scores will become an accepted part of clinical practice in the future..."
Article in The Conversation by two Australian professors of Public Health (discussion refers to both monogenic and polygenic risks, AFAICT):
"As DNA testing becomes cheaper, it becomes more feasible to screen large numbers of healthy people for their risk of disease."
"We modelled the health and economic benefits of offering population DNA screening in Australia, focusing on young adults aged 18-25 years (about 2.6 million Australians). ... At A$200 per test (which could be realistic in the near future), savings in treatment costs could outweigh screening costs, saving the health-care system money and saving lives."
MyHeritage DNA (T2D, Heart Disease and Breast Cancer) joins 23andMe (T2D) in offering polygenic risk scores using common SNPs in their health reports. FDA regulatory stance allows DTC (Direct To Consumer) reports of this type as long as they are provided as information to be discussed with a physician, and not to diagnose a condition or prescribe care.
Someone tweeted me this photo from a recent conference presentation: increase of AUC (predictive power) with sample size. ~100k cases enough to capture most of common SNP heritability for diseases such as Testicular or Breast Cancer?
Highly recommended interview with MIT professor Tomaso Poggio, which I listened to recently on a plane. IIRC, I largely agreed with his positions except that I'm a bit more optimistic about AGI. I think his estimate for AGI was 100 or 200 years from now, whereas I think by the end of my lifetime is a distinct possibility.
Poggio (trained in theoretical physics) starts by describing the effect that Special Relativity had on him as a kid. It is a striking realization that pure thought experiments of the kind originally formulated by Einstein can have such far-reaching implications. See Physics as a Strange Attractor:
I suspect that Special Relativity, because it is easy to introduce (no mathematics beyond algebra is required), yet deep and beautiful and counterintuitive, stimulates many people of high ability to become interested in physics.
I notice (perhaps unsurprisingly) a lot of similarities in Poggio's views and those of his former student Demis Hassabis of DeepMind.
Tomaso Poggio is a professor at MIT and is the director of the Center for Brains, Minds, and Machines. Cited over 100,000 times, his work has had a profound impact on our understanding of the nature of intelligence, in both biological neural networks and artificial ones. He has been an advisor to many highly-impactful researchers and entrepreneurs in AI, including Demis Hassabis of DeepMind, Amnon Shashua of MobileEye, and Christof Koch of the Allen Institute for Brain Science. This conversation is part of the Artificial Intelligence podcast and the MIT course 6.S099: Artificial General Intelligence. The conversation and lectures are free and open to everyone. Audio podcast version is available on https://lexfridman.com/ai/
Ron Unz is the publisher of the Unz Review, a controversial but widely read alternative media site hosting opinion outside of the mainstream, including from both the far right and the far left. Unz studied theoretical physics at Harvard, Cambridge and Stanford. He founded the software company Wall Street Analytics, acquired by Moody’s in 2006, and was behind the 1998 ballot initiative that ended bilingual education in California.
In mathematics, a manifold is a topological space that locally resembles Euclidean space near each point.
Steve Hsu and Corey Washington have been friends for almost 30 years, and between them hold PhDs in Neuroscience, Philosophy, and Theoretical Physics. Join them for wide ranging and unfiltered conversations with leading writers, scientists, technologists, academics, entrepreneurs, investors, and more.
Steve Hsu is VP for Research and Professor of Theoretical Physics at Michigan State University. He is also a researcher in computational genomics and founder of several Silicon Valley startups, ranging from information security to biotech. Educated at Caltech and Berkeley, he was a Harvard Junior Fellow and held faculty positions at Yale and the University of Oregon before joining MSU.
Corey Washington is Director of Analytics in the Office of Research and Innovation at Michigan State University. He was educated at Amherst College and MIT before receiving a PhD in Philosophy from Stanford and a PhD in a Neuroscience from Columbia. He held faculty positions at the University Washington and the University of Maryland. Prior to MSU, Corey worked as a biotech consultant and is founder of a medical diagnostics startup.
