SMSTPred types methyltransferase substrates from sequence alone: MCC 0.94 under homology-disjoint splits, ahead of sequence and structure homology search
Synopsis
SMSTPred is a multi-label classifier that reads the substrate type of SAM-dependent methyltransferases, DNA, RNA, protein or small-molecule metabolite, from protein language model embeddings without structural input. Under cross-validation with non-overlapping homology clusters it reaches a mean MCC of 0.94, above sequence-homology nearest neighbor MMseqs2 at 0.848, structure-homology Foldseek at 0.887 and a profile HMM at 0.519. Applied to 115,850 candidate methyltransferases from 1,249 reference proteomes, its predictions show a higher small-molecule share in plants and Actinomycetota, while the plant expansion mostly follows proteome size.
Interpretation
Substrate type can now be read directly from an amino-acid sequence, without a structure or an alignable close relative first. Previously substrate type came only from transferring the label of the nearest annotated enzyme by homology search, yet the SAM-binding core shared by Class I methyltransferases does not determine substrate, and most database entries lack experimentally supported catalytic-activity annotation. Over sixteen splits clustered at 30% sequence identity with no homology cluster shared between training and test, the deployed ProtT5 multilayer perceptron averaged MCC 0.940, against 0.848 for MMseqs2, 0.887 for Foldseek and 0.519 for the profile HMM, and those three differences remained significant after Holm correction.
The method compresses a whole sequence into one fixed-length vector, then lets a small network give each of four substrate classes an independent probability. The ProtT5 protein language model is trained self-supervised on hundreds of millions of sequences and maps each protein to the mean of its per-residue representations; the head is two fully connected layers producing four logits trained with binary cross-entropy, so the labels do not exclude one another. The deployed version is an ensemble of six networks of the same architecture, each trained on a different fold, whose mean probability is thresholded at a fixed 0.5, with the standard deviation across the six as uncertainty and a cosine-similarity cutoff of 0.81 to the training set as the applicability-domain flag.
Spread across the tree of life, substrate composition shifts by lineage, and the plant excess of small-molecule methyltransferases mostly comes from larger proteomes. The reviewed Swiss-Prot set is RNA-biased, whereas the predictions make small-molecule metabolites the most abundant class at about 40.9%, suggesting the earlier composition reflects curation history rather than natural abundance. Among 115,850 candidate enzymes from 1,249 reference proteomes, about 72% of a typical actinomycete's methyltransferases are called small-molecule against about 35% in other bacteria; plants show a median 42% against 34% in non-plant eukaryotes, yet methyltransferases make up a similar share of the plant proteome (1.14% versus 1.02%).
Perspective
The demonstrated scope is four-way substrate typing of Class I SAM-dependent methyltransferases, with training labels drawn from Swiss-Prot catalytic-activity annotations backed by literature or curated rules, and evaluation limited to homology-disjoint cross-validation plus predictions over 1,249 reference proteomes; the authors frame dark-cluster substrate calls as leads and note that fixing a reaction still requires experiment.
The authors state that lineage-level trends are consistent with, but do not by themselves establish, differences in methylation strategy between lineages; a higher plant level at equal proteome size is suggested but not resolved, and the plant and fungal fitted intervals overlap; of the 53 dark clusters CLEAN gives a reliable call for only 10 and an unreliable one for 27, and the two methods converge on a specific chemistry for four.
