ME-PFP

Consider a protein with no close annotated homolog. Sequence search returns little that is useful, yet the protein carries a recognizable InterPro domain and sits in a well-mapped corner of its organism's interaction network — two independent lines of functional evidence a sequence-only predictor never sees. ME-PFP is built on the observation that these evidence sources fail in different places, so a predictor reading all three can cover the gaps in any one of them.

ME-PFP assigns Gene Ontology (GO) terms to proteins from three feature channels: sequence representations taken from a pretrained protein language model, protein domain annotations, and protein–protein interaction data. Rather than concatenating the three into a single vector, it gives each modality its own attention-based feature extractor, shaped to the structure of that data type, and each extractor emits a full multi-label score vector over the ontology's GO terms. Those three score vectors are then combined by a learned dynamic weighting layer. The weighting is what makes this more than a vote: it can lean on domain evidence for a protein whose interaction row is sparse, and on the network for a protein whose domain composition says little.

The model was developed by Haoxing Luo, Yue Hu and colleagues at Xinjiang University in Urumqi, spanning its School of Software and its School of Computer Science and Technology, with Lei Deng, who is affiliated with both Xinjiang University and Central South University, and published in the Journal of Chemical Information and Modeling in 2026. It continues the same group's earlier work on multisource feature fusion, and sits alongside MKFGO, which attacks the same task by fusing five prediction pipelines at the decision level.

#Key Features

• Modality-specific extractors: Each channel gets an attention mechanism matched to its data. The sequence branch applies channel attention over a pooled embedding before a multi-scale 1-D convolution stack; the domain branch runs large- and small-kernel convolutions in parallel and learns a soft split between them.
• Dynamic weighting fusion: The three per-modality GO score vectors are concatenated and passed through two linear layers, so the contribution of each channel is learned per GO term rather than fixed in advance.
• Frozen language model embeddings: The sequence channel consumes precomputed ESM embeddings read from disk. The language model is external and is not retrained, which keeps the trainable part of the system small.
• Multi-label output over the full ontology: Every branch, and the fusion head, ends in a sigmoid over the complete GO term set for the ontology being predicted, so a protein receives a graded score for every term rather than a single class.
• Explicit per-organism scoping: Separate models are trained for each organism and each GO ontology, with input dimensions fixed per organism.

#Technical Details

The released implementation is a compact convolutional ensemble rather than a large pretrained network: each module terminates in a linear layer sized to the GO term count, and a weight classifier fuses the three. Domain identifiers are looked up in a 14,243-entry embedding table with 128-dimensional vectors, and a protein is represented by up to 357 domain entries for human and 40 for yeast. Interaction evidence enters as a 1024-dimensional per-protein vector derived from the organism's PPI network. Separate models are built for the biological process, molecular function and cellular component ontologies, and separately again for human and for yeast — so annotating a query protein requires a language model embedding, domain identifiers, and a row in that organism's PPI representation. Against sequence-based and multisource-fusion baselines on the benchmark data sets, the authors report an average improvement of 13.23% on the human data set and 11.11% on the yeast data set.

#Applications

The model targets the routine annotation problem faced by groups working on incompletely characterized proteomes: assigning candidate GO terms to proteins that homology transfer leaves unlabeled, so downstream enrichment analysis, target triage and hypothesis generation have something to work with. Because it draws on interaction and domain evidence as well as sequence, it suits well-studied organisms where those resources are dense — the human and yeast settings the authors trained for — and not a newly sequenced genome with no interaction network.

#Impact

ME-PFP is a worked argument for decision-level fusion of heterogeneous evidence in function prediction, and for letting a learned weighting rather than a fixed rule decide how much each evidence type is worth for each term. Its reach is bounded by its own design: the per-organism, per-ontology construction means the reported models answer for human and yeast, and extending to another organism means rebuilding the domain vocabulary and interaction features and retraining. The implementation is public under the project's earlier name, MSPredGO, but it ships as training and evaluation code with no trained weights, so reproducing the reported results means assembling the feature inputs and training from scratch.