Michigan State University senior vice president Stephen Hsu, a theoretical physicist and the founder of Genomic Prediction, demonstrates how the machine learning revolution, combined with the dramatic fall in the cost of human genome sequencing, is driving a transformation in our relationship with our genes. Stephen and Azeem Azhar explore how the technology works, what predictions can and cannot yet be made (and why), and the ethical challenges created by this technology.
In this podcast, Azeem and Stephen also discuss:
FDA approval of the first genetic treatment for monogenic conditions and the work towards developing treatments for polygenic conditions like diabetes and cancer.
How this technology might exacerbate existing social inequalities or create new ones; is it just an issue of access, or does it go further?
Developing best practice protocols for governance and regulation of genomic technologies.
In the interview I mention that the number of genomics papers on polygenic risk scores has exploded just in the last year or so:
Several people asked me to comment on this paper, which appeared recently on biorxiv. It seems to be an update of earlier (simulation) analyses by Gwern [16] and Shulman and Bostrom [15] (cited in the paper) on potential gains from embryo selection using quantitative trait predictors (e.g., height, cognitive ability). In the paper the authors analyze real families using actual genetic and phenotype data.
The main limitations given current technology are the number of embryos available from which to select, and the accuracy of the polygenic predictors. The latter will almost certainly improve significantly for some traits in the near future, and for all traits eventually. The number of embryos available for selection may also increase if new methods allow oocytes (eggs) to be produced using stem cell technology (already demonstrated in mice; video).
Genome-wide association studies have led to the development of polygenic score (PS) predictors that explain increasing proportions of the variance in human complex traits. In parallel, progress in preimplantation genetic testing now allows genome-wide genotyping of embryos generated via in vitro fertilization (IVF). Jointly, these developments suggest the possibility of screening embryos for polygenic traits such as height or cognitive function. There are clear ethical, legal, and societal concerns regarding such a procedure, but these cannot be properly discussed in the absence of data on the expected outcomes of screening. Here, we use theory, simulations, and real data to evaluate the potential gain of PS-based embryo selection, defined as the expected difference in trait value between the top-scoring embryo and an average, unselected embryo. We observe that the gain increases very slowly with the number of embryos, but more rapidly with increased variance explained by the PS. Given currently available polygenic predictors and typical IVF yields, the average gain due to selection would be ≈2.5cm if selecting for height, and ≈2.5 IQ (intelligence quotient) points if selecting for cognitive function. These mean values are accompanied by wide confidence intervals; in real data drawn from nuclear families with up to 20 offspring each, we observe that the offspring with the highest PS for height was the tallest only in 25% of the families. We discuss prospects and limitations of PS-based embryo selection for the foreseeable future.
The authors of the paper seem to define "utility" in terms of expected gain in trait value. However, there is also utility in eliminating very negative outcomes, even if they have small probability. This does not shift the average very much but may still be highly desirable. For example, the odds of my house being destroyed by fire or earthquake in the next decade are small, but the outcome is negative enough that I will act to insure against it. If there is a 1% chance of a $100k house being destroyed, the expected loss is only $1k over the period. But without insurance the outcome might be devastating to a family.
One can compare this to screening for Down Syndrome, which has an incidence of roughly 1% (depending on parental age, etc.) but very serious consequences (see podcast discussion below).
At Genomic Prediction we have focused on screening against disease risk rather than on selection for quantitative traits, for both ethical and practical reasons. Even noisy (imperfect) predictors allow the identification of individuals who are high risk outliers -- e.g., are 5x times more likely to get the disease than a typical person.
When considering disease risk the key metric is not the polygenic score itself, because odds ratios are nonlinear functions of the score (or score percentile). For example (note, this is entirely hypothetical), consider 3 embryos with disease risk percentile scores (e.g., Breast Cancer, Type 1 Diabetes, Atrial Fibrillation, Coronary Artery Disease) given by column:
#1 33 57 64 51
#2 62 39 36 49
#3 26 22 52 99.5
Even though the linear averages of the four risk percentiles for all three embryos are similar (contrived to be near 50), embryo #3 has unusually high risk for one condition (e.g., Coronary Artery Disease) and embryos #1 and #2 might be preferred.
