Deep evolutionary mining and structural distillation transform human interactome prediction

Cellular functions emerge from ensembles of PPIs that coordinate signaling, metabolism, and gene regulation.Over the past two decades, large-scale experimental efforts have mapped fragments of the human interactome using yeast two-hybrid (Y2H) [1] , and affinity purification-mass spectrometry (AP-MS) [2] .While large-scale experimental efforts like HuRI [3] , BioPlex [4] have reported hundreds of thousands of putative PPIs.These datasets remain noisy and incomplete.Computational approaches broaden coverage by combining homology transfer, interface modeling, and functional association.Yet, they still face a trade-off between scalability and mechanistic resolution.Deep-learning-based structure prediction has begun to reshape this landscape, as AlphaFold2 [5] is often used to generate accurate models for monomeric proteins and many complexes.Cong et al. in 2019 achieved 50%-70% recall via deep prokaryotic MSAs in yeast [6] .Humphreys et al. in 2021 leveraged fungal diversity for stable complex prediction [7] .In these microbial studies, concatenated multiple sequence alignments (MSAs) of orthologs across thousands of species provide strong interfacial coevolutionary signals, which direct coupling analysis or deep networks can decode to distinguish true PPIs from random pairs.However, these efforts yielded higher recall but struggled with precision for transient interactions due to a simpler network.A recent study by Zhang and colleagues [8] presents a significant conceptual and methodological shift in this domain.They integrate three forms of information that have not previously been combined at this scale.These include evolutionary depth extracted from raw eukaryotic sequencing data, structural regularities obtained through large-scale distillation of domain interactions from AlphaFold monomer models, and a learning architecture designed specifically for interaction discrimination.As shown in Fig. 1, we can find how this system integrates evolutionary and structural cues into a unified screening architecture.Figure 1a demonstrates how genomic-scale multiple sequence alignments strengthen coevolutionary signals, while Fig. 1b shows the dual pipeline design that evaluates candidate protein pairs.One of the most substantial advances in this study concerns the construction of deeply enriched multiple sequence alignments.Standard MSAs rely on curated reference proteomes, but many human proteins have limited orthologs, weakening coevolutionary signals.Prior tools like DeepMSA2 [9] incorporated iterative searches through genomic and metagenomic databases but prioritized prokaryotic or broad metagenomic data, often yielding shallower alignments for eukaryotic proteins.By contrast, the framework introduced by Zhang and colleagues retrieves coding sequences directly from raw genomic and transcriptomic data for more than 20,000 eukaryotic species.Using custom assembly and ortholog search procedures, the authors construct omicMSAs with far greater diversity than conventional methods.Conventional MSAs rely on annotated sequences, yielding shallow alignments for eukaryotes.In contrast, omicMSAs mine 30 PB of raw data from 21,415 species,