search

mattapow / dodonaphy

hyperbolic embeddings for variational approximation of phylogenetic posteriors
3 stars 1 forks source link

Alignment and tree in pytorch #10

Closed mattapow closed 2 years ago

mattapow commented 2 years ago

The work in End-to-end learning of multiple sequence alignments with differentiable Smith-Waterman 2021 formed a differentiable alignment in pytorch. Could we combing the alignment and tree building process into one differentiable framework for optimisation? This is all based of _softsort type algorithms.

Q: What's the selling point other than say 'look what we can do'? Inter-dependence of MSA and tree: could phylogenetic analysis biased if the MSA is wrong? Give gradient feedback to MSA from tree.

mattapow commented 2 years ago

Probably no practical value as alignment uncertainty is generally low, whereas computational cost and complexity of this is not worth the trouble.