|
|
<!DOCTYPE html> |
|
|
<html lang="en"> |
|
|
|
|
|
<head> |
|
|
<meta charset="utf-8"> |
|
|
<meta http-equiv="X-UA-Compatible" content="IE=edge"> |
|
|
<meta name="viewport" content="width=device-width, initial-scale=1"> |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
<title>Molecule Generation with Fragment Retrieval Augmentation</title> |
|
|
<meta name="description" content=""> |
|
|
|
|
|
|
|
|
<link rel="stylesheet" href="assets/main.css"> |
|
|
<link rel="canonical" href="index.html"> |
|
|
|
|
|
|
|
|
<link rel="alternate" type="application/rss+xml" title="Seul Lee" href="https://f-rag.github.io/feed.xml"> |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
<meta property="og:title" content="Molecule Generation with Fragment Retrieval Augmentation"> |
|
|
<meta property="og:site_name" content="Seul Lee"> |
|
|
<meta property="og:url" content="https://f-rag.github.io/"> |
|
|
<meta property="og:description" content=""> |
|
|
|
|
|
|
|
|
<meta name="twitter:card" content="summary"> |
|
|
|
|
|
<meta name="twitter:title" content="Molecule Generation with Fragment Retrieval Augmentation"> |
|
|
<meta name="twitter:description" content=""> |
|
|
|
|
|
|
|
|
|
|
|
<link rel="dns-prefetch" href="https://fonts.gstatic.com"> |
|
|
<link rel="preconnect" href="https://fonts.gstatic.com" crossorigin> |
|
|
<link href="https://fonts.googleapis.com/css2?family=Bitter:ital,wght@0,400;0,700;1,400&display=swap" rel="stylesheet"> |
|
|
|
|
|
|
|
|
|
|
|
</head> |
|
|
|
|
|
|
|
|
<body> |
|
|
|
|
|
<header class="site-header"> |
|
|
|
|
|
<div class="wrapper"> |
|
|
|
|
|
|
|
|
|
|
|
<nav class="site-nav"> |
|
|
|
|
|
</nav> |
|
|
|
|
|
</div> |
|
|
|
|
|
</header> |
|
|
|
|
|
|
|
|
<main class="page-content" aria-label="Content"> |
|
|
<div class="wrapper"> |
|
|
<article class="post" itemscope itemtype="http://schema.org/BlogPosting"> |
|
|
|
|
|
<header class="post-header"> |
|
|
|
|
|
<h1 class="post-title" itemprop="name headline">Molecule Generation with Fragment Retrieval Augmentation</h1> |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
</header> |
|
|
|
|
|
<div class="post-content" itemprop="articleBody"> |
|
|
<p style="text-align:center"> |
|
|
<span style="font-weight:bold"> |
|
|
Conference on Neural Information Processing Systems (NeurIPS), 2024<br /> |
|
|
</span> |
|
|
Seul Lee<sup>1*</sup> |
|
|
Karsten Kreis<sup>2</sup> |
|
|
Srimukh Prasad Veccham<sup>2</sup> |
|
|
Meng Liu<sup>2</sup><br /> |
|
|
Danny Reidenbach<sup>2</sup> |
|
|
Saee Paliwal<sup>2</sup> |
|
|
Arash Vahdat<sup>2†</sup> |
|
|
Weili Nie<sup>2†</sup><br /> |
|
|
<span style="color:gray"> |
|
|
1 <img height="25px" alt="KAIST" src="assets/2024-09-17-f-rag/kaist.svg" /> |
|
|
2 <img height="20px" alt="NVIDIA" src="assets/2024-09-17-f-rag/nvidia.svg" /><br /> |
|
|
* Work done during an internship at NVIDIA |
|
|
† Equal advising<br /> |
|
|
seul.lee@kaist.ac.kr<br /> |
|
|
{kkreis, sveccham, menliu, dreidenbach, saeep, avahdat, wnie}@nvidia.com<br /> |
|
|
</span> |
|
|
<a href="https://arxiv.org/abs/2411.12078" style="color:teal">[Paper]</a> |
|
|
</p> |
|
|
|
|
|
<p><br /></p> |
|
|
|
|
|
<p align="center"><img width="800px" alt="f-RAG" src="assets/2024-09-17-f-rag/concept.png" /></p> |
|
|
|
|
|
<h3><em>f</em>-RAG: Fragment Retrieval-Augmented Generation</h3> |
|
|
<hr /> |
|
|
