Approach to machine learning for extraction of real-world data variables from electronic health records.

Background: As artificial intelligence (AI) continues to advance with breakthroughs in natural language processing (NLP) and machine learning (ML), such as the development of models like OpenAI's ChatGPT, new opportunities are emerging for efficient curation of electronic health records (EHR) into real-world data (RWD) for evidence generation in oncology. Our objective is to describe the research and development of industry methods to promote transparency and explainability. Methods: We applied NLP with ML techniques to train, validate, and test the extraction of information from unstructured documents (e.g., clinician notes, radiology reports, lab reports, etc.) to output a set of structured variables required for RWD analysis. This research used a nationwide electronic health record (EHR)-derived database. Models were selected based on performance. Variables curated with an approach using ML extraction are those where the value is determined solely based on an ML model (i.e. not confirmed by abstraction), which identifies key information from visit notes and documents. These models do not predict future events or infer missing information. Results: We developed an approach using NLP and ML for extraction of clinically meaningful information from unstructured EHR documents and found high performance of output variables compared with variables curated by manually abstracted data. These extraction methods resulted in research-ready variables including initial cancer diagnosis with date, advanced/metastatic diagnosis with date, disease stage, histology, smoking status, surgery status with date, biomarker test results with dates, and oral treatments with dates. Conclusion: NLP and ML enable the extraction of retrospective clinical data in EHR with speed and scalability to help researchers learn from the experience of every person with cancer.

Broad-scale variation in human genetic diversity levels is predicted by purifying selection on coding and non-coding elements.

Analyses of genetic variation in many taxa have established that neutral genetic diversity is shaped by natural selection at linked sites. Whether the mode of selection is primarily the fixation of strongly beneficial alleles (selective sweeps) or purifying selection on deleterious mutations (background selection) remains unknown, however. We address this question in humans by fitting a model of the joint effects of selective sweeps and background selection to autosomal polymorphism data from the 1000 Genomes Project. After controlling for variation in mutation rates along the genome, a model of background selection alone explains ~60% of the variance in diversity levels at the megabase scale. Adding the effects of selective sweeps driven by adaptive substitutions to the model does not improve the fit, and when both modes of selection are considered jointly, selective sweeps are estimated to have had little or no effect on linked neutral diversity. The regions under purifying selection are best predicted by phylogenetic conservation, with ~80% of the deleterious mutations affecting neutral diversity occurring in non-exonic regions. Thus, background selection is the dominant mode of linked selection in humans, with marked effects on diversity levels throughout autosomes.

The impact of genetic modifiers on variation in germline mutation rates within and among human populations.

Mutation rates and spectra differ among human populations. Here, we examine whether this variation could be explained by evolution at mutation modifiers. To this end, we consider genetic modifier sites at which mutations, "mutator alleles," increase genome-wide mutation rates and model their evolution under purifying selection due to the additional deleterious mutations that they cause, genetic drift, and demographic processes. We solve the model analytically for a constant population size and characterize how evolution at modifier sites impacts variation in mutation rates within and among populations. We then use simulations to study the effects of modifier sites under a plausible demographic model for Africans and Europeans. When comparing populations that evolve independently, weakly selected modifier sites (2Nes≈1), which evolve slowly, contribute the most to variation in mutation rates. In contrast, when populations recently split from a common ancestral population, strongly selected modifier sites (2Nes≫1), which evolve rapidly, contribute the most to variation between them. Moreover, a modest number of modifier sites (e.g. 10 per mutation type in the standard classification into 96 types) subject to moderate to strong selection (2Nes>1) could account for the variation in mutation rates observed among human populations. If such modifier sites indeed underlie differences among populations, they should also cause variation in mutation rates within populations and their effects should be detectable in pedigree studies.

Changes in life history and population size can explain the relative neutral diversity levels on X and autosomes in extant human populations.

