TMbed: transmembrane proteins predicted through language model embeddings

Abstract

Background Despite the immense importance of transmembrane proteins (TMP) for molecular biology and medicine, experimental 3D structures for TMPs remain about 4–5 times underrepresented compared to non-TMPs. Today’s top methods such as AlphaFold2 accurately predict 3D structures for many TMPs, but annotating transmembrane regions remains a limiting step for proteome-wide predictions. Results Here, we present TMbed, a novel method inputting embeddings from protein Language Models (pLMs, here ProtT5), to predict for each residue one of four classes: transmembrane helix (TMH), transmembrane strand (TMB), signal peptide, or other. TMbed completes predictions for entire proteomes within hours on a single consumer-grade desktop machine at performance levels similar or better than methods, which are using evolutionary information from multiple sequence alignments (MSAs) of protein families. On the per-protein level, TMbed correctly identified 94 ± 8% of the beta barrel TMPs (53 of 57) and 98 ± 1% of the alpha helical TMPs (557 of 571) in a non-redundant data set, at false positive rates well below 1% (erred on 30 of 5654 non-membrane proteins). On the per-segment level, TMbed correctly placed, on average, 9 of 10 transmembrane segments within five residues of the experimental observation. Our method can handle sequences of up to 4200 residues on standard graphics cards used in desktop PCs (e.g., NVIDIA GeForce RTX 3060). Conclusions Based on embeddings from pLMs and two novel filters (Gaussian and Viterbi), TMbed predicts alpha helical and beta barrel TMPs at least as accurately as any other method but at lower false positive rates. Given the few false positives and its outstanding speed, TMbed might be ideal to sieve through millions of 3D structures soon to be predicted, e.g., by AlphaFold2.

LambdaPP: Fast and accessible protein-specific phenotype predictions

“…predicting CATH (63) classes (26), LambdaPP will be updated to extend its breadth. All feature prediction methods integrated into the LambdaPP webserver currently use ProtT5 (24) that, in our hands, outperformed ESM-1b (55) and others (4; 9; 25; 24) for numerous applications (38-40; 43; 65; 13; 26; 73). This consistency also increases speed as the generation of embeddings becomes a limiting step.…”

“… Panel A: residue level features: secondary structure, transmembrane topology, disordered residues, small molecule, nucleic or metal binding residues, residue conservation and average variation (24; 40; 43; 13; 30); Panel B: sequence-level features: predicted subcellular localization (65), and an excerpt of predicted GO-annotations (39); Panel C: effect of SAVs (wild-type sequence on x-axis, mutations on y-axis; darker color=higher effect) (43); and Panel D : predicted 3D structure (46). Interactive version at https://embed.predictprotein.org/o/Q9NZC2.…”

“…Per-protein features predicted solely with ProtT5 embeddings as input currently include subcellular location (65), and Gene Ontology terms (GO) (39). Per-residue predictions solely with ProtT5 embeddings as input include: conservation (43); helical transmembrane regions, transmembrane beta barrels, along with signal peptides (13); binding for various ligands (40); intrinsically disordered regions (30); secondary structure (24). LambdaPP also predicts the effect of introducing single amino acid variants (SAV) in the input sequence upon molecular function, which uses the predicted conservation with a BLOSUM62-score (27) of the SAV as input (43).…”

“…Starting with an amino acid sequence, LambdaPP orchestrates the prediction of (1) protein 3D structure by ColabFold (46); (2) per-protein features: Gene Ontology (GO) annotations: goPredSim (39), subcellular location: LA (65); (3) per-residue features: binding residues: bindEmbed21DL (40), conservation: ProtT5cons (43), disorder: SETH (30), secondary structure: ProtT5-sec (24), helical and barrel transmembrane (TM) regions: TMbed (13); and (4) variant effect scores: VESPAl (43). …”

From sequence to function through structure: deep learning for protein design

“…In particular, we unconditionally (i.e., without priors on family, function or structure) generated a set of 100,000 protein sequences using ProtGPT2, and predicted secondary structure [22], Gene Ontology (GO) terms [44], residue ability to bind small molecules, nucleotides or metals [47], protein subcellular localization [46], transmembrane topology [41], residue conservation [42], residue disorder [43] and CATH family [45]. Remarkably, this generated a repertoire of 100,000 protein sequences with ∼12 predicted features of structure and function from a single script in approximately 3.5 hours ( Supplement 1 ).…”