update readme

README.md

@@ -542,11 +542,8 @@ Of these
 *** **IMPORTANT! Please [register your use](https://forms.gle/wuHv8MKmEu4EEMA99) of these data so that we (the Rocklin Lab) can continue to release new useful datasets!! This will take 10 seconds!!** ***
 
 
-
 ## Quickstart Usage
 
-
-
 ### Install HuggingFace Datasets package
 
 Each subset can be loaded into python using the Huggingface [datasets](https://huggingface.co/docs/datasets/index) library.
@@ -565,7 +562,7 @@ then, from within python load the datasets library
 
 ### Load model datasets
    
-To load one of the `MegaScale` model datasets, use `datasets.load_dataset(...)`:
+To load one of the `MegaScale` model datasets (see available datasets below), use `datasets.load_dataset(...)`:
 
     >>> dataset_tag = "dataset3_single"
     >>> dataset3_single = datasets.load_dataset(
@@ -643,13 +640,14 @@ Curated set of `325,132` ΔG measurements at `17,093` sites in `365` domains
 
 **`dataset3_single`**:
 The single point mutations in `dataset3`
-  * Using the train/val/test splits defined in ThermoMPNN
-  * 
+  * Using the train/val/test splits defined in ThermoMPNN [(Dieckhaus, et al., 2024)](https://www.pnas.org/doi/abs/10.1073/pnas.2314853121)
 
-**`dataset3_single_CV`**:
+**`dataset3_single_cv`**:
 The single point mutations in `dataset3`
-  * Using the 5-fold cross validation splits (`train_[0-4]`/`val_[0-4]`/`test_[0-4]`) defined in ThermoMPNN
+  * Using the 5-fold cross validation splits (`train_[0-4]`/`val_[0-4]`/`test_[0-4]`) defined in ThermoMPNN [(Dieckhaus, et al., 2024)](https://www.pnas.org/doi/abs/10.1073/pnas.2314853121)
 
+**`AlphaFold_model_PDBs`**:
+AlphaFold predicted structures of wildtype domains (even if structures exist in the Protein Databank)
 
 ### Target Selection
 Targets consist of natural, designed, and destabilized wild-type
@@ -666,7 +664,7 @@ These were then processed by
     which was used to trim amino acids from the N- and C termini that had a low number of contacts with any other residues.
   * selected domains with up to 72 amino acids after excluding N- or C-terminal flexible loops
 
-XXX **designed targets** were selected from
+**designed targets** were selected from
   * previous Rosetta designs with ααα, αββα, βαββ, and ββαββ topologies (40 to 43 amino acids)
   *  new ββαα proteins designed using Rosetta (47 amino acids)
   * new domains designed by trRosetta hallucination (46 to 69 amino acids)
@@ -715,37 +713,9 @@ Each target was analyzed and given a single quality category score G0-G11, which
   * G11: Probably cleaved in folded state(s)
 
 
-~~~~~~~~~~~~~~~~~~~~~~~~~~~~
-
-1.8 million measurements in total
-
-
-
-We determine ΔG using each sequence’s
-   * measured K50, a predicted sequence-specific K50 for the unfolded state (K50,U)
-   * a universal K50 for the folded state (K50,F)
-
-
-published studies using purified protein samples for 1,188 variants of 10 proteins (Fig. 1g and Supplementary Fig. 1 for more details on GB129)
-Our measurements for these sequences were all performed in libraries of 244,000–900,000 total sequences.
-
-
-
-Other Datasets for comparison
-  * ProthermDB
-    * Thermodynamic data
-    * Thermal proteome profiling
-  * Rocklin2017
-
-
-Tsuboyama2023_Dataset2_Dataset3_20230416.csv
-  * All sequences in dataset 2 and dataset 3 are included
-  * All sequences in this file have an inferred ΔG estimate
-  * only sequences in dataset 3 have a tabulated ΔΔG estimate
-  * datasets 2 and 3 include a very small number of sequences with low-quality data (wide confidence intervals)
-    because these sequences come from mutational scans that are high quality overall
-  * low-quality data (including mutant data filtered in Stage 3) have been filtered out and
-    replaced by a "–"" symbol in the columns labelled ‘_ML’ (for machine learning).
-
-predicting wild-type amino acids from the folding stabilities (ΔG) of each protein variant
-  * 99,156 ΔG measurements (5,214 sites in 90 non-redundant natural domains)
+## ThermoMPNN splits
+ThermoMPNN is a message passing neural network that predicts protein ΔΔG of mutation
+based on ProteinMPNN [(Dauparas et al., 2022)](https://www.science.org/doi/10.1126/science.add2187).
+ThermoMPNN uses in part data from the MegaScale dataset. From the mutations in `dataset2`,
+272,712 mutations across 298 proteins were curated that were single point mutants, reliable,
+and where the baseline is wildtype.

src/02.1_gather_ThermoMPNN_splits.py

@@ -6,8 +6,6 @@ import pickle
 
 
 
-
-
 # Gabe requested that the splits defined in ThermoMPNN of the MegaScale dataset be the default splits
 
 # ThermoMPNN/datsets.py