Quantifying the utility to the family from this kind of screening is much more complex than for quantitative traits such as height or cognitive ability.
For more on ethical questions related to genetic engineering and embryo selection, see this podcast discussion with Sam Kerstein, chair of the philosophy department at the University of Maryland.
Janelia is the research campus of the Howard Hughes Medical Institute (HHMI), located near Washington DC. It is reputed to be heaven on earth for scientists :-)
I'll be visiting there next week (see title and abstract below). If you're at Janelia and want to meet with me there is still a little space on my schedule. Or just come to the talk and try to grab me afterward.
My talk is Tuesday May 7 12:30 – 1:30.
Genomic Prediction of Complex Traits and Disease Risks via AI and Large Genomic Datasets
Abstract: The talk is divided into two parts. The first gives an overview of the rapidly advancing area of genomic prediction of disease risks using polygenic scores. We can now identify risk outliers (e.g., with 5 or 10 times normal risk) for about 20 common disease conditions, ranging from diabetes to heart diseases to breast cancer, using inexpensive SNP genotypes (i.e., as offered by 23andMe). We can also predict some complex quantitative traits (e.g., adult height with accuracy of few cm, using ~20k SNPs). I discuss application of these results in precision medicine as well as embryo selection in IVF, and give some details concerning genetic architecture. The second part covers the AI/ML used to build these predictors, with an emphasis on "sparse learning" and phase transitions in high dimensional statistics.
Corey and Steve speak with Samuel Kerstein, Professor of Philosophy and expert in Medical Ethics at the University of Maryland. They discuss the ethics of genome engineering and preimplantation embryo selection, and the inequality and narrowing of human diversity that might result from widespread adoption of these technologies. Among the topics covered: Why genome engineering at this time is immoral. Should we always pick the healthiest embryo? In the future will parents have a moral obligation to engineer their children? Will there be an arms race between countries to engineer their populations? Is Star Trek’s Khan a more advanced person (Steve) or just another smart psychopath (Sam) or both?
In mathematics, a manifold is a topological space that locally resembles Euclidean space near each point.
Steve Hsu and Corey Washington have been friends for almost 30 years, and between them hold PhDs in Neuroscience, Philosophy, and Theoretical Physics. Join them for wide ranging and unfiltered conversations with leading writers, scientists, technologists, academics, entrepreneurs, investors, and more.
Steve Hsu is VP for Research and Professor of Theoretical Physics at Michigan State University. He is also a researcher in computational genomics and founder of several Silicon Valley startups, ranging from information security to biotech. Educated at Caltech and Berkeley, he was a Harvard Junior Fellow and held faculty positions at Yale and the University of Oregon before joining MSU.
Corey Washington is Director of Analytics in the Office of Research and Innovation at Michigan State University. He was educated at Amherst College and MIT before receiving a PhD in Philosophy from Stanford and a PhD in a Neuroscience from Columbia. He held faculty positions at the University Washington and the University of Maryland. Prior to MSU, Corey worked as a biotech consultant and is founder of a medical diagnostics startup.
In a high corner office, overlooking Cambridge and the Harvard campus.
How big a role is deep learning playing right now in building genomic predictors?
So far, not a big one. Other ML methods perform roughly on par with DL. The additive component of variance is largest, and we have compressed sensing theorems showing near-optimal performance for capturing it. There are nonlinear effects, and eventually DL will likely be useful for learning multi-loci features. But at the moment everything is limited by statistical power, and nonlinear features are even harder to detect than additive ones. ...
The bottom line is that with enough statistical power predictors will capture the expected heritability for most traits. Are people in your field ready for this?
Some are, but for others it will be very difficult.
Conference on AI and Genomics / Precision Medicine (Boston).
I enjoyed your talk. I work for [leading AgBio company], but my PhD is in Applied Math. We've been computing Net Merit for bulls using SNPs for a long time. The human genetics people have been lagging...
