frugivores ( 25 ). The cross-validated multi-
trait models achieved high accuracy, strongly
outperforming traditional approaches based
on body mass allometries or size matching
fitted by using generalized linear models.
For an independent tenth of the dataset not
used to fit the model, we could predict which
interactions occur with 88% accuracy (table
S1 and fig. S2). Further, the boosted regres-
sion tree models used to characterize seed
dispersal distance and impacts of gut pas-
sage on germination also substantially out-
performed the generalized linear models
(table S1). We show these functional outcomes
for each interaction observed in the network
database (Fig. 1), but these estimates can also
be made for any other plant-animal species
pair, given directly measured or phylogeneti-
cally imputed traits. Such estimates of species
interactions and dispersal can be used to
improve forecasting of biotic responses to
climate change ( 26 ).
Having created predictive models of seed
dispersal interactions on the basis of observedinteractions in existing ecosystems, we next
sought to assess our ability to predict inter-
actions in novel communities. Although the
capacity to do so would advance ecological
forecasting, it is currently not possible to
validate predictions for future novel commu-
nity scenarios. Further, the capacity to develop
predictions based on current observations and
standard, easily measured traits is uncertain
because interaction probabilities may be de-
termined by subtle relationships between other
behavioral, chemical, and morphological factorsSCIENCEscience.org 14 JANUARY 2022¥VOL 375 ISSUE 6577 211
Fig. 2. Validating novel network assembly forecasts using observed
interactions with recently introduced species.(AtoH) Interactions
between plants (rows) and animals (columns) in eight local network
examples, with lowercase letters on the map showing study locations. Large
black squares show observed interactions among native species. Introduced
plant and animal names are shown in blue and native species in black, with
the blue gradient showing the interaction probability predicted by a model
trained on native species interactions only. The small black squares indicate
interactions involving introduced species observed in the field. Points show
study locations for local network studies. Full species names are shown in
table S2. Photos show examples of interactions involving introduced species
predicted to be relatively likely and that were not observed in the local
networks but have been reported elsewhere. (I)Pitangus sulphuratusand
Schefflera arboricola, species in local network f (Photos: S. Andrade),
and (J) showsHemiphaga novaeseelandiaeandCrataegus monogyna, species
in local network h (Photos: S. Attwood).RESEARCH | REPORTS