- 1) Generate training and test sets for human, fly.
- 2) Reproduce the fly results from the talk in January with the new code.
- 3) Check accuracy on human when trained on human and compare to fly (training & prediction).
- 4) Compute cross-training accuracy human version on fly test set and fly version on human test set.
- 5) Train wth a mixed human and fly training set and evaluate on human and fly test set.
- 6) Conditional on relative success: Compile set of labeled non-animal alignments, tree and training and test sets.
- 7) Optional: compare to a second tool, PAML that computes omega=dN/dS
- a) Implement
--to_phylocsfoption. - b) Implement
clamsa train. - c) Implement model specific training callbacks (drawing rate matrix progression for example)
- d) Implement
recover_modelfunction to load a model from itsTrial ID - e) Implement
clamsa predict:- Present statistics on the datasets
- Evaluate learned models on specified datasets
- f) ClaMSA runs with the same input and similar output as PhyloCSF (drop-in replacement). This also enables the uniform evaluation of all competitors and is a check that the very same data is used for testing and the very same scripts for evaluation.
- Find name soon so figures could be final, e.g.
- ClaMSA
- DiscrEvo: discriminative evolutionary model
- ELEMO: End-to-end Learning of Evolutionary MOdels
- some acronym from: differentiable rate matrix pruning model discriminative ML evolution tree Markov selection
- obtain biological results
With modest effort, we could screen the negatively labeled MSAs that are most confidently predicted as positive for candidates of missing genes. Finding an unannotated human gene is a long shot, but I could search genome-wide with a non-human vertebrate as reference. This could lead to a figure of one - hopefully clear - MSA example of a yet missing gene and some statistics to estimate missing protein-coding genes. This may be a soft requirement for submitting to high-ranking journals as "Bioinformatics" (Oxford, impact factor 5.6).