Skip to content

Latest commit

 

History

History
 
 

paper-data

README.md

Here we provide the data associated with the paper "Language models generalize beyond natural proteins" (2022) by Robert Verkuil*, Ori Kabeli*, Yilun Du, Basile I. M. Wicky, Lukas F. Milles, Justas Dauparas, David Baker, Sergey Ovchinnikov, Tom Sercu, and Alexander Rives.

Free Generations (Section 3)

Designs from the Free Generations section of the paper (Section 3) along with their statistics and pdb files can be found at free_generations_full.db and can be loaded using:

import pandas as pd; pd.read_sql('free_generations_full', 'sqlite:///free_generations_full.db')

Designs with wetlab validation (Section 2 and 3)

  • data.csv - Load scalar data in data.csv with pd.read_csv.
  • data.hdf5 For long-form data, download data.hdf5 from this link and load with pd.read_hdf.
# Design information
Design ID - {F,G}{0-267} unique identifier for each (LM or AlphaFold) design evaluated. 8 Nan values correspond to 8 ground truth sequences tested.
Experiment Name - Label for the testing pool to which the design / ground-truth sequence belongs. See Supplement; Section 1.6 for a full description of submitted sequences. These pools (minus ground-truth sequences) have experimental results shown in fig. S11.
Design Model - 228x LM, 20x AlphaFold, 20x AF+ngram, 8x Ground Truth.
Target ID - PDB ID of de novo target for all fixed backbone designs, 'Generation' for all free generations.
Sequence - Designed sequence

# In Silico Evaluation
*AlphaFold predicted PDB file - Structure prediction from AlphaFold (5x pTM models, select best by pLDDT -> Amber Relax).
AlphaFold RMSD - (AlphaFold-predicted) RMSD to target backbone for fixed backbone designs, Nan for free generations
AlphaFold pLDDT - (AlphaFold-predicted) Avg pLDDT for the predicted structure

# Experimental Evaluation
# Results from experimental testing.  Final classifications are in the booleans: {Soluble, Success, Success+Monodisperse}.
Total Yield - Actual total soluble yield (in mg) from the 4x1mL prep. (Actual yield is closer to ~2x, we can only inject 1/2 of the total product onto the column.)
yield_per_Leq - Total Yield, adjusted to 1 L of culture equivalent
*Elution Volume (mL) - Array of x-values for plotting of the SEC trace.
*Chromatographic Absorbance at 280nm - Array of y-values for plotting of the SEC trace.
*Elution Volume (mL) (raw) - Raw version, data is not truncated, lengths may differ between rows.
*Chromatographic Absorbance at 280nm (raw) - Raw version, data is not truncated, lengths may differ between rows.
Soluble - Total Yield > 0.05 mg.
Success - Soluble and SEC peak at the expected elution volume.
Success+Monodisperse - SEC peak *only* at the expected elution volume.

# Jackhmmer results
# See Supplement, Section 1.5 for verbose details.
# In short: Summary statistics of Jackhmmer searches (-n 1 --seed 0) of the designed sequence against UniRef90.  Hits that were removed from ESM2's train set were removed from consideration here.  See `.txt` files for ID's of these omitted sequences.
min Jackhmmer E-value - Minimum (best-domain) E-value
max Jackhmmer Seq-id (significant hits only) - Maximum Sequence identity over all significant (best domain E-value < 1) hits.
max Jackhmmer TM-score (top-10 hits only) - Maximum TM-score of the ≈top-10 (by best-domain E-value) hits.  (Purging was applied after top-10, so the number considered may be slightly lower, counts were rarely reduced below 7).

(* denotes long-form data only available in data.hdf5)

artificial_sequence_purge_ids.txt

ID's of sequences removed due to being annotateed "artificial sequence" by the UniProt website when 2021_04 was the latest release.

uniref90_jackhmmer_purge_ids.txt

ID's of sequences removed by Jackhmmer search (-n 1 --seed 0) of UniRef90 when given the de novo target set as queries.

Minimal structure projection

A small new model head was constructed on top of ESM2, which is this linear projection layer. For a given sequence the projection measures the compatibility of the internal representations of the language model with a structure. The linear projection layer is automatically downloaded when running the lm_design code.

Reference

If using this work, please cite:

@article{verkuil2022language,
  author={Robert Verkuil\*, Ori Kabeli\*, Yilun Du, Basile I. M. Wicky, Lukas F. Milles, Justas Dauparas, David Baker, Sergey Ovchinnikov, Tom Sercu, and Alexander Rives},
  title={Language models generalize beyond natural proteins},
  year={2022},
  journal={bioRxiv},
  note={bioRxiv 2022.12.21.521521},
  url={https://doi.org/10.1101/2022.12.21.521521},
}