This repository is for the paper submitted to the 2021 NeurIPS Benchmark track.

1. collect_splits contains notebooks to process RAW datasets collected from various sources.
2. splits contains all splits, a brief description of their processing and the logic behind train/test splits
3. baselines contains code used to compute baselines

A .gitignored folder called data contains RAW data used to produce all splits. As the folder size is substantial, it could not be shipped with GitHub. However, it can be accessed here: http://data.bioembeddings.com/public/FLIP

The goal of the splits in this repository is to assess how well machine learning devices using protein sequence inputs can represent different dimensions relevant for protein design. The main place to find out about the splits is the splits folder. Each set contains a zip file with one or more "splits", where different splits may be different train/test splits based on biological or statistical intuition.

Splits are associated with a semaphore which indicates for what they may be used:

• 🟢: active splits can be used to evaluate accuracy of your machine learning models
• 🟠: splits that should not be used to make performance comparisons, as may give overestimations, or because other active splits have similar discriminative ability
• 🔴: splits that should not be used / considered obsolete. Please do not use these to report performance.