Deep Recurrent Neural Network for Protein Function Prediction from Sequence

Key results

1. ShownOn randomly held-out UniProt sequences the classifiers separated each of four function classes - ferritin-like iron sequestration, cytochrome P450, serine and cysteine proteases, and G-protein coupled receptors - from the rest of SwissProt almost perfectly.

   Positives were keyword-matched in UniProt, negatives were all other reviewed entries (around 550K sequences), split 80/20, with negatives fed in four or more class-balanced chunks for at least five epochs each (ADAM, mini-batch 64). Precision, recall, F1 and ROC area under the curve on the deliberately unbalanced test split are described as close to unity; the per-class values live in tables absent from this text. · In-class Model Training and Validation

2. ShownGeneralising to phylogenetically distinct families with the same function was far harder, and extra depth helped: recall for the 12-subunit dps cage rose from 0.13 to 0.36 when the recurrent stack was tripled, without losing precision.

   Trained on non-haem ferritin (IPR001519) with bacterioferritin (IPR002024) and tested on dps (IPR002177). Training on Cas9 (IPR028629, 1,201 sequences) and testing on Cpf1 (IPR027620, 55 sequences) gave recall 0.59 when predictions from the first and last 800 residues were averaged, while either terminus alone lost precision. Left-out P450 families were called with sensitivities not quoted in this text. · Out-of-class Training and Prediction

3. ShownTen predicted iron-sequestering proteins that carried no prior functional annotation all reduced intracellular free iron when expressed in E. coli, and some, notably the 'algae', 'human' and 'archaea' candidates, also made the cells more magnetically retentive.

   Candidates from diverse domains of life were expressed from a rhamnose-inducible plasmid in E. coli BW25113 carrying a fiu-promoter sfGFP iron sensor: all ten raised fluorescence significantly, equivalent to lower free iron, against uninduced cells (P < 0.05, two-tailed Student's t-test), though the 'potato' protein barely shifted it. 'Algae', 'human' and 'archaea' also increased magnetic column retention. · Experimental Validation of Predicted Function; Construction of expression vectors for predicted…; Iron level characterization by genetic sensor

4. ShownScanning UniRef100 (54 million sequences), the models returned thousands of hits per function that an iterated jackhmmer search missed; most carried an independent family or gene-ontology assignment elsewhere, and a further subset had no UniProt annotation.

   jackhmmer was seeded with FTNA_ECOLI, CP21A_HUMAN, SEPR_HUMAN and FFAR2_HUMAN and run for at least five iterations at the default reporting threshold E = 10.0. Clustal Omega alignment of the unannotated ferritin-class predictions showed many lineages with no obvious shared pattern. Actual hit counts appear only in a table missing from this text. · Database Search and Prediction; In-class Model Training and Validation

5. ShownThe sequence model outperformed classical classifiers trained on hand-computed descriptors, though the comparison is reported only qualitatively.

   Logistic regression and random forest were fitted to 51 ProtParam features (amino acid composition, length, isoelectric point, molecular weight, stability index, hydrophobicity, GRAVY) with grid-searched hyperparameters on the same four datasets: logistic regression trained fastest and predicted worst, random forest did better but stayed below the RNN. No accuracy figures for the baselines appear in the text. · Comparison of RNN to Other Machine Learning…