Albert Bou

Albert Bou, Matteo Bettini, Sebastian Dittert et al.

PyTorch has ascended as a premier machine learning framework, yet it lacks a native and comprehensive library for decision and control tasks suitable for large development teams dealing with complex real-world data and environments. To address this issue, we propose TorchRL, a generalistic control library for PyTorch that provides well-integrated, yet standalone components. We introduce a new and flexible PyTorch primitive, the TensorDict, which facilitates streamlined algorithm development across the many branches of Reinforcement Learning (RL) and control. We provide a detailed description of the building blocks and an extensive overview of the library across domains and tasks. Finally, we experimentally demonstrate its reliability and flexibility and show comparative benchmarks to demonstrate its computational efficiency. TorchRL fosters long-term support and is publicly available on GitHub for greater reproducibility and collaboration within the research community. The code is open-sourced on GitHub.

Jon M Laurent, Albert Bou, Michael Pieler et al.

Optimism for accelerating scientific discovery with AI continues to grow. Current applications of AI in scientific research range from training dedicated foundation models on scientific data to agentic autonomous hypothesis generation systems to AI-driven autonomous labs. The need to measure progress of AI systems in scientific domains correspondingly must not only accelerate, but increasingly shift focus to more real-world capabilities. Beyond rote knowledge and even just reasoning to actually measuring the ability to perform meaningful work. Prior work introduced the Language Agent Biology Benchmark LAB-Bench as an initial attempt at measuring these abilities. Here we introduce an evolution of that benchmark, LABBench2, for measuring real-world capabilities of AI systems performing useful scientific tasks. LABBench2 comprises nearly 1,900 tasks and is, for the most part, a continuation of LAB-Bench, measuring similar capabilities but in more realistic contexts. We evaluate performance of current frontier models, and show that while abilities measured by LAB-Bench and LABBench2 have improved substantially, LABBench2 provides a meaningful jump in difficulty (model-specific accuracy differences range from -26% to -46% across subtasks) and underscores continued room for performance improvement. LABBench2 continues the legacy of LAB-Bench as a de facto benchmark for AI scientific research capabilities and we hope that it continues to help advance development of AI tools for these core research functions. To facilitate community use and development, we provide the task dataset at https://huggingface.co/datasets/futurehouse/labbench2 and a public eval harness at https://github.com/EdisonScientific/labbench2.

Ludovico Mitchener, Angela Yiu, Benjamin Chang et al.

Data-driven scientific discovery requires iterative cycles of literature search, hypothesis generation, and data analysis. Substantial progress has been made towards AI agents that can automate scientific research, but all such agents remain limited in the number of actions they can take before losing coherence, thus limiting the depth of their findings. Here we present Kosmos, an AI scientist that automates data-driven discovery. Given an open-ended objective and a dataset, Kosmos runs for up to 12 hours performing cycles of parallel data analysis, literature search, and hypothesis generation before synthesizing discoveries into scientific reports. Unlike prior systems, Kosmos uses a structured world model to share information between a data analysis agent and a literature search agent. The world model enables Kosmos to coherently pursue the specified objective over 200 agent rollouts, collectively executing an average of 42,000 lines of code and reading 1,500 papers per run. Kosmos cites all statements in its reports with code or primary literature, ensuring its reasoning is traceable. Independent scientists found 79.4% of statements in Kosmos reports to be accurate, and collaborators reported that a single 20-cycle Kosmos run performed the equivalent of 6 months of their own research time on average. Furthermore, collaborators reported that the number of valuable scientific findings generated scales linearly with Kosmos cycles (tested up to 20 cycles). We highlight seven discoveries made by Kosmos that span metabolomics, materials science, neuroscience, and statistical genetics. Three discoveries independently reproduce findings from preprinted or unpublished manuscripts that were not accessed by Kosmos at runtime, while four make novel contributions to the scientific literature.

Nikolai Schapin, Carles Navarro, Albert Bou et al.

Binding affinity optimization is crucial in early-stage drug discovery. While numerous machine learning methods exist for predicting ligand potency, their comparative efficacy remains unclear. This study evaluates the performance of classical tree-based models and advanced neural networks in protein-ligand binding affinity prediction. Our comprehensive benchmarking encompasses 2D models utilizing ligand-only RDKit embeddings and Large Language Model (LLM) ligand representations, as well as 3D neural networks incorporating bound protein-ligand conformations. We assess these models across multiple standard datasets, examining various predictive scenarios including classification, ranking, regression, and active learning. Results indicate that simpler models can surpass more complex ones in specific tasks, while 3D models leveraging structural information become increasingly competitive with larger training datasets containing compounds with labelled affinity data against multiple targets. Pre-trained 3D models, by incorporating protein pocket environments, demonstrate significant advantages in data-scarce scenarios for specific binding pockets. Additionally, LLM pretraining on 2D ligand data enhances complex model performance, providing versatile embeddings that outperform traditional RDKit features in computational efficiency. Finally, we show that combining 2D and 3D model strengths improves active learning outcomes beyond current state-of-the-art approaches. These findings offer valuable insights for optimizing machine learning strategies in drug discovery pipelines.

Siddharth Narayanan, James D. Braza, Ryan-Rhys Griffiths et al.

