Increasing the greediness of generative flow networks through action-values
Bibliographic record
Abstract
In recent years, deep learning has emerged as a highly promising tool in drug discovery, offering significant reductions in both time and cost associated with the process.Specifically, a novel generative method known as Generative Flow Networks (GFlowNets) has demonstrated promising results by showcasing its ability to generate a diverse pool of candidates for small molecule generation tasks.Recognizing the potential of GFlowNets, it is essential to investigate and improve their ability to generate high reward and diverse samples, which are crucial for drug development process.This thesis proposes an approach to address this challenge in complex and multi-dimensional scenarios, namely Q-learning GFlowNets (QGFN) -an approach which allows to control the greediness of a GFlowNet by leveraging the well-known reinforcement learning (RL) technique of action values.QGFN increases exploration diversity while preserving the ability to generate high-reward candidates.Empirical results show that QGFN effectively generates high-reward samples in a variety of tasks without sacrificing diversity.This work demonstrates significant improvements that can translate into practical applications in the field of drug discovery, contributing to its advancement.i Abrégé Ces dernières années, l'apprentissage profond est apparu comme un outil très prometteur dans la découverte de médicaments, offrant des réductions significatives du temps et des coûts associés au processus.Plus précisément, une nouvelle méthode générative connue sous le nom de Generative Flow Networks (GFlowNets) a démontré des résultats prometteurs en démontrant sa capacité à générer un pool diversifié de candidats pour des tâches de génération de petites molécules.Reconnaissant le potentiel des GFlowNets, il est essentiel d'étudier et d'améliorer leurs performances, leur rendement élevé et leurs échantillons diversifiés, ce qui est crucial pour le processus de développement de médicaments.Cette thèse propose une approche pour relever ce défi dans des scénarios complexes et multidimensionnels, à savoir QGFN -Controllable Greediness with Action Values.Les QGFN sont conçus pour accroître la diversité de l'exploration tout en préservant les candidats les plus rémunérateurs.Avec QGFN, nous introduisons la combinaison du réseau Q avec QGFN comme une politique plus gourmande pour échantillonner des trajectoires, avec une gourmandise contrôlée par ce réseau Q. Les résultats empiriques montrent que QGFN génère efficacement des échantillons très rémunérateurs dans diverses tâches sans sacrifier la diversité.De plus, ces travaux démontrent des améliorations significatives qui peuvent se traduire par des applications pratiques dans le domaine de la découverte de médicaments, contribuant ainsi à son avancement.List of Tables 5.1 Fragment-based molecule task: Reward and Diversity at inference after training. . . . . . . . . . . . . .
Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.
How this classification was reachedexpand
Full frame machine prediction
Teacher imitationNot calibrated prevalence, not ground truth. Human validation pending. The Gemma side is a direct model label for every work in the frame, read from the title-only record. The Codex side is a classifier learned from the 10,348 direct Codex labels and calibrated to design-weighted sample rates; fields without enough sample support carry no Codex call. Candidate is the union of the two sides; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels.
Distilled classifier scores by category (both heads)
| Category | Codex | Gemma |
|---|---|---|
| Metaresearch | 0.001 | 0.006 |
| Meta-epidemiology (narrow) | 0.001 | 0.000 |
| Meta-epidemiology (broad) | 0.001 | 0.001 |
| Bibliometrics | 0.001 | 0.000 |
| Science and technology studies | 0.000 | 0.001 |
| Scholarly communication | 0.001 | 0.002 |
| Open science | 0.001 | 0.002 |
| Research integrity | 0.001 | 0.002 |
| Insufficient payload (model declined to judge) | 0.009 | 0.001 |
Machine scores (provisional)
The two teacher heads of the student model, read on this work. A score orders the frame for review; it never asserts a category, and the validation status ships verbatim with every row.
Baseline scores from an immature model (maturity gate not passed, 7 training rounds). Scores rank; they never assert a category.
score_only:v0-immature-baseline · verbatim from the scoring run: score_only means the number may rank works, and no category label ships from itClassification
machine, unvalidatedMachine predicted; a candidate call from one source (direct Gemma or distilled Codex), not a consensus.
How this classification was reached, model by model and score by score, is at the end of the page under "How this classification was reached".