<p>Many fragment-based molecule generation methods show limited exploration beyond the existing fragments in the database as they only reassemble or slightly modify the given ones. To tackle this problem, we propose a new fragment-based molecule generation framework with retrieval augmentation, namely <strong>Fragment Retrieval-Augmented Generation (<em>f</em>-RAG)</strong>. <em>f</em>-RAG is based on a pre-trained molecular generative model (<a href="https://arxiv.org/abs/2310.10773"><span style="color:teal">SAFE-GPT</span></a>) that proposes additional fragments from input fragments to complete and generate a new molecule. Given a fragment vocabulary, <em>f</em>-RAG retrieves two types of fragments: (1) <strong>hard fragments</strong>, which serve as building blocks that will be explicitly included in the newly generated molecule, and (2) <strong>soft fragments</strong>, which serve as reference to guide the generation of new fragments through a trainable <strong>fragment injection module</strong>. To extrapolate beyond the existing fragments, <em>f</em>-RAG updates the fragment vocabulary with generated fragments via an iterative refinement process which is further enhanced with post-hoc genetic fragment modification. <em>f</em>-RAG can achieve an improved exploration-exploitation trade-off by maintaining a pool of fragments and expanding it with novel and high-quality fragments through a strong generative prior.</p> |
|
|
|
|
|
<p><br /></p> |
|
|
|
|
|
<h3>Fragment-based Drug Discovery with <em>f</em>-RAG</h3> |
|
|
<hr /> |
|
|
<p>On the <a href="https://arxiv.org/abs/2206.12411"><span style="color:teal">PMO benchmark</span></a>, <em>f</em>-RAG outperforms the previous methods in terms of the sum of the AUC top-10 values and achieves the highest AUC top-10 values in 12 out of 23 tasks. Furthermore, even though the essential considerations in drug discovery (e.g., diversity, novelty, and synthesizability) often conflict with each other, <em>f</em>-RAG exhibits the best balance across them, demonstrating its applicability as a promising tool for drug discovery.</p> |
|
|
|
|
|
<p align="center"><img width="700px" src="assets/2024-09-17-f-rag/table.png" /></p> |
|
|
<p align="center"><img width="100%" src="assets/2024-09-17-f-rag/tradeoff.png" /></p> |
|
|
|
|
|
<p><br /></p> |
|
|
|
|
|
<h3>Evolution of Generated Molecules over Iterations</h3> |
|
|
<hr /> |
|
|
<p><em>f</em>-RAG dynamically updates the fragment vocabulary with newly generated fragments. The fragment vocabulary and therefore the generated molecules are iteratively refined throughout generation.</p> |
|
|
|
|
|
<p align="center"> |
|
|
<img width="300px" src="assets/2024-09-17-f-rag/frag.gif" /> |
|
|
<img width="300px" src="assets/2024-09-17-f-rag/frag.png" /> |
|
|
<br /> |
|
|
<img width="300px" src="assets/2024-09-17-f-rag/mol.gif" /> |
|
|
<img width="300px" src="assets/2024-09-17-f-rag/mol.png" /> |
|
|
</p> |
|
|
|
|
|
<p><br /></p> |
|
|
|
|
|
<h3>Exploration with Dynamic Vocabulary Update</h3> |
|
|
<hr /> |
|
|
<p><em>f</em>-RAG effectively improves the exploration-exploitation trade-off in drug discovery by utilizing existing fragments while dynamically updating the fragment vocabulary with newly proposed fragments. <em>f</em>-RAG can discover molecules that have better target properties than the top molecule in the training set with the dynamic update.</p> |
|
|
|
|
|
<p align="center"><img width="250px" src="assets/2024-09-17-f-rag/kde.png" /></p> |
|
|
|
|
|
<h3>Citation</h3> |
|
|
<hr /> |
|
|
<div class="language-plaintext highlighter-rouge"><div class="highlight"><pre class="highlight"><code>@article{lee2024frag, |
|
|
title = {Molecule Generation with Fragment Retrieval Augmentation}, |
|
|
author = {Seul Lee and Karsten Kreis and Srimukh Prasad Veccham and Meng Liu and |
|
|
Danny Reidenbach and Saee Paliwal and Arash Vahdat and Weili Nie}, |
|
|
journal = {Advances in Neural Information Processing Systems}, |
|
|
year = {2024} |
|
|
} |
|
|
</code></pre></div></div> |
|
|
|
|
|
</div> |
|
|
|
|
|
|
|
|
|
|
|
</article> |
|
|
|
|
|
</div> |
|
|
</main> |
|
|
|
|
|
<footer class="site-footer"> |
|
|
|
|
|
<div class="wrapper"> |
|
|
|
|
|
<p> |
|
|
|
|
|
|
|
|
© Seul Lee</a> |
|
|
|
|
|
</p> |
|
|
|
|
|
</div> |
|
|
|
|
|
</footer> |
|
|
|
|
|
|
|
|
</body> |
|
|
|
|
|
</html> |
|
|
|
