Looks very interesting. Can I think of this like a more extreme form of the evotuning from UniRep or doi.org/10.1101/2024... except it uses one sequence instead of the sequence plus homologs?

Bioconductor R package: bioconductor.org/packages/MPAC

Shiny app to explore results in manuscript: connect.doit.wisc.edu/content/122/

MPAC uses PARADIGM as the probabilistic model but makes many improvements:
- data-driven omic data discretization
- permutation testing to eliminate spurious predictions
- full workflow and downstream analyses in an R package
- Shiny app for interactive visualization

I found out that Neurosnap offers ESMFold via API neurosnap.ai/service/ESMF...

I may test how many calls are possible with the free academic plan to see if it is worthwhile to update my repo.

The main GitHub repo github.com/gitter-lab/m... links to the extensive resources for running Rosetta simulations at scale to generate new training data, training METL models, running our models, and accessing our datasets. 8/

We can use METL for low-N protein design. We trained METL on Rosetta simulations of GFP biophysical attributes and only 64 experimental examples of GFP brightness. It designed fluorescent 5 and 10 mutants, including some with mutants entirely outside training set mutations. 7/

A powerful aspect of pretraining on biophysical simulations is that the simulations can be customized to match the protein function and experimental assay. Our expanded simulations of the GB1-IgG complex with Rosetta InterfaceAnalyzer improve METL predictions of GB1 binding. 6/

We also benchmark METL on four types of difficult extrapolation. For instance, positional extrapolation provides training data from some sequence positions and tests predictions at different sequence positions. Linear regression completely fails in this setting. 5/

We compare these approaches on deep mutational scanning datasets with increasing training set sizes. Biophysical pretraining helps METL generalize well with small training sets. However, augmented linear regression with EVE scores is great on some of these assays. 4/

METL models pretrained on Rosetta biophysical attributes learn different protein representations than general protein language models like ESM-2 or protein family-specific models like EVE. These new representations are valuable for machine learning-guided protein engineering. 3/

Most protein language models train on natural protein sequence data and use the underlying evolutionary signals to score sequence variants. Instead, METL trains on @rosettacommons.bsky.social data, learning from simulated biophyiscal attributes of the sequence variants we select. 2/

Paper: arxiv.org/abs/2507.12574
GitHub: github.com/gitter-lab/A...
Datasets: doi.org/10.5281/zeno...

7/

There are many more results and controls in the paper. Here's how the best (most negative) docking scores change when we use relevant assays, irrelevant assays, or no assays as context for generation with GPT-4o. In the majority of cases, but not all, relevant context helps. 6/

This generally has the desired effects across multiple LLMs and queried protein targets, with the caveat that our core results are based on AutoDock Vina scores. Assessing generated molecules with docking is admittedly frustrating. 5/

We embed the BioAssay data into a vectorbase, retrieve initial candidate assays, and do further LLM-based filtering and summarization. We select some active and inactive molecules from the BioAssay data table. This is all used for in-context learning and molecule generation. 4/