PredictProtein - Predicting Protein Structure and Function for 29 Years

1The sequence is known for far more proteins than experimental annotations of function or structure. This sequence-annotation gap existed when PredictProtein started in 1992, and has kept expanding ever since. Unannotated sequences contribute crucial evolutionary information to neural networks predicting secondary structure that seeded PredictProtein (PP) at the European Molecular Biology Laboratory (EMBL) in 1992, the first fully automated, query-driven Internet server providing evolutionary information and structure prediction for any protein. Many other methods predicting aspects of protein function and structure have since joined under the PP roof now hosted by the Luxembourg Centre of Systems Biomedicine (LCSB).

2PP offers an array of structure and function predictions most of which combine machine learning with evolutionary information; now enhanced by a faster alignment algorithm. A few prediction methods now also use embeddings from protein Language Models (LMs). Embeddings are much faster to obtain than evolutionary information, yet for many tasks, perform almost as well, or even better. All PP methods join at PredictProtein.org with interactive visualizations; for some methods, more advanced visualizations are linked. The reliability of PredictProtein, its speed, the continuous integration of well-validated, top methods, and its intuitive interface have attracted thousands of researchers over 29 years of steady operation.

3multiple sequence alignments (MSAs) and position-specific scoring matrices (PSSMs) computed by a combination of pairwise BLAST, PSI-BLAST, and MMseqs2 on query vs. PDB and query versus UniProt. For each residue in the query, the following per-residue predictions are assembled: secondary structure (RePROF/PROFsec and ProtBertSec); solvent accessibility (RePROF/PROFacc); transmembrane helices and strands (TMSEG and PROFtmb); protein disorder (Meta-Disorder); backbone flexibility (relative B-values; PROFbval); disulfide bridges (DISULFIND); sequence conservation (ConSurf/ConSeq); protein-protein, protein-DNA, and protein-RNA binding residues (ProNA2020); PROSITE motifs; effects of sequence variation (single amino acid variants, SAVs; SNAP2). For each query per-protein predictions include: transmembrane topology (TMSEG); binary protein-(DNA|RNA|protein) binding (protein binds X or not; ProNA2020); Gene Ontology (GO) term predictions (goPredSim); subcellular localization (LocTree3); Pfam domain scans, and some biophysical features. Under the hood, PP computes more results (SOM: PredictProtein Methods; Supplementary Table S1), either as input for frontend methods, or for legacy support.

4goPredSim predicts GO terms by transferring annotations from the most embedding-similar protein. Embeddings are obtained from SeqVec; similarity is established through the Euclidean distance between the embedding of a query and a protein with experimental GO annotations. Replicating the conditions of CAFA3 in 2017, goPredSim achieved Fmax values of 37 ± 2%, 52 ± 2% and 58 ± 2% for BPO (biological process), MFO (molecular function), and CCO (cellular component), respectively. Using Gene Ontology Annotation (GOA) to test 296 proteins annotated after February 2020, goPredSim appeared to reach even slightly higher values that were confirmed by CAFA4.

5ProtBertSec predicts secondary structure in three states (helix, strand, other) using ProtBert embeddings derived from training on BFD with almost 3 × 109 proteins. On a hold-out set from CASP12, ProtBertSec reached a three-state per-residue accuracy of Q3 = 76 ± 1.5%. Although below the state-of-the-art (NetSurfP-2.0 at 82%), this method performed on-par with other MSA-based methods, despite itself not using MSAs.

6ProNA2020 predicts whether or not a protein interacts with other proteins, RNA or DNA (binary), and if so, where it binds (which residues). The binary per-protein predictions rely on homology and machine learning models employing profile-kernel SVMs and on embeddings from an in-house implementation of ProtVec. Per-residue predictions (where) use simple neural networks due to data shortage. ProNA2020 correctly predicted 77 ± 1% of the proteins binding DNA, RNA or protein. In proteins known to bind other proteins, RNA or DNA, ProNA2020 correctly predicted 69 ± 1%, 81 ± 1% and 80 ± 1% of binding residues, respectively.

7Most time-consuming for PP was the search for related proteins in ever growing databases. MMseqs2 finds related sequences blazingly fast and seeds a PSI-BLAST search. The query sequence is sent to a dedicated MMseqs2 server that searches for hits against cluster representatives within the UniClust30 and PDB reduced to 70% pairwise percentage sequence identity (PIDE). All hits and their respective cluster members are turned into a MSA and filtered to the 3000 most diverse sequences.