Caught up now, though. And first derivative (sample size growth rate) is much larger...
Yes. It's funny because sperm is priced by Net Merit and when we or USDA revise models some farmers or breeders get very angry because the value of their bull can change a lot!
A Harvard Square restaurant.
I last saw Roman at the Fellows spring dinner, many years ago. I was back from Yale to see friends. He was drinking, with serious intent. He told me about working with Wilson at Cornell. He also told me an old story about Jeffrey and the Higgs mechanism. Jeffrey almost had it, soon after his work on the Goldstone boson. But Sidney talked him out of it -- something to the effect of "if you can only make sense of it in unitary gauge, it must be an artifact" ... Afterwards, at MIT they would say When push comes to shove, Sidney is wrong. ...
Genomics is in the details now. Lots of work to be done, but conceptually it's clear what to do. I wouldn't say that about AGI. There are still important conceptual breakthroughs that need to be made.
The Dunster House courtyard, overlooking the Charles.
We used to live here, can you let us in to look around?
I remember it all -- the long meals, the tutors, the students, the concerts in the library. Yo Yo Ma and Owen playing together.
A special time, at least for us. But long vanished except in memory.
Wheeler used to say that the past only exists as memory records.
Not very covariant! Why not a single four-manifold that exists all at once?
The Ritz-Carlton.
Flying private is like crack. Once you do it, you can't go back...
It's not like that. They never give you a number. They just tell you that the field house is undergoing a renovation and there's a naming opportunity. Then your kid is on the right list. They've been doing this for a hundred years...
Card had to do the analysis that way. Harvard was paying him...
I went to the session on VC for newbies. Now I realize "valuation" is just BS... Now you see how it really works...
Then Bobby says "What's an LP? I wanna be an LP because you gotta keep them happy."
Let me guess, you want a dataset with a million genomes and FICO scores?
I've helped US companies come to China for 20+ years. At first it was rough. Now if I'm back in the states for a while and return, Shenzhen seems like the Future. The dynamism is here.
To most of Eurasia it just looks like two competing hegemons. Both systems have their pluses and minuses, but it's not an existential problem...
Sure, Huawei is a big threat because they won't put in backdoors for the NSA. Who was tapping Merkel's cellphone? It was us...
Humans are just smart enough to create an AGI, but perhaps not smart enough to create a safe one.
Maybe we should make humans smarter first, so there is a better chance that our successors will look fondly on us. Genetically engineered super-geniuses might have a better chance at implementing Asimov's Laws of Robotics.
Babbage Podcast by The Economist. (21 minutes. Sorry, can't embed the player.)
Economist Senior Editor Kenneth Cukier takes a look at what it means to be human. He speaks to leading scientists, doctors and philosophers to ask if ethics and regulations are able to keep up with the technology of genetic engineering.
If I understand correctly the issue is how to realize something like backprop when most of the information flow is feed-forward (as in real neurons). How do you transport weights "non-locally"? The L2 optimization studied here doesn't actually transport weights. Rather, the optimized solution realizes the same set of weights in two places...
Center for Brains, Minds and Machines (CBMM)
Published on Apr 3, 2019
Speaker: Dr. Jon Bloom, Broad Institute
Abstract: When trained to minimize reconstruction error, a linear autoencoder (LAE) learns the subspace spanned by the top principal directions but cannot learn the principal directions themselves. In this talk, I'll explain how this observation became the focus of a project on representation learning of neurons using single-cell RNA data. I'll then share how this focus led us to a satisfying conversation between numerical analysis, algebraic topology, random matrix theory, deep learning, and computational neuroscience. We'll see that an L2-regularized LAE learns the principal directions as the left singular vectors of the decoder, providing a simple and scalable PCA algorithm related to Oja's rule. We'll use the lens of Morse theory to smoothly parameterize all LAE critical manifolds and the gradient trajectories between them; and see how algebra and probability theory provide principled foundations for ensemble learning in deep networks, while suggesting new algorithms. Finally, we'll come full circle to neuroscience via the "weight transport problem" (Grossberg 1987), proving that L2-regularized LAEs are symmetric at all critical points. This theorem provides local learning rules by which maximizing information flow and minimizing energy expenditure give rise to less-biologically-implausible analogues of backproprogation, which we are excited to explore in vivo and in silico. Joint learning with Daniel Kunin, Aleksandrina Goeva, and Cotton Seed.
