Science - USA (2022-02-25)

science.org SCIENCE

INSIGHTS | PERSPECTIVES

trees can now be inferred for thousands of
genomes ( 3 – 7 ). Analyses of these trees have
already improved our knowledge of hu-
man demography ( 4 , 5 , 8 , 9 ) and promise
to improve our ability to identify targets
of natural selection ( 5 , 8 , 10 – 12 ). Further,
high-quality ancient genomes can be di-
rectly integrated in the tree, which greatly
increases the certainty of evolutionary
processes by supporting inferences with
data from the past. This is particularly
helpful to study populations that are ex-
tinct or that contributed little to modern
populations, such as Neanderthals and
Denisovans (archaic humans) ( 13 , 14 ), and
tree-based studies are helping to create a
better understanding of the nature of in-
terbreeding between modern and archaic
groups ( 5 , 8 , 9 ).
Unfortunately, the DNA in ancient re-
mains is usually highly degraded and most
ancient genomes are not of sufficient qual-
ity to be fully incorporated into the trees.
Wohns et al. present an approach that aims
to skirt inaccuracies introduced by ancient
DNA by using such ancient genomes only
to help time the emergence of alleles (see
the figure). This allows the use of hundreds
of ancient genomes while limiting the ef-
fects of errors. They generated an impres-
sive, unified human genealogy from >3500
modern and high-quality ancient genomes
from >215 different human populations,
using >3000 additional ancient genomes
to improve inferences from the trees. With
this unified genealogy, key events in his-
tory, such as population size changes,
splits, or migrations, become clearly ap-
parent. They identify well-resolved events,
such as the out-of-Africa migration, and
suggest multiple severe reductions in pop-
ulation size through human history.
Further, Wohns et al. propose a new
method that, applied to the unified gene-
alogy, allows them to estimate ancestral
geographical locations of evolutionary
events. Although there have been other
efforts to place past events geographically
( 15 ), Wohns et al. use a nonparametric
method to estimate the theoretical loca-
tion of inferred human ancestors simply
as the midpoint of the geographic coordi-
nates of their descendants—with these cal-
culations extended backward up the tree
to reach a theoretical common ancestor of
all individuals. Using this method, Wohns
et al. infer an average ancestral location of
all sampled humans in Northeast Africa
by 72,000 years ago and until the oldest

common ancestors of all individuals. This

simple method works well to refine known

ancestral locations and, as sampling im-

proves, it has the potential to identify cur-

rently unknown human movements. More

generally, the trees generated in this study

will undoubtedly prove useful to those

studying human evolution.

Although tree-recording methods rep-

resent an exciting and promising avenue

of harmonizing datasets across time and

space, as demonstrated by Wohns et al.,

they are not without their limitations.

There remains uncertainty in evolutionary

parameters and ancient genomes, and it

is unlikely that the use of low-quality an-

cient DNA will ever be without a degree of

error, except when using rare high-quality

genomes or if inferring missing data be-

comes substantially more accurate. Still,

as genomic datasets continue to grow,

genealogical methods will be increasingly

useful to represent the wealth of genomic

data. Perhaps most importantly, as larger

numbers of modern and ancient genomes

from underrepresented populations be-

come available, our understanding of hu-

man demography, currently still biased to

well-sampled populations, will improve in

both detail and scope.

The power and resolution of tree-record-

ing methods promise to help clarify the

evolutionary history of humans and other

species. It is likely that the most powerful

ways to infer evolutionary history going

forward will have their foundations firmly

set in these methods. j

REFERENCES AND NOTES

1. A. W. Wo h n s et al., Science 375 , eabi8264 (2022).

2. M. D. Rasmussen, M. J. Hubisz, I. Gronau, A. Siepel,
   PLOS Genet. 10 , e1004342 (2014).


3. J. Kelleher, A. M. Etheridge, G. McVean, PLOS Comput.
   Biol. 12 , e1004842 (2016).


4. J. Kelleher et al., Nat. Genet. 51 , 1330 (2019).


5. L. Speidel, M. Forest, S. Shi, S. R. Myers, Nat. Genet. 51 ,
   1321 (2019).


6. K. Harris, Nat. Genet. 51 , 1306 (2019).


7. P. Ralph, K. Thornton, J. Kelleher, Genetics 215 , 779
   (2020).


8. L. Speidel et al., Mol. Biol. Evol. 38 , 3497 (2021).


9. N. K. Schaefer, B. Shapiro, R. E. Green, S c i. A d v. 7 ,
   abc0776 (2021).


10. Y. Field et al., Science 354 , 760 (2016).


11. A. J. Stern, P. R. Wilton, R. Nielsen, PLOS Genet. 15 ,
    e1008384 (2019).


12. A. J. Stern, L. Speidel, N. A. Zaitlen, R. Nielsen, A m. J.
    Hum. Genet. 108 , 219 (2021).


13. M. Meyer et al., Science 338 , 222 (2012).


14. K. Prüfer et al., Nature 505 , 43 (2014).


15. M. M. Osmond, G. Coop, bioRxiv
    10.1101/2021.07.13.452277 (2021).

ACKNOWLEDGMENTS

J.R. is funded by the National Institute for Health Research

Great Ormond Street Hospital Biomedical Research Centre.

A.A. is funded by University College London’s Wellcome

Institutional Strategic Support Fund 3 (204841/Z/16/Z).

10.1126/science.a bo0498

(^1) UCL Genetics Institute, Department of Genetics, Evolution
and Environnment, University College London, London,
UK.^2 Genetics and Genomic Medicine Programme, Great
Ormond Street Institute of Child Health, University College
London, London, UK. Email: [email protected]
MEDICINE
Anticipating
antibiotic
resistance
Machine learning can use
clinical history to lower the
risk of infection recurrence
By J ean-Baptiste Lugagne1,2 and
Mary J. Dunlop1,2

W

hen a patient is diagnosed with a

bacterial infection, clinicians per-

form antibiotic susceptibility tests

in the laboratory and use the re-

sults to prescribe an appropriate

antibiotic. But recurrence rates

are high; for example, ~25% of women with

urinary tract infections (UTIs) develop an-

other infection within 6 months ( 1 , 2 ). On

page 889 of this issue, Stracy et al. ( 3 ) reveal

that recurrent infections are often driven

by a different strain than the original infec-

tion. Although correctly prescribing a sus-

ceptibility-matched antibiotic reduces re-

currence rates overall, it also increases the

chances of developing a resistant infection

caused by a different strain. Thus, treat-

ment is a double-edged sword, in which

susceptibility-matched treatments can pave

the way for resistant strains lurking in the

microbiota. The authors find that this risk

can be minimized with a data-driven ap-

proach that incorporates population-wide

statistics and the patient’s personal history

through a machine learning model to make

antibiotic recommendations.

Key to this study is a large longitudinal

dataset containing information about UTIs

and wound infections, including antibiotic

susceptibility profiles and prescribed anti-

biotics, for male and female patients of all

ages in Israel’s Maccabi Healthcare Services

between 2007 and 2019. In most cases, an-

tibiotic treatment of the initial infection is

effective. However, in 10% of UTIs and 6%

of wound infections, patients experience

early recurrence, returning with another in-

fection within 28 days of the original. Some

of these early recurrences are the result of

erroneous prescription of an antibiotic that

the infection was resistant to, as can hap-

(^1) Department of Biomedical Engineering, Boston University,
Boston, MA, USA.^2 Biological Design Center, Boston
University, Boston, MA, USA. Email: [email protected]
818 25 FEBRUARY 2022 • VOL 375 ISSUE 6583