Agentic AI platforms for autonomous training and rule induction of human-human and virus-human protein-protein interactions

This paper demonstrates the capability of agentic AI to autonomously orchestrate two specialized platforms: one that builds high-accuracy predictive machine learning models for human-human and virus-human protein-protein interactions, and another that induces interpretable, human-readable rules that align with those models' findings.

The Gist

The "Digital Scientist" Duo: Teaching AI to Decode the Body’s Secret Handshakes

Imagine your body is a massive, bustling city. In this city, millions of tiny workers (called proteins) are constantly running around. For the city to function, these workers need to find each other and perform specific tasks—like building a bridge or delivering a package. When two workers meet and work together, it’s called a Protein-Protein Interaction (PPI).

If the workers (proteins) are doing their jobs, the city thrives. But if a "bad actor" (like a virus) sneaks into the city and starts shaking hands with the wrong workers, it can hijack the city’s systems and cause chaos.

The Problem:
Scientists want to know exactly who is shaking hands with whom. If we can predict these "handshakes," we can design medicines to stop viruses or understand diseases. However, mapping these interactions is incredibly hard. It’s like trying to map every single handshake in a city of billions, where half the people are wearing masks and many of the handshakes are never recorded.

---

The Solution: Two AI "Super-Detectives"

Instead of having humans do all the tedious work, the researchers built two different types of Agentic AI. Think of "Agentic AI" not just as a smart calculator, but as a Digital Scientist—an AI that can plan, execute, and check its own work.

1. The "Predictor" Detective (The Pattern Matcher)

The first AI platform is like a detective who is an expert at spotting patterns.

• How it works: This detective has a team of five specialized mini-agents. One goes to the library to find data; another checks if the data is fake; another translates protein "languages" into math; another designs a mathematical model; and the last one runs the tests.
• The Goal: To look at two proteins and say, "I’m 87% sure these two are going to shake hands."
• The Result: It was incredibly accurate at predicting how human proteins interact and how viruses "trick" human proteins into interacting.

2. The "Rule-Maker" Detective (The Philosopher)

The second AI platform is different. It doesn't just want to guess if a handshake will happen; it wants to know WHY.

• How it works: This detective doesn't just look at patterns; it looks for Laws of Nature. It looks at where the proteins live (are they in the same room?), what they look like (do they have similar shapes?), and their history.
• The Goal: To write down simple, human-readable rules, like: "If two workers are in the same room and wearing similar uniforms, they are likely to work together."
• The Result: It discovered that human proteins usually interact if they are in the same part of the cell and share similar "family traits." For viruses, it found the rules are much more complex, involving how the virus mimics human "uniforms" to blend in.

---

Why is this a big deal? (The "Double-Check")

The most brilliant part of this study is the Cross-Check.

Imagine you have a student who is great at multiple-choice tests (The Predictor) but struggles to explain their answers. Then you have another student who is great at writing essays (The Rule-Maker). If both students arrive at the exact same conclusion using different methods, you can be much more confident that the answer is correct.

The researchers found that the "Pattern Matcher" and the "Rule-Maker" agreed! This proves the AI wasn't just "guessing" or "cheating" by looking at shortcuts; it was actually learning the real biological laws of how life works.

The Bottom Line

We have moved from using AI as a simple tool to using AI as an autonomous research partner. This "Digital Scientist" can help us map the microscopic world faster than ever, potentially leading to new vaccines and treatments by understanding the secret handshakes of life and death.

Technical Summary

Technical Summary: Agentic AI Platforms for Autonomous Training and Rule Induction of PPI

1. Problem Statement

Protein-Protein Interactions (PPI) are fundamental to biological processes, medical countermeasures, and host-pathogen research. However, predicting PPIs—specifically human-human and virus-human interactions—is hindered by two major challenges:

• Data Scarcity and Imbalance: There is a significant lack of experimentally confirmed negative data labels and an incomplete set of positive datasets.
• Model Interpretability: Traditional Machine Learning (ML) models often act as "black boxes," making it difficult to distinguish whether a model has learned genuine biological laws or is merely exploiting data artifacts (data leakage).

2. Methodology

The researchers employed a "coding agent" (based on ChatGPT 5.2 and 5.4 Thinking) to architect two distinct Agentic AI platforms. These platforms mimic the workflow of human scientists, moving from data planning to execution and explanation.

A. Platform 1: Autonomous Training of Predictive ML Models
This platform consists of five specialized AI agents:

1. Data Collector: Harvests positive data from databases (IntAct, BioGRID, HVIDB, etc.) and synthesizes negative data using compartment-based exclusion and Negatome retrieval. It implements a three-way protein-disjoint split (70:10:20) to prevent data leakage.
2. Data Verifier: Checks for hallucinations, class imbalance, and potential data leakage.
3. Feature Embedder: Converts amino acid sequences into multimodal representations using Protein Language Models (PLMs) (e.g., ESM-2) concatenated with physicochemical autocovariance descriptors (Sandberg z-scale and Eisenberg hydrophobicity).
4. Model Designer: Explores various architectures, including MLPs, Transformers (pair, token, and cross), and Two-Tower models.
5. Executor: Conducts training using the AdamW optimizer and performs validation using metrics like MCC (Matthews Correlation Coefficient). It also applies SHAP (Shapley Additive Explanations) to interpret feature importance.

B. Platform 2: Rule Induction for Interpretability
To cross-check the ML models, a second platform was built to replace the "Model Designer" and "Executor" with a Rule Induction Agent.

• Instead of gradient-based learning, this agent operates on stable scalar signals (e.g., cosine similarity, compartment overlap, pathway/domain Jaccard indices, and graph-neighborhood contexts).
• It employs two search strategies: Greedy forward selection (for compact rules) and Sparse Logistic Rule Induction (for complex, multi-component rules).

3. Key Contributions

• Agentic Orchestration: Demonstrates that AI agents can autonomously manage the entire ML lifecycle, from raw data curation to complex model validation.
• Dual-Track Validation: Introduces a novel framework where predictive ML models are cross-checked against explicit, human-readable biological rules to ensure scientific validity.
• Robust Evaluation Framework: Implements stringent safeguards, such as protein-disjoint cross-folding, temperature scaling, and hard-negative mining, to ensure generalizability to unseen proteins.

4. Results

• Predictive Performance:
  • Human-Human PPI: The ensemble achieved 87.3% accuracy and a 74.6% MCC.
  • Virus-Human PPI: Despite noisier data, the model achieved 86.5% accuracy and a 73.0% MCC.
• Rule Induction & Interpretability:
  • Human-Human Rules: The agent induced a highly compact two-rule system: interactions are favored when proteins are not in disjoint compartments and when they share domain/family organizations.
  • Virus-Human Rules: The agent induced a more complex set of 60 weighted rules, highlighting that virus-human interaction is a "host-protein problem" dominated by human protein localization and feature-space proximity.
• Convergence: SHAP analysis of the ML models aligned with the rules induced by the second platform, confirming that the models learned biological regularities (like spatial co-accessibility and domain compatibility) rather than artifacts.

5. Significance

This work marks a significant step toward autonomous scientific discovery. By coupling predictive power with mechanistic interpretability, the authors demonstrate that Agentic AI can not only predict biological outcomes but also "understand" and explain the underlying biological laws. This framework provides a reproducible and scalable blueprint for applying AI to complex biological domains, such as drug discovery and pathogen research, where transparency and data integrity are paramount.