Predicting Protein-Protein Interactions in the Human Proteome: Methods and Challenges

Predicting protein-protein interactions (PPIs) at proteome scale is possible, but a predicted interaction is a candidate for testing, not a confirmed biological relationship. A 2025 study combined deeper sequence alignments with a model trained on domain interactions to screen human protein pairs. Its result offers a way to prioritize experiments, while the limits of the evidence still matter Predicting protein-protein interactions in the human proteome (2025).

PPI prediction draws on several kinds of evidence. Coevolutionary methods look for amino-acid changes that appear together across related species. Structure-based methods assess whether protein domains could form a compatible interface. Curated databases provide existing interaction evidence and annotations that can help assess new candidates. Each signal has weaknesses, so agreement between independent sources can make a prediction more useful.

How the 2025 approach works

Coevolutionary analysis depends on multiple sequence alignments (MSAs), which group related protein sequences across species. When two proteins interact, changes at contacting residues may be correlated over evolutionary time. Statistical methods can use those patterns to identify possible contacts. This approach depends on having enough related sequences; shallow alignments can produce unreliable signals.

The 2025 study addressed that limitation by using unassembled genomic data to create deeper alignments. The paper reports that this produced alignments about seven times deeper than standard reference-database MSAs. More sequence data can improve the evidence available for coevolutionary analysis, although it cannot guarantee a clear signal for every protein pair Predicting protein-protein interactions in the human proteome (2025).

The study also used predicted protein structures to train a deep-learning network on domain-domain interactions. Many proteins contain distinct domains, and an interaction may depend on only one part of each protein. A model trained to recognize domain-level contacts can focus on those parts rather than treating every pair as a single, undivided structure. The study describes training data augmented with information from predicted structures; those structures are computational models, not direct experimental observations.

The researchers screened 200 million human protein pairs and reported roughly 18,000 predicted interactions. These figures describe the scope and output of that study, not a complete map of the human interactome. They also do not mean that every reported pair has been experimentally confirmed.

One example: evaluating a candidate pair

Suppose a model predicts that proteins A and B interact, but no experimental record is available. Treat the result as a hypothesis and check what supports it.

First, inspect the score and the evidence behind it. A strong coevolutionary signal and a plausible structural interface provide different kinds of support. If both point to the same contact, the candidate may be a stronger experimental target. If the signals disagree, the mismatch can help identify what remains uncertain. For example, a structural model may poorly represent a flexible or disordered region.

Next, check annotations and existing evidence. UniProt (https://www.uniprot.org) can help identify protein domains and other reviewed annotations. STRING (https://string-db.org) collects interaction evidence from several sources and assigns confidence scores. A database result can support a prediction, but it does not automatically confirm a direct physical interaction. A prediction with no database support remains a candidate for follow-up.

Finally, choose an experiment suited to the question. An assay that detects association may help test whether the proteins occur in the same complex, while other methods can provide information about the contact site. The right choice depends on the proteins, the proposed interface, and the biological conditions under study. No single computational score determines which assay will be decisive.

What the predictions can and cannot tell you

A low score does not prove that two proteins never interact. The model may lack enough sequence information, may not represent the relevant protein state, or may miss an interaction that occurs only in a particular tissue or condition. A high score also needs care: predicted structures and training data have limits, and computational support is not experimental validation.

The method is likely to be more informative for some interactions than others. Domain-mediated contacts may fit the model’s training approach well. Contacts that depend on flexible regions, chemical modifications, or conditions absent from the training evidence may be harder to predict. These are general limitations of the evidence used, not reasons to dismiss every prediction in those categories.

The 200 million pairs screened represent a large analysis, but the human protein pair space is larger still. The study’s output should therefore be read as a set of candidates selected by a particular method, not as a census of all interactions. A pair absent from the predicted set may be a true non-interaction, or it may simply be unresolved by the available data and model.

Curated resources help put predictions in context. UniProt offers reviewed protein annotations, while STRING combines interaction evidence from multiple sources. Their records can guide candidate selection, but database coverage is uneven and different evidence types support different conclusions. For a novel candidate, the absence of a database record is not proof that the interaction is false.

Key Takeaways

• The 2025 study used deeper MSAs and a domain-based deep-learning model to screen 200 million human protein pairs. It reported roughly 18,000 predicted interactions Predicting protein-protein interactions in the human proteome (2025).
• Coevolutionary signals, predicted structural compatibility, and curated annotations provide different forms of evidence. A prediction supported by multiple signals is a stronger candidate for testing, but it is not confirmed.
• UniProt (https://www.uniprot.org) and STRING (https://string-db.org) can help assess annotations and existing evidence.
• Low scores do not establish that an interaction is absent. High scores do not replace experimental validation.
• Use predictions to prioritize candidates and plan experiments.