iiinova — where your digital self keeps evolving

Amin Emad

@aminmemad.bsky.social

Associate Professor at #McGill University, Director of Computational Biology and Artificial Intelligence (COMBINE) lab (www.ece.mcgill.ca/~aemad2) #Mila, Quebec AI Institute

Posts Replies Media

Amin Emad @aminmemad.bsky.social · 13/02/2026

8/ 📖 Read the full paper 🔗 rdcu.be/e3PGD

rdcu.be

A flaw in using pretrained protein language models in protein–protein interaction inference models

Nature Machine Intelligence - The usage of pretrained protein language models (pLMs) is rapidly growing. However, Szymborski and Emad find that pretrained pLMs can be a source of data leakage in...

000

Amin Emad @aminmemad.bsky.social · 13/02/2026

7/ 🔧 What’s next? We propose stricter evaluation pipelines for any paired biological dataset and highlight the need for pLMs that truly understand protein context and generalize to out-of-distribution proteins and tasks.

100

Amin Emad @aminmemad.bsky.social · 13/02/2026

6/ 📚 Take home messages 1. Pretrained pLMs can leak data → over optimistic PPI scores. 2. Leakage is task specific; non paired tasks may stay unaffected. 3. SqueezeProt is a new pLM that is 5x faster, no test in train. 4. Current models still fail on some real-world tasks.

100

Amin Emad @aminmemad.bsky.social · 13/02/2026

5/ 🌍 Real world generalization fails. Both pLM based and classic models stumble on: - Human–SARS CoV 2 interaction predictions - Point mutation effects on binding affinity

110

Amin Emad @aminmemad.bsky.social · 13/02/2026

4/ 🧬 Does leakage affect other tasks? When we tested the same pLMs on protein keyword annotation (a non paired task), performance stayed stable. Leakage might be specific to paired sequence problems like PPI.

100

Amin Emad @aminmemad.bsky.social · 13/02/2026

3/ 🐇 New ultra-fast pLM with no data leakage! To run these experiments we created SqueezeProt, a new pLM which is 5x faster than the competition with similar performance.

100

Amin Emad @aminmemad.bsky.social · 13/02/2026

2/ 🚨 Data leakage discovered! We found that existing pretrained pLMs unintentionally leak information from training data into downstream PPI tasks, inflating test scores. The gap can be >10 % MCC!

110

Amin Emad @aminmemad.bsky.social · 13/02/2026

🧵 1/ 🎉New paper alert! Pretrained protein language models (#pLMs) are all the hype, but are they really helping us predict protein- protein interactions? 🤔Dive into our thread to see why you should read the full study @natmachintell.nature.com. ⬇️ 🔗 rdcu.be/e3PGD

133