Solving complex real-world tasks requires cycles of actions and observations. This is particularly true in science, where tasks require many cycles of analysis, tool use, and experimentation. Language agents are promising for automating intellectual tasks in science because they can interact with tools via natural language or code. Yet their flexibility creates conceptual and practical challenges for software implementations, since agents may comprise non-standard components such as internal reasoning, planning, tool usage, as well as the inherent stochasticity of temperature-sampled language models. Here, we introduce Aviary, an extensible gymnasium for language agents. We formalize agents as policies solving language-grounded partially observable Markov decision processes, which we term language decision processes. We then implement five environments, including three challenging scientific environments: (1) manipulating DNA constructs for molecular cloning, (2) answering research questions by accessing scientific literature, and (3) engineering protein stability. These environments were selected for their focus on multi-step reasoning and their relevance to contemporary biology research. Finally, with online training and scaling inference-time compute, we show that language agents backed by open-source, non-frontier LLMs can match and exceed both frontier LLM agents and human experts on multiple tasks at up to 100x lower inference cost.

Albert Bou, Morgan Thomas, Sebastian Dittert et al.

In recent years, reinforcement learning (RL) has emerged as a valuable tool in drug design, offering the potential to propose and optimize molecules with desired properties. However, striking a balance between capabilities, flexibility, reliability, and efficiency remains challenging due to the complexity of advanced RL algorithms and the significant reliance on specialized code. In this work, we introduce ACEGEN, a comprehensive and streamlined toolkit tailored for generative drug design, built using TorchRL, a modern RL library that offers thoroughly tested reusable components. We validate ACEGEN by benchmarking against other published generative modeling algorithms and show comparable or improved performance. We also show examples of ACEGEN applied in multiple drug discovery case studies. ACEGEN is accessible at \url{https://github.com/acellera/acegen-open} and available for use under the MIT license.

Siddharth M. Narayanan, James D. Braza, Ryan-Rhys Griffiths et al.

Reasoning models are large language models that emit a long chain-of-thought before answering, providing both higher accuracy and explicit reasoning for their response. A major question has been whether language model reasoning generalizes beyond mathematics, programming, and logic, where most previous work has focused. We demonstrate that reasoning models can be post-trained for chemistry without additional domain pretraining, and require substantially less data compared to contemporary domain-specific models. We report ether0, a 24B parameter LLM (based on Mistral-Small-24B) that can reason in natural language and respond with chemical structures. This reasoning model was trained with reinforcement learning on 640,730 experimentally-grounded chemistry problems across 375 tasks ranging from synthesizability, to blood-brain barrier permeability, to human receptor activity, to scent. Our model exceeds general-purpose chemistry models, frontier models, and human experts on molecular design tasks. It is also more data efficient relative to specialized models. We anticipate that this method can be applied to train data-efficient language models specialized for tasks across a wide variety of scientific domains.

Morgan Thomas, Albert Bou, Jose Carlos Gómez-Tamayo et al.

Chemical language models, combined with reinforcement learning (RL), have shown significant promise to efficiently traverse large chemical spaces for drug discovery. However, the performance of various RL algorithms and their best practices for practical drug discovery are still unclear. Here, starting from the principles of the REINFORCE algorithm, we investigate the effect of different components from RL theory including experience replay, hill-climbing, baselines to reduce variance, and alternative reward shaping. We propose a new regularization method more aligned to REINFORCE than current standard practices, and demonstrate how RL hyperparameters can be fine-tuned for effectiveness and efficiency. Lastly, we apply our learnings to practical drug discovery by demonstrating enhanced learning efficiency on frontier binding affinity models by using Boltz2 as a reward model. We share our RL models used in the ACEGEN repository, and hope the experiments here act as a guide to researchers applying RL to chemical language models for drug discovery.

Morgan Thomas, Albert Bou, Gianni De Fabritiis

Chemical Language Models (CLMs) leveraging reinforcement learning (RL) have shown promise in de novo molecular design, yet often suffer from mode collapse, limiting their exploration capabilities. Inspired by Test-Time Training (TTT) in large language models, we propose scaling TTT for CLMs to enhance chemical space exploration. We introduce MolExp, a novel benchmark emphasizing the discovery of structurally diverse molecules with similar bioactivity, simulating real-world drug design challenges. Our results demonstrate that scaling TTT by increasing the number of independent RL agents follows a log-linear scaling law, significantly improving exploration efficiency as measured by MolExp. In contrast, increasing TTT training time yields diminishing returns, even with exploration bonuses. We further evaluate cooperative RL strategies to enhance exploration efficiency. These findings provide a scalable framework for generative molecular design, offering insights into optimizing AI-driven drug discovery.

Albert Bou, Sebastian Dittert, Gianni De Fabritiis

Advancing reinforcement learning (RL) requires tools that are flexible enough to easily prototype new methods while avoiding impractically slow experimental turnaround times. To match the first requirement, the most popular RL libraries advocate for highly modular agent composability, which facilitates experimentation and development. To solve challenging environments within reasonable time frames, scaling RL to large sampling and computing resources has proved a successful strategy. However, this capability has been so far difficult to combine with modularity. In this work, we explore design choices to allow agent composability both at a local and distributed level of execution. We propose a versatile approach that allows the definition of RL agents at different scales through independent reusable components. We demonstrate experimentally that our design choices allow us to reproduce classical benchmarks, explore multiple distributed architectures, and solve novel and complex environments while giving full control to the user in the agent definition and training scheme definition. We believe this work can provide useful insights to the next generation of RL libraries.