Paul Gardner

Zentrum für Politische Schönheit @politicalbeauty.bsky.social [ View ] Jun 09, 2024 at 10:54 UTC

11 replies 176 reposts 583 likes

Paul Gardner @ppgardne.bsky.social [ View ] May 21, 2024 at 00:01 UTC

Thanks. Its a negative/critical take, but justified given the results we see IMHO. Hopefully it's useful to a few folk.

0 replies 0 reposts 0 likes

Paul Gardner @ppgardne.bsky.social [ View ] May 20, 2024 at 21:24 UTC

Glad to hear it. Evolutionary selection remains one of the strongest signatures of function (& is required by one definition of function).

0 replies 0 reposts 1 likes

Paul Gardner @ppgardne.bsky.social [ View ] May 20, 2024 at 20:49 UTC

Finally, popularity turns out to be a terrible way to select a software tool. We have once again found that popular tools can be both inaccurate and slow on realistic, independent datasets. #citationmetrics 10/n

0 replies 2 reposts 6 likes

Paul Gardner @ppgardne.bsky.social [ View ] May 20, 2024 at 20:48 UTC

It is important to match confounding variables like homology, length & C+G in your +ve & -ve sets. Also, new genomes/transcriptomes have sequence & assembly error, this should be modelled too! #MachineLearning 9/n

1 replies 0 reposts 0 likes

Paul Gardner @ppgardne.bsky.social [ View ] May 20, 2024 at 20:47 UTC

We emphasize the importance of using realistic, diverse datasets for tool training and testing to avoid biases and improve tool reliability. #DataScience 8/n

1 replies 0 reposts 1 likes

Paul Gardner @ppgardne.bsky.social [ View ] May 20, 2024 at 20:46 UTC

Timing data showed that modern tools are not only often inaccurate but also computationally inefficient. Some widely used tools are both slow and inaccurate! #ComputationalBiology 7/n

1 replies 0 reposts 3 likes

Paul Gardner @ppgardne.bsky.social [ View ] May 20, 2024 at 20:45 UTC

We note a large discrepancy between self-reported tool accuracies and our independent evaluations, suggesting a need for more rigorous validation in the field.  6/n

1 replies 0 reposts 2 likes

Paul Gardner @ppgardne.bsky.social [ View ] May 20, 2024 at 20:44 UTC

The tools using evolutionary conservation (alignment inputs), like RNAcode and PhyloCSF, had much higher accuracy over single-sequence methods in identifying protein-coding regions. #EvolutionaryBiology 5/n

1 replies 0 reposts 1 likes

Paul Gardner @ppgardne.bsky.social [ View ] May 20, 2024 at 20:43 UTC

The results show that only three of nine popular tools significantly outperform a naive scoring scheme (longest stop-free subsequence). The underperforming tools are widely used by the lncRNA community. 4/n

1 replies 1 reposts 3 likes

Paul Gardner @ppgardne.bsky.social [ View ] May 20, 2024 at 20:39 UTC

To test this further we selected control datasets excluding any previous training dataset and including coding exons as a positive set with length-matched intergenic and shuffled sequences as negative sets. 3/n

1 replies 0 reposts 0 likes

Paul Gardner @ppgardne.bsky.social [ View ] May 20, 2024 at 20:38 UTC

After attempting to annotate several plant transcriptomes we noticed some discrepancies between several popular de novo protein-coding annotation tools, and homology-based results. 2/n

1 replies 1 reposts 2 likes

Paul Gardner @ppgardne.bsky.social [ View ] May 20, 2024 at 20:37 UTC

I’m happy to share our preprint  
“Flawed machine-learning confounds coding sequence annotation“
doi.org/10.1101/2024... 
A collaboration with Otago and Plant & Food scientists #Genomics #Bioinformatics #ncRNA #lncRNA #protein 1/n

3 replies 3 reposts 8 likes