Protein Tasks · InstaDeepAI
true-cds-protein-tasks
A genomic benchmark evaluating machine learning models on five protein tasks paired with true coding sequences.
What true-cds-protein-tasks is
true-cds-protein-tasks is a biological benchmark curated by InstaDeepAI that compiles five protein function and structure prediction tasks paired directly with authentic nucleotide coding sequences (CDS). Because genomic language models operate on nucleotide inputs, assessing their protein understanding typically requires authentic coding sequences rather than synthetic back-translations.
The suite covers four regression problems and one residue-level classification task: avGFP green fluorescent protein fluorescence prediction, secondary structure prediction, melting point prediction, de novo protein stability prediction, and TEM-1 beta-lactamase activity prediction.
By compiling genuine CDS data with stringent splitting schemes, the benchmark allows researchers to test whether genomic models capture both amino-acid properties and codon-level effects such as synonymous variation.
How it was made
The dataset is structured across six configurations distributed as Parquet tables. Secondary Structure Prediction (SSP) uses multi-label classification to assign each residue to one of eight secondary structure classes derived from crystallographic or NMR data, filtered to 25% sequence identity and evaluated across the CASP12, CB513, and TS115 test sets. Melting Point Prediction (MPP) derives from mass-spectrometry thermostability atlas data clustered at 20% sequence identity into train and test splits.
Beta-lactamase activity prediction investigates single-codon substitution fitness from deep mutational scanning of the TEM-1 gene conferring ampicillin resistance, presented in both a 'complete' version retaining raw synonymous CDS values and a 'unique' subset of non-degenerate sequences. Fluorescence prediction measures avGFP mutant log-fluorescence, splitting sequences so that low-order mutants (three or fewer mutations) are in the training set and high-order mutants (four or more) form the test set. Stability prediction evaluates computationally designed proteins exposed to protease, quantifying stability via the change in EC50.
Key facts
- Data format
- Parquet
- Sequence modality
- Nucleotide coding sequence (CDS) and amino acid
- Number of tasks
- 5 tasks across 6 configurations
- Task types
- 4 sequence-level regression tasks, 1 residue-level multi-label classification task
What you can use it for
- Genomic Secondary Structure Prediction
- Benchmarking residue-level eight-state secondary structure classification directly from genuine coding sequences on CASP12, CB513, and TS115 test sets.
- Synonymous Codon Variant Effect Prediction
- Comparing the performance of genomic models on the complete versus unique TEM-1 beta-lactamase splits to study the effect of degenerate, synonymous coding mutations on antibiotic resistance prediction.
- Out-of-Distribution Mutational Extrapolation
- Testing whether genomic models trained on avGFP variants with three or fewer mutations can accurately predict the log-fluorescence of variants with four or more mutations.
- Protein Thermostability and Stability Modelling
- Predicting global protein melting temperatures from sequence clusters, as well as EC50 stability shifts in computationally designed de novo proteins.
Explore the data
The schema, real rows, full-text search and a SQL filter — read in your browser, nothing downloaded.
Explore
Read the real rows without downloading anything
Reading rows…
LiveRead from InstaDeepAI/true-cds-protein-tasks at the moment you asked. Nothing is cached or stored — every row above came from that request.
How the columns are distributed
Over 9,202 rows of beta_lactamase_complete/train
label
float- min
- 8.3e-4
- median
- 0.45
- mean
- 0.54
- max
- 3.11
sequence
string_text- min
- 858
- median
- 858
- mean
- 858
- max
- 858
Strengths and limits
Good at
- Pairs established protein fitness benchmarks with authentic nucleotide coding sequences rather than artificial back-translations.
- Provides both complete and non-degenerate splits for beta-lactamase to study the impact of synonymous codon redundancy on model predictions.
- Incorporates strict sequence identity thresholds and mutational distance splits to ensure rigorous out-of-distribution evaluation.
Watch out for
- The secondary structure prediction task does not include a predefined validation set, requiring users to split one manually from the training partition.
- The benchmark is restricted to five specific protein tasks, leaving out other common biological phenotypes.
- Most degenerate coding sequences were pruned from the fluorescence task rather than systematically retained for synonymous mutation analysis.
Questions
Can I look inside true-cds-protein-tasks without downloading it?
Yes. This page reads its schema, real rows and per-column statistics in the browser, with full-text search and a SQL filter. Nothing is downloaded to your machine.
What licence is true-cds-protein-tasks under?
Its authors declare cc-by-nc-sa-4.0. Read the terms on the original page before reusing the data.
What is the difference between the complete and unique beta-lactamase splits?
The complete split contains all coding sequence samples and their raw fitness measurements, excluding those identical to sequences in the test set. The unique split selects a random maximal subset of non-degenerate coding sequences to enable comparisons of how synonymous codons affect model performance.
How are the test splits organised for secondary structure prediction?
Secondary structure prediction has no default validation split but includes three independent, standard test splits: test_casp12, test_cb513, and test_ts115.
How is stability defined in the stability task?
Stability is calculated on a log10 scale as the difference between the observed EC50 (the protease concentration where half of cells express folded proteins) and the predicted EC50 of the unfolded state.
How to cite it
arXiv:2206.02096
Written from true-cds-protein-tasks's own dataset card. Statements citing a figure that does not appear there are removed before publication. Licence and running details come from our registry.