Is it redcurrant? Fond memories of my dacha where we were collecting it in 10-liter buckets

Finally: nucleosome positioning as highlighting with a marker in a book. Just as highlighted words or sentences stand out differently, where nucleosomes sit can make parts of the genome easier or harder to read. This analogy took me a while to come up with! 4/4

Next: DNA methylation as accents on letters. In languages like French, an accent can change how a letter or word is read; methylation marks can change how cells read DNA, without changing the underlying sequence. Has anyone seen this analogy used before? 3/4

First: the genome as a book. DNA sequence is the text, written with four letters - A, T, G and C. This is a standard analogy, but it gives students a clear starting point before introducing gene regulation. 2/4

It's important to distinguish, what this process is not:
- not asking AI to draft a text for you
- not asking it to fix your text for you
- not giving up your own style, even if imperfect
- not agreeing blindly with all its suggestions

9/9 The model and ranking criteria need to be transparent. If multiple AI-based evaluation systems emerge, which seems likely, their assumptions, limitations, and outputs should be easy to compare and interpret. At the same time, no AI model should be fully autonomous without human supervision, IMO

8/9 The cohort of experts providing feedback for model training probably needs to be diverse and balanced. That also means there should be broad agreement on its composition. One possible approach could be to include, for example, all scientists with a PhD, or the whole current peer-reviewer pool.

7/9 A score based on an AI model trained with feedback from many human experts is likely more stable than a score created by one person. But how many experts are enough? My guess is: having just several experts is not enough. For most fields, likely thousands would be needed.

6/9 On the technical side, no single score is perfect, and one can envision many different AI scores. You can already experiment with ChatGPT to define a score (I tried it on some of my own papers). Scaling this to many manuscripts would need a bit more programming, but it is already doable.

5/9 If we post preprints knowing they may be ranked, it becomes a different game from posting preprints simply to share quick, often important, results that may not yet have matured into a solid story. This may have implications for open science, not necessarily positive ones.

4/9 My sense is that all research that is out there in the open can already be scrutinised by anyone, with or without AI. We cannot really “consent” to our work being ranked. Any openly available work can potentially be evaluated for different purposes.

3/9 This could have major financial and policy implications. For example, the UK’s national research evaluation exercise (REF 2021) cost an estimated £471 million, most of it borne by universities. AI assistance could make parts of this process cheaper.

2/9 It seems to me that Pandora’s box is open and cannot really be closed, potentially leading to many unintended consequences. AI can, conceptually, assist in evaluating research, e.g. in the form of papers or preprints. You can already ask ChatGPT to do this, or develop a more specialist model.