Chen, Hu, Sun, Tan, Wang, Yu, Zong, Hong, Xiao, Shen, King and Li (2022) — ARXIV-2204.00300, read as NOTE-252.
Standing
The first RNA foundation model, and the record’s entry point to biological
sequence modelling. It is the RNA half of a recipe LIT-505
established for proteins: mask tokens, train on the whole unannotated
database, attach a small head per task. The extends: relation is not a
family resemblance — the paper cites it as reference 66 and takes its
downstream ResNet32 from it verbatim.
Held for what it measures, not for what it recommends. The recommendation — pretrain self-supervised on the unlabelled pool when labels are scarce — is one this record already carries implicitly across every practice that touches pretraining, and this paper adds an instance in a new domain rather than evidence about the claim. No practice is filed from it, and NOTE-252 gives the reason in numbers.
The paper is honest where it matters most. Its own Discussion says “the improvement brought by RNA-FM in the functional tasks seems more slight compared with the gain in the structural tasks” and offers a hypothesis for why. That sentence is worth more than the abstract, and it is not in the abstract.