Keenan Cornelius is a world class blackbelt in jiujitsu. His skill level is sufficiently high that he can verbalize his tactics in real time as he rolls with lower belts.
"He's trying to bump me off mount, so I'm shifting my weight to my left knee to keep my weight off of his hips. But once he gives me an opening I'm going to slide under his lapel to finalize the choke..." Sometimes he is several moves ahead of his opponent!
Roy Dean does something similar, but with narration added in post-production, here. Roy's video is more precise (for one thing, he's not out of breath) but what Keenan is doing is super impressive :-)
Hossenfelder is a Research Associate at the Frankfurt Institute of Advanced Studies. Her research areas include particle physics and quantum gravity. She discusses the current state of theoretical physics, and her recent book Lost in Math: How Beauty Leads Physics Astray.
In mathematics, a manifold is a topological space that locally resembles Euclidean space near each point.
Steve Hsu and Corey Washington have been friends for almost 30 years, and between them hold PhDs in Neuroscience, Philosophy, and Theoretical Physics. Join them for wide ranging and unfiltered conversations with leading writers, scientists, technologists, academics, entrepreneurs, investors, and more.
Steve Hsu is VP for Research and Professor of Theoretical Physics at Michigan State University. He is also a researcher in computational genomics and founder of several Silicon Valley startups, ranging from information security to biotech. Educated at Caltech and Berkeley, he was a Harvard Junior Fellow and held faculty positions at Yale and the University of Oregon before joining MSU.
Corey Washington is Director of Analytics in the Office of Research and Innovation at Michigan State University. He was educated at Amherst College and MIT before receiving a PhD in Philosophy from Stanford and a PhD in a Neuroscience from Columbia. He held faculty positions at the University Washington and the University of Maryland. Prior to MSU, Corey worked as a biotech consultant and is founder of a medical diagnostics startup.
Compute polygenic EA scores for 3000 US high school students of European ancestry. Track individual progress from 9th to 12th grade, focusing on mathematics courses. The students are out-of-sample: not used in training of predictor. In fact, a big portion (over half?) of individuals used in predictor training are not even from the US -- they are from the UK/EU.
Results: predictor captures about as much variance as family background (SES = Social Economic Status). Students with lower polygenic scores are less likely to take advanced math (e.g., Geometry and beyond).
Typical education paths of individuals with, e.g., bottom few percentile polygenic score are radically different from those in the top percentiles, even after controlling for SES. For example, consider only rich kids or kids at superior schools and compare educational trajectory vs polygenic score. Looks like (bottom figure) odds ratio for taking Geometry in 9th grade is about 4-6x higher for top polygenic score kids.
...we address this question using student polygenic scores, which are DNA-based indicators of propensity to succeed in education8. We integrated genetic and official school transcript data from over 3,000 European-ancestry students from U.S. high schools. We used polygenic scores as a molecular tracer to understand how the flow of students through the high school math pipeline differs in socioeconomically advantaged versus disadvantaged schools. Students with higher education polygenic scores were tracked to more advanced math already at the beginning of high school and persisted in math for more years...
...including family-SES and school-SES as covariates attenuated the association between the education-PGS and mathematics tracking in the 9th-grade only by about 20% (attenuated from b = 0.583, SE = .034, to b = 0.461, SE = .036, p < 2 × 10-16, Supplementary Table S3). Note that the association with genetics was roughly comparable in magnitude to the association with familySES...
David Skrbina is a philosopher at the University of Michigan. He and Ted Kaczynski published the book Technological Slavery, which elaborates on the Unabomber manifesto and contains about 100 pages of correspondence between the two which took place over almost a decade. Skrbina discusses his and Kaczynski's views on deep problems of technological society, and whether violent opposition to it is justified.
