BioML Seminar 4.5 - Antoine Koehl on Deep Models of Protein Evolution in Time

BioML Seminar 4.5 - Antoine Koehl on Deep Models of Protein Evolution in Time

by Antoine Koehl

Life & Natural Systems Biology PSALife sciences: general issuesPSAXComputational biology
🎙 Antoine Koehl 👥 14K 📅 June 26, 2026 ⏱ 54 min 👁 137 📄 original study 🧭 2026-08-16
Available in: English (current) Français

Keywords

protein evolutiondeep learningphylogeneticsgenerative modelsPEINT

Summary

Antoine Koehl presents PEINT, a deep learning framework for modeling protein evolution over time. He begins by motivating the need for realistic evolutionary models, highlighting limitations of classical approaches that assume independent sites. He reviews the history of protein evolution models, from Dayhoff’s PAM matrices to maximum likelihood methods like WAG, and introduces the CherryML framework that simplifies likelihood computation by pairing sequences. PEINT uses a transformer architecture with a pre-trained protein language model (ESM-2) to learn conditional transition probabilities between sequences given evolutionary time. The model is trained on cherries extracted from TrRosetta alignments, and evaluated on held-out data. Results show that PEINT outperforms classical models like WAG and LG in terms of per-site likelihood, especially at short evolutionary distances. The talk concludes with potential applications in phylogenetic inference and protein engineering.

133 words

Critical Evaluation

Value of the Information & Strength of the Argument

The talk provides a clear and well-structured argument for the development of PEINT. It systematically identifies the limitations of existing models, particularly the independent-sites assumption, and explains how CherryML overcomes computational barriers. The use of a transformer with a pre-trained encoder is well-justified, leveraging advances in protein language models. The evaluation against classical models demonstrates the added value of the approach. The speaker’s background in structural biology adds credibility to the discussion of protein evolution.

Scientific Rigor, Source Quality, Title Accuracy

The presentation is scientifically rigorous, with proper attribution to prior work (Dayhoff, JTT, WAG, LG, CherryML). The methodology is clearly described, including data preparation, model architecture, and evaluation. The title accurately reflects the content. No external sources are cited beyond the speaker’s own work and the mentioned models, but the description provides no additional references.

145 words

Title / Content Match

The title accurately reflects the content: a seminar on deep models of protein evolution over time.

Quality & Reliability

8/10

Presentation of original research with clear methodology, references to established models (Dayhoff, JTT, WAG, LG), and use of standard datasets (TrRosetta). Speaker is a postdoc at UC Berkeley with relevant expertise. Limitations and assumptions are discussed.

Key Moments

Markers derived by PSI from the transcript: the creator did not define chapters.

Cited Sources

  • TrRosetta — Dataset of multiple sequence alignments used for training and evaluation.
  • ESM-2 — Pre-trained protein language model used as encoder.

Concurring Sources

  • CherryML — The CherryML framework that PEINT builds upon.

Contribution & Novelties

PEINT introduces a novel deep learning approach to model protein evolution that captures site interactions, overcoming limitations of classical independent-sites models. It leverages the CherryML framework for tractable likelihood computation and uses a transformer with a pre-trained protein language model. This enables realistic simulation of evolutionary trajectories and has potential applications in phylogenetic inference and protein engineering.

Pour aller plus loin :

  • Protein language models — Overview of protein language models.
  • Phylogenetics — Background on phylogenetic inference.
  • Generative models — General concept of generative models.

85 words

Radar Profile

The radar profile shows high scores across all dimensions, indicating a well-balanced and reliable presentation. The talk is technically deep, provides substantial information, and is scientifically rigorous.

Reliability 8/10

💬 No comments were provided for analysis.