In human populations, the relative levels of neutral diversity on the X and autosomes differ markedly from each other and from the naïve theoretical expectation of 3/4. Here we propose an explanation for these differences based on new theory about the effects of sex-specific life history and given pedigree-based estimates of the dependence of human mutation rates on sex and age. We demonstrate that life history effects, particularly longer generation times in males than in females, are expected to have had multiple effects on human X-to-autosome (X:A) diversity ratios, as a result of male-biased mutation rates, the equilibrium X:A ratio of effective population sizes, and the differential responses to changes in population size. We also show that the standard approach of using divergence between species to correct for male mutation bias results in biased estimates of X:A effective population size ratios. We obtain alternative estimates using pedigree-based estimates of the male mutation bias, which reveal that X:A ratios of effective population sizes are considerably greater than previously appreciated. Finally, we find that the joint effects of historical changes in life history and population size can explain the observed X:A diversity ratios in extant human populations. Our results suggest that ancestral human populations were highly polygynous, that non-African populations experienced a substantial reduction in polygyny and/or increase in the male-to-female ratio of generation times around the Out-of-Africa bottleneck, and that current diversity levels were affected by fairly recent changes in sex-specific life history.

Life History Effects on Neutral Diversity Levels of Autosomes and Sex Chromosomes.

Understanding the determinants of neutral diversity patterns on autosomes and sex chromosomes provides a bedrock for the interpretation of population genetic data; in particular, differences between the two informs our understanding of sex-specific demographic and mutation processes. While sex-specific age-structure and variation in reproductive success have long been known to affect neutral diversity, theoretical descriptions of these effects were complicated and lacking in generality, stymying attempts to relate diversity patterns of species with their life history. Here, we derive general yet simple expressions for these effects. In particular, we show that life history effects on X-to-autosome ratios of pairwise diversity levels (X:A diversity ratios) depend only on the male-to-female ratios of mutation rates, generation times, and reproductive variances. Our results reveal that changing the male-to-female ratio of generation times has opposite effects on X:A ratios of diversity and divergence. They also explain how sex-specific life histories modulate the response of X:A diversity ratios to changes in population size. More generally, they clarify that sex-specific life history-generation times in particular-should have marked effects on X:A diversity ratios in many taxa and enable further investigation of these effects.

Overlooked roles of DNA damage and maternal age in generating human germline mutations.

The textbook view that most germline mutations in mammals arise from replication errors is indirectly supported by the fact that there are both more mutations and more cell divisions in the male than in the female germline. When analyzing large de novo mutation datasets in humans, we find multiple lines of evidence that call that view into question. Notably, despite the drastic increase in the ratio of male to female germ cell divisions after the onset of spermatogenesis, even young fathers contribute three times more mutations than young mothers, and this ratio barely increases with parental age. This surprising finding points to a substantial contribution of damage-induced mutations. Indeed, C-to-G transversions and CpG transitions, which together constitute over one-fourth of all base substitution mutations, show genomic distributions and sex-specific age dependencies indicative of double-strand break repair and methylation-associated damage, respectively. Moreover, we find evidence that maternal age at conception influences the mutation rate both because of the accumulation of damage in oocytes and potentially through an influence on the number of postzygotic mutations in the embryo. These findings reveal underappreciated roles of DNA damage and maternal age in the genesis of human germline mutations.

Life history effects on the molecular clock of autosomes and sex chromosomes.

One of the foundational results in molecular evolution is that the rate at which neutral substitutions accumulate on a lineage equals the rate at which mutations arise. Traits that affect rates of mutation therefore also affect the phylogenetic "molecular clock." We consider the effects of sex-specific generation times and mutation rates in species with two sexes. In particular, we focus on the effects that the age of onset of male puberty and rates of spermatogenesis have likely had in hominids (great apes), considering a model that approximates features of the mutational process in mammals, birds, and some other vertebrates. As we show, this model can account for a number of seemingly disparate observations: notably, the puzzlingly low X-to-autosome ratios of substitution rates in humans and chimpanzees and differences in rates of autosomal substitutions among hominine lineages (i.e., humans, chimpanzees, and gorillas). The model further suggests how to translate pedigree-based estimates of human mutation rates into split times among extant hominoids (apes), given sex-specific life histories. In so doing, it largely bridges the gap reported between estimates of split times based on fossil and molecular evidence, in particular suggesting that the human-chimpanzee split may have occurred as recently as 6.6 Mya. The model also implies that the "generation time effect" should be stronger in short-lived species, explaining why the generation time has a major influence on yearly substitution rates in mammals but only a subtle one in human pedigrees.