In mathematics, a manifold is a topological space that locally resembles Euclidean space near each point.
Steve Hsu and Corey Washington have been friends for almost 30 years, and between them hold PhDs in Neuroscience, Philosophy, and Theoretical Physics. Join them for wide ranging and unfiltered conversations with leading writers, scientists, technologists, academics, entrepreneurs, investors, and more.
Steve Hsu is VP for Research and Professor of Theoretical Physics at Michigan State University. He is also a researcher in computational genomics and founder of several Silicon Valley startups, ranging from information security to biotech. Educated at Caltech and Berkeley, he was a Harvard Junior Fellow and held faculty positions at Yale and the University of Oregon before joining MSU.
Corey Washington is Director of Analytics in the Office of Research and Innovation at Michigan State University. He was educated at Amherst College and MIT before receiving a PhD in Philosophy from Stanford and a PhD in a Neuroscience from Columbia. He held faculty positions at the University Washington and the University of Maryland. Prior to MSU, Corey worked as a biotech consultant and is founder of a medical diagnostics startup.
Big chickens! Modern breeds are about four times larger than those raised in the 1950s. I wonder how many population SDs of change that represents? About 40?
In this exclusive interview, Stephen Hsu (Michigan State University and co-founder of Genomic Prediction) discusses the application of polygenic risk scores (PRS) for complex traits in pre-implantation genetic screening. Interview conducted by Julianna LeMieux (GEN).
GEN: What motivated you to start Genomic Prediction?
STEVE HSU: It has a very long history. Laurent Tellier is the CEO and we’ve known each other since 2010. We’d been working on the background science of how to use machine learning to look at lots of genomes and then learn to predict phenotypes from that information.
We were betting on the continuing decline in cost for genotyping, and it paid off because now there are millions of genotypes available for analysis. We’d always thought that one of the best and earliest applications of this would be embryo selection because we can help families have a healthy child.
GEN: How did you first get interested in genomics in general, given your educational background in physics?
HSU: I was interested in genetics and evolution, molecular biology, since I was a kid. I grew up in the ’70s and ’80s and already at that time there was a lot of attention focused on the molecular biology revolution, recombinant DNA. We were always told physics is a very mature subject and biology is the subject of the future and it will just explode eventually with these new molecular techniques.
When I got to college and I took some classes in molecular biology, I realized that a lot of the deep questions—like how do you actually decipher a genome and figure out which pieces of the genetic code have direct consequences in phenotypes or complex traits?—would not be answerable with the technology of that time. So I put it aside and did theoretical physics, but got re-interested around the time I met Laurent. I became aware of the super exponential cost curve for genotyping, sequencing in particular. I realized, if this continues for another ten years or so, we’re going to be able to answer all these interesting questions I’ve been thinking about since I was a kid.
...
GEN: How do you generate a polygenic risk score for different diseases? Of the eight diseases listed on the Genomic Prediction website, are those diseases that your lab has basically generated that data for?
HSU: Many of them were produced by my research group, but the current best-performing breast cancer predictor actually comes from a large international consortium that works on breast cancer…
We use the same data that people would use for GWAS [genome-wide association studies]. For example, we might have 200,000 controls and 20 or 30,000 cases of people in their 50s and 60s who are old enough that they would have been diagnosed for diabetes (or something) if they had it. The algorithm knows which ones are the cases and which are the controls, and it also has about 1 million SNPs from each person, typically what you get from an Affymetrix or an Illumina array.
It is a learning algorithm that tries to tune its internal model so that it best predicts whether someone is actually a case or a control. There’s a bunch of fancy math involved in this—a high-dimensional optimization. You are basically finding the model that best predicts the data.
It is different from GWAS because GWAS is very simple—you look at a particular gene or SNP and you say is there statistical evidence that this particular SNP is associated with whether you have diabetes? You get a yes/no answer. If the P value is significant enough then you say we found a hit.
