Models have always been central to inferring molecular evolution and to
reconstructing phylogenetic trees. Their use typically involves the
development of a mechanistic framework reflecting our understanding of the
underlying biological processes, such as nucleotide substitutions, and the
estimation of model parameters by maximum likelihood or Bayesian
inference. However, deriving and optimizing the likelihood of the data is
not always possible under complex evolutionary scenarios or even tractable
for large datasets, often leading to unrealistic simplifying assumptions
in the fitted models. To overcome this issue, we coupled stochastic
simulations of genome evolution with a new supervised deep learning model
to infer key parameters of molecular evolution. Our model is designed to
directly analyze multiple sequence alignments and estimate per-site
evolutionary rates and divergence, without requiring a known phylogenetic
tree. The accuracy of our predictions matched that of likelihood-based
phylogenetic inference, when rate heterogeneity followed a simple gamma
distribution, but it strongly exceeded it under more complex patterns of
rate variation, such as codon models. Our approach is highly scalable and
can be efficiently applied to genomic data, as we showed on a dataset of
26 million nucleotides from the clownfish clade. Our simulations also
showed that the integration of per-site rates obtained by deep learning
within a Bayesian framework led to significantly more accurate
phylogenetic inference, particularly with respect to the estimated branch
lengths. We thus propose that future advancements in phylogenetic analysis
will benefit from a semi-supervised learning approach that combines
deep-learning estimation of substitution rates, which allows for more
flexible models of rate variation, and probabilistic inference of the
phylogenetic tree, which guarantees interpretability and a rigorous
assessment of statistical support.