PhyloCNN: Improving tree representation and neural network architecture for deep learning from trees in phylodynamics and diversification studies

Phylodynamics and diversification studies using complex evolutionary
models can be challenging, especially with traditional likelihood-based
approaches. As an alternative, likelihood-free simulation-based approaches
have been proposed due to their ability to incorporate complex models and
scenarios. Here, we propose a new simulation-based deep learning (DL)
method capable of selecting birth-death models and accurately estimating
their parameters in both phylodynamics and diversification studies. We use
a convolutional approach, where trees are encoded using the neighborhood
of all nodes and leaves of the input phylogeny. We also developed a
dedicated neural network architecture called PhyloCNN. Using simulations,
we compared the accuracy of PhyloCNN when using a variable number of
neighbors to describe the local context of nodes and leaves. The number of
neighbors had a greater impact when considering smaller training sets,
with a broader context showing higher accuracy, especially for complex
evolutionary models. Compared to other recently developed DL approaches,
PhyloCNN showed higher or similar accuracies for all parameters when used
with training sets one or two orders of magnitude smaller (10,000 to
100,000 simulated training trees, instead of millions). PhyloCNN also
compared favorably with state-of-the-art likelihood-based methods. We
applied PhyloCNN with compelling results to two real-world phylodynamics
and diversification datasets, related to HIV superspreaders in Zurich and
to primates and their ecological role as seed dispersers. The high
accuracy and computational efficiency of PhyloCNN open new possibilities
for phylodynamics and diversification studies that need to account for
idiosyncratic phylogenetic histories with specific parameter spaces and
sampling scenarios.