We also made some improvements with genomic language model, Evo 2, but in this case the interpretation was less clear. See the preprint for more details. Code for using LFB will made available shortly. 10/10

This provides evidence that better fitness estimation can be achieved at negligible computational cost by bridging the gap between likelihood and fitness at inference time. 9/n

This trend held across DMS assay types and mutational depth, and also on prediction of clinical variants. 8/n

On ProteinGym, LFB provided significant improvements across model classes and sizes and we saw that larger better fit models provided better predictions in general.
proteingym.org 7/n

We found under an Ornstein–Uhlenbeck model of evolution that our LFB should be lower variance than the standard estimate by marginalising the effect of drift. 6/n

We tried a simple strategy — averaging predictions over sequences under similar selective pressures to effectively reduce the impact of unwanted non-fitness related correlations — likelihood fitness bridging (LFB). 5/n

We wondered whether we might be able to improve predictions from existing models without any further training. 4/n

Weinstein et al show that better fit sequence models can perform worse at fitness estimation due to phylogenetic structure:
openreview.net/forum?id=CwG...
And in practice we are seeing that pLMs don’t improve with lower perplexities:
openreview.net/forum?id=UvP... www.biorxiv.org/content/10.1... 3/n

Protein language models are showing promise in variant effect prediction - but there’s emerging evidence likelihood based zero shot fitness estimation is breaking down. See this excellent summary from @pascalnotin.bsky.social: pascalnotin.substack.com/p/have-we-hi... 2/n