8Users query PredictProtein.org by submitting a protein sequence. Results are available in seconds for sequences that had been submitted recently (cf. PPcache next section), or within up to 30 min if predictions are recomputed. Per-residue predictions are displayed online via ProtVista, which allows to zoom into any sequential protein region (Supplementary Figure S1), and are grouped by category (e.g. secondary structure), which can be expanded to display more detail (e.g. helix, strand, other). On the results page, links to visualize MSAs through AlignmentViewer are available. More predictions can be accessed through a menu on the left, e.g. Gene Ontology Terms, Effect of Point Mutations and Subcellular Localization. Prediction views include references and details of outputs, as well as rich visualizations, e.g. GO trees for GO predictions and cell images with highlighted predicted locations for subcellular localization predictions.

9The PP backend lives at LCSB, allowing for up to 48 parallel queries. Results are stored on disc in the PPcache. Users submitting sequences for which results were over the last two years obtain results immediately. Using the bio-embeddings pipeline, the PPcache is enriched by embeddings and embedding-based predictions on the fly. For all methods displayed on the frontend, JSON files compliant with ProtVista are available via REST APIs (SOM: Programmatic access), and are in use by external services such as the protein 3D structure visualization suite Aquaria and by MolArt.

10All results displayed on and downloadable from PP are available through Docker and as source code for local and cloud execution (available at github.com/rostlab).

11We demonstrate PredictProtein.org tools through predictions of the novel coronavirus SARS-CoV-2 (NCBI:txid2697049) nucleoprotein (UniProt identifier P0DTC9/NCAP_SARS2; Figure 1; SOM: Use Case; Supplementary Figure S2). NCAP_SARS2 has 419 residues and interacts with itself (experimentally verified). Sequence similarity and automatic assignment via UniRule suggest NCAP is RNA-binding (binding with the viral genome), binding with the membrane protein M (UniProt identifier P0DTC5/VME1_SARS2), and is fundamental for virion assembly. goPredSim transferred GO terms from other proteins for MFO (RNA-binding; GO:0003723; ECO:0000213) and CCO (compartments in the host cell and viral nucleocapsid; GO:0019013; GO:0044172; GO:0044177; GO:0044220; GO:0030430; ECO:0000255) matching annotations found in UniProt. While it missed the experimentally verified MFO term identical protein binding (GO:0042802), goPredSim predicted protein folding (GO:0006457) and protein ubiquitination (GO:0016567) suggesting the nucleoprotein to be involved in biological processes requiring protein binding. ProNA2020 predicts RNA-binding regions, the one with highest confidence between I84 (Isoleucine at position 84) and D98 (Aspartic Acid at 98) (protein sequence in SOM: Use Case). While high resolution experimental data on binding is not available, an NMR structure of the SARS-CoV-2 nucleocapsid phosphoprotein N-terminal domain in complex with 10mer ssRNA (PDB identifier 7ACT) supports the predicted RNA-binding site (Figure 2). Additionally, SNAP2 predicts single amino acid variants (SAVs) in that region to likely affect function, reinforcing the hypothesis that this is a functionally relevant site. Although using different input information (evolutionary vs. embeddings), RePROF and ProtBertSec both predict an unusual content exceeding 70% non-regular (neither helix nor strand) secondary structure, suggesting that most of the nucleoprotein might not form regular structure. This is supported by Meta-Disorder predicting 53% overall disorder. Secondary structure predictions match well high-resolution experimental structures of the nucleoprotein not in complex (e.g., PDB 6VYO; 6WJI). Both secondary structure prediction methods managed to zoom into the ordered regions of the protein and predicted e.g., the five beta-strands that are formed within the sequence range I84 (Isoleucine) to A134 (Alanine), and the two helices formed within the sequence range spanned from F346 (Phenylalanine) to T362 (Tyrosine).

12For almost three decades (preceding Google) PredictProtein (PP) has been offering predictions for proteins. PP is the oldest prediction Internet server, online for 6-times as long as most other servers. It pioneered combining machine learning with evolutionary information and making predictions freely accessible online. While the sequence-annotation gap continues to grow, the sequence-structure gap might be bridged in the near future. For the time being, servers such as PP, providing a one-stop solution to predictions from many sustained, novel tools are needed. Now, PP is the first server to offer fast embedding-based predictions of structure (ProtBertSec) and function (goPredSim). By slashing runtime for PSSMs from 72 to 4 min through MMseqs2 and better LCSB hardware, PP also delivers evolutionary information-based predictions fast! Instantaneously if the query is in the precomputed PPcache. For heavy use, the offline Docker containers provide predictors out-of-the-box. At no cost to users, PredictProtein offers to quickly shine light on proteins for which little is known using well validated prediction methods.

13Freely accessible webserver PredictProtein.org; Source and docker images: github.com/rostlab.