The day in conversation for 2026-11-01

The discussion: Ines: I was thinking about that ESM2 comparison—it’s compelling that it maintains competitiveness when we only have scarce functional labels; I wonder if that performance generalizes across entirely different protein families, or is it just exploiting some shared latent space bias. Marcus: What concerns me is the dependency on those embeddings being sufficiently rich; if the underlying sequence representation misses a critical evolutionary signal, then the high accuracy reported on those limited supervision datasets might be an artifact of domain overlap rather than true predictive power for novel structures. Yuki: That leads into GenoBERT, which seems to tackle a different kind of scarcity—missing genotype data. The claim is that using self-attention to capture long-range linkage disequilibrium allows it to impute missing alleles with high accuracy on datasets like 1KGP, but I’m curious about the computational cost; how robust is that imputation if we introduce even more complex, non-linear interactions between loci?