InstaDeepAI/true-cds-protein-tasks

Dataset Description

The benchmark consists of five proteins tasks (4 regression and one amino-acid level classification) frequent in the literature which have the associated true CDS seqeunces for each protein. The motivation for this benchmark is to compile a set of protein tasks on which genomic models can be evaluated with the highest reliability.

Tasks Overview

Each of the true CDS protein tasks can be loaded by passing the corresponding name into huggingface load_dataset function.

Task	name	Sample Output	# Train Seqs	# Validation Seqs	# Test Seqs
avGFP Fluorescence Prediction	fluorescence	{sequence, label}	21464	5366	27217
Secondary Structure Prediction (SSP)	ssp	{sequence, label}	7780	NA	334 (3 test sets)
Melting Point Prediction (MPP)	melting_point	{sequence, label}	9432	1064	1648
Stability Prediction	stability	{sequence, label}	53700	2512	12851
Beta-Lactamase Activity (Complete Split)	beta_lactamase_complete	{sequence, label}	9202	2322	1080
Beta-Lactamase Activity (Unique Split)	beta_lactamase_unique	{sequence, label}	3457	865	1080

Splits

Each task has one validation, train and test set, except for SSP. SSP has one training set, and 3 independent test sets. The validation set can simply be randomly split from the training set.

Loading a Dataset Example

from datasets import load_dataset

task_name = "fluorescence"

dataset = load_dataset(
    "InstaDeepAI/true-cds-protein-tasks",
    name=task_name,
)

For SSP, the test splits are named test_casp12, test_cb513, and test_ts115:

from datasets import load_dataset

# Load SSP training data
ssp_train = load_dataset("InstaDeepAI/true-cds-protein-tasks", name="ssp", split="train")

# Load a specific test set
ssp_casp12 = load_dataset("InstaDeepAI/true-cds-protein-tasks", name="ssp", split="test_casp12")

Dataset Tasks

Secondary Structure Prediction (SSP)

The task is a multi-label classification task where each input amino-acid is associated with one of 8 labels, denoting which secondary structure that residue is a part of. All secondary structures were empirically derived using crystallography or NMR. The structural data for the training and validation sets were collected by Klausen et al. Crystal structures were retrieved from Protein Data Bank and filtered with a 25% sequence similarity threshold, a resolution at least as fine as 2.5 angstrom and a length of at least 20 amino acids. Following the work of Klausen we used splits filtered at 25% sequence identity to ensure generalization, and evaluated on 3 independent test sets: CASP12, CB513, TS115.

---

Melting Point Prediction

Melting point prediction (MPP) is a sequence-level regression task that evaluates a model’s ability to predict a measure of melting temperature.

The data orginates from the thermostability atlas, which was originally measured and compiled using a mass spectrometry-based proteomic approach. We follow the same “mixed” splits described in FLIP which seek to avoid over-emphasis of large clusters. Sequences are clustered at 20% identity with 80% of clusters assigned to the train dataset and 20% of clusters assigned to the test dataset.

---

Beta-Lactamase Activity Prediction

Beta-Lactamase is a regression task exploring the fitness landscape of all single codon substitutions in the TEM-1 gene. Labels indicate the ability of mutant genes to confer ampicillin resistance. The data for this task is from Firnberg et al which systematically examined fitness landscape of all single codon mutations in the TEM-1 Beta-lacatamase gene synthesized in native host E. coli. The TEM-1 gene is known to confer antibiotic resistance, and fitness is taken to be a function of this resistance. In particular, gene fitness was measured by splitting the library of mutants onto thirteen sub-libraries exposed to increasing levels of ampicillin concentration. Fitness was calculated as a weighted average of allele counts over plates normalized by WT fitness.

Since beta-lactamase task consists of all single codon mutations, the dataset contains many degenerate coding sequences. In PEER labels were averaged over degenerate coding sequences in the original dataset, however this process removes much data and does not allow us to study gLMs on degenerate sequences. Consequently, we propose two training datasets, sharing a single test set.

Complete Dataset

The Complete set contains all CDS samples except those that are degenerate with respect to any CDS in the test set. Fitness values are the raw CDS fitness values.

Unique Dataset

The Unique set contains a random, maximal, subset of the non-degenerate coding sequences. Comparing the Unique to the Complete allows the study of impact of degenerate CDS on gLM performance. Notably, we use the raw CDS fitness values of the CDS, rather than those averaged over degenerate sequences.

---

Fluorescence Prediction

This task evaluates a model’s ability to predict log-fluorescence of higher-order mutant green fluorescent protein (avGFP) sequences.

Original data is from an Sarkisyan et al, an experimental study of the avGFP fitness landscape. The library was generated via random mutagensis of the wildtype sequence and synthesized in E. coli. Inspired from the TAPE and PEER benchmarks, we restrict the training set to amino-acid sequences with three or fewer mutations from parent GFP sequences, while the test set is all sequences with four or more mutations.

Like the Beta-lactamase task, random mutagenesis of the avGFP gene led to CDS which were degenerate. However, since this process less systematic, this was true of a much smaller fraction of the sequences. In the training set and validation sets there were 54,025 uniquely translating CDS of the 58,417 total sequences. Since most sequences were non-degenerate we selected a random maximal subset and did not study the degenerate sequences. Notably, since the test set was higher order mutants, there was only one amino acid sequence (SS26C:SN168H:SD188V:SS200G) of the 27,217 which did not have a unique coding sequence. We removed randomly one of the two corresponding CDS. Notably, this means the test set is near identical to that used in TAPE and PEER benchmarks.

---

Stability Prediction

This regression task evaluates how well models predict stability around a small region of high-fitness sequences. The train and validation originate from a multi-round experiment and consist of a broad selection of de novo computationally designed proteins composing a small number of topologies. The test set consists of the neighborhoods of single codon mutations around a few of the most stable candidates.

The data for this task originates from Rocklin et al in which stability is measured as a function of the resistance to increasing levels of protease. In particular, the designed libraries were synthesized in yeast and exposed to different concentrations of protease. At each level of protease the fraction of proteins remaining folded was measured, and these values were used to infer the EC₅₀: the value at which half of cells express proteins that pass a defined stability threshold. The stability of a protein is then defined as the difference between the EC₅₀ value of the protein and that of the predicted EC₅₀ in the unfolded state, calculated in a log₁₀ scale.