That problem mathematically is very different from the problem we solved. We are actually doing an optimization in a million-dimensional space to find simultaneously all the SNPs that should be activated in our predictor. This is all in the technical weeds but it is just different mathematics…
We think we can actually predict risk by doing this high-dimensional optimization. Initially, people just thought we were crazy. We wrote theoretical papers predicting how much data would you need to be able to accurately predict height or something like that. ... [ AND THOSE PREDICTIONS WERE CORRECT ... ]
This 54-minute documentary traces the writer James Salter's lifelong love affair with France, unforgettably expressed in his 1967 masterpiece, A Sport and a Pastime. The film captures the great purity of Salter's prose and the essence of his power to evoke the erotic. Salter's own reflections on his writing and life offer rich insights for reader and writer alike.
If you enjoy it you may also like this history of The Paris Review, also free on Prime video.
The excerpt below is from his 1993 interview for Paris Review's The Art of Fiction.
INTERVIEWER
When A Sport and a Pastime came out you were hailed as “celebrating the rites of erotic innovation” and yet also criticized for portraying such “vigorous ‘love’ scenes.” What did you think of all that?
SALTER
The eroticism is the heart and substance of the book. That seems obvious. I meant it to be, to use a word of Lorca’s, “lubricious” but pure, to describe things that were unspeakable in one sense, but at the same time, irresistible. Having traveled, I also was aware that voyages are, in a large sense, a search for, a journey toward love. A voyage without that is rather sterile. Perhaps this is a masculine view, but I think not entirely. The idea is of a life that combines sex and architecture—I suppose that’s what the book is, but that doesn’t explain it. It’s more or less a guide to what life might be, an ideal.
INTERVIEWER
People seem to have different opinions of what the book is about.
SALTER
I listen occasionally to people explaining the book to me. Every few years there’s an inquiry from a producer who would like to make a movie of it. I’ve turned the offers down because it seems to me ridiculous to try and film it. To my mind the book is obvious. I don’t see the ambiguity, but there again, you don’t know precisely what you are writing. Besides, how can you explain your own work? It’s vanity. To me it seems you can understand the book, if there’s been any doubt, by reading the final paragraph:
As for Anne-Marie, she lives in Troyes now, or did. She is married. I suppose there are children. They walk together on Sundays, the sunlight falling upon them. They visit friends, talk, go home in the evening, deep in the life we all agree is so greatly to be desired.
That paragraph, the final sentence, is written in irony, but perhaps not read that way. If you don’t see the irony, then the book is naturally going to have a different meaning for you.
INTERVIEWER
It has been said that Dean’s desire for Anne-Marie is also a desire for the “real” France. It’s a linked passion.
SALTER
France is beautiful, but his desire is definitely for the girl herself. Of course she is an embodiment. Even when you recognize what she is, she evokes things. But she would be desirable to him even if she didn’t.
INTERVIEWER
There’s a postmodern side to the book. The narrator indicates that he’s inventing Dean and Anne-Marie out of his own inadequacies.
SALTER
That’s just camouflage.
INTERVIEWER
What do you mean?
SALTER
This book would have been difficult to write in the first person—that is to say if it were Dean’s voice. It would be quite interesting written from Anne-Marie’s voice, but I wouldn’t know how to attempt that. On the other hand, if it were in the third person, the historic third, so to speak, it would be a little disturbing because of the explicitness, the sexual descriptions. The question was how to paint this, more or less. I don’t recall how it came to me, but the idea of having a third person describe it, somebody who is really not an important part of the book but merely serving as an intermediary between the book and the reader, was perhaps the thing that was going to make it possible; and consequently, I did that. I don’t know who this narrator is. You could say it’s me; well, possibly. But truly, there is no such person. He’s a device. He’s like the figure in black that moves the furniture in a play, so to speak, essential, but not part of the action.
...
INTERVIEWER
What do you think is the ultimate impulse to write?
SALTER
To write? Because all this is going to vanish. The only thing left will be the prose and poems, the books, what is written down. Man was very fortunate to have invented the book. Without it the past would completely vanish, and we would be left with nothing, we would be naked on earth.
