LLPS datasets & benchmarking: bringing order to condensate chaos 🧬📊

We present curated datasets of proteins involved in liquid–liquid phase separation (LLPS), enabling robust benchmarking and improved machine learning predictions in condensate biology.

Keywords

LLPS, liquid-liquid phase separation, biomolecular condensates, protein datasets, machine learning, bioinformatics, benchmarking

---

📌 Project highlights

• 🧬 First integrated datasets of LLPS proteins (drivers, clients, negatives)
  
• ⚖️ Introduces standardized negative datasets (structured + disordered proteins)
  
• 📊 Benchmark of 16 LLPS prediction tools - the most comprehensive to date
  
• 🧠 Reveals limitations of current ML models for phase separation
  
• 🌐 Public resource: datasets + website for reproducible research

---

🎉 New paper out! And this one is a big one, not just another tool, but a foundation for the whole field 😄

👉 Comprehensive protein datasets and benchmarking for liquid–liquid phase separation studies

---

🔗 Data & resources

This is not just a paper, it’s a community resource for building and benchmarking LLPS models.

• 🌐 Dataset website
  
• 💻 GitHub

---

🎧 Audio summary

Not everyone wants to read about κ parameters and sticker–spacer distributions over coffee ☕😄

👉 We’ve attached a short audio explanation 🎧 to make things easier.

🔬 What is this about?

Liquid–liquid phase separation (LLPS) is a process where proteins form biomolecular condensates, membrane-less compartments that organize cellular processes.

These condensates are crucial for:

• gene regulation
  
• stress response
  
• disease mechanisms (e.g. neurodegeneration)

But studying LLPS computationally has a major problem:

👉 data is messy, inconsistent, and fragmented across databases

---

🚨 The problem we tackled

Existing LLPS databases:

• use different definitions
  
• contain inconsistent annotations
  
• lack reliable negative examples

👉 This makes:

• machine learning unreliable
  
• benchmarking unfair
  
• results hard to compare

---

🧠 What we built

We created high-confidence datasets of proteins involved in LLPS, based on:

🧬 Protein roles

• Drivers → form condensates independently
  
• Clients → join existing condensates
  
• Negatives → proteins not associated with LLPS

👉 Importantly, we introduced two types of negative datasets:

• structured proteins (PDB-like)
  
• disordered proteins (DisProt-like)

This is critical because: 👉 disorder alone ≠ LLPS (big source of bias!)

---

⚙️ How it works

We used:

• 📚 integration of multiple LLPS databases
  
• 🧹 strict filtering criteria for data quality
  
• 🧬 annotation of:
  • disorder
    
  • prion-like domains
    
  • physicochemical features

Then we:

👉 analyzed protein properties
👉 and benchmarked prediction tools

---

📊 Benchmarking LLPS predictors

We evaluated 16 bioinformatics tools on independent datasets.

Key findings:

• 🤖 ML-based tools perform best
  
• ⚠️ Many tools still have high false positive rates
  
• 🧠 Models often confuse:
  • disorder
    
  • with actual phase separation

👉 Translation: we are still not great at predicting LLPS reliably 😄

---

🧬 Key biological insights

From the analysis:

• 🧩 LLPS is multi-factorial, not driven by a single feature
  
• ⚡ properties like:
  • charge distribution
    
  • aggregation propensity
    
  • interaction regions
    matter more than simple sequence features
    
• 🧠 drivers and clients are different, but subtle

👉 And importantly: you cannot model LLPS properly without good negative data

---

🚀 Why this matters

This work:

• 📊 sets a new standard for benchmarking
  
• 🧠 improves machine learning reliability
  
• 🔬 enables better biological interpretation of condensates

👉 In short: we finally have a clean dataset to stop guessing and start learning properly

---

💚 BioGenies context

This project connects directly to our interests in:

• amyloids 🧬
  
• protein aggregation 🧊
  
• phase separation ⚡

👉 and gives us a solid foundation for future tools and models