Recent advances in large language models (LLMs) have shown great potential to accelerate drug discovery. However, the specialized nature of biochemical data often necessitates costly domain-specific fine-tuning, posing critical challenges. First, it hinders the application of more flexible general-purpose LLMs in cutting-edge drug discovery tasks. More importantly, it impedes the rapid integration of the vast amounts of scientific data continuously generated through experiments and research. To investigate these challenges, we propose CLADD, a retrieval-augmented generation (RAG)-empowered agentic system tailored to drug discovery tasks. Through the collaboration of multiple LLM agents, CLADD dynamically retrieves information from biomedical knowledge bases, contextualizes query molecules, and integrates relevant evidence to generate responses -- all without the need for domain-specific fine-tuning. Crucially, we tackle key obstacles in applying RAG workflows to biochemical data, including data heterogeneity, ambiguity, and multi-source integration. We demonstrate the flexibility and effectiveness of this framework across a variety of drug discovery tasks, showing that it outperforms general-purpose and domain-specific LLMs as well as traditional deep learning approaches.

Diffusion models have shown promising generative capabilities across diverse domains, yet aligning their outputs with desired reward functions remains a challenge, particularly in cases where reward functions are non-differentiable. Some gradient-free guidance methods have been developed, but they often struggle to achieve optimal inference-time alignment. In this work, we newly frame inference-time alignment in diffusion as a search problem and propose Dynamic Search for Diffusion (DSearch), which subsamples from denoising processes and approximates intermediate node rewards. It also dynamically adjusts beam width and tree expansion to efficiently explore high-reward generations. To refine intermediate decisions, DSearch incorporates adaptive scheduling based on noise levels and a lookahead heuristic function. We validate DSearch across multiple domains, including biological sequence design, molecular optimization, and image generation, demonstrating superior reward optimization compared to existing approaches.

Predicting transcriptional responses to novel drugs provides a unique opportunity to accelerate biomedical research and advance drug discovery efforts. However, the inherent complexity and high dimensionality of cellular responses, combined with the extremely limited available experimental data, makes the task challenging. In this study, we leverage single-cell foundation models (FMs) pre-trained on tens of millions of single cells, encompassing multiple cell types, states, and disease annotations, to address molecular perturbation prediction. We introduce a drug-conditional adapter that allows efficient fine-tuning by training less than 1% of the original foundation model, thus enabling molecular conditioning while preserving the rich biological representation learned during pre-training. The proposed strategy allows not only the prediction of cellular responses to novel drugs, but also the zero-shot generalization to unseen cell lines. We establish a robust evaluation framework to assess model performance across different generalization tasks, demonstrating state-of-the-art results across all settings, with significant improvements in the few-shot and zero-shot generalization to new cell lines compared to existing baselines.

Bridging biomolecular modeling with natural language information, particularly through large language models (LLMs), has recently emerged as a promising interdisciplinary research area. LLMs, having been trained on large corpora of scientific documents, demonstrate significant potential in understanding and reasoning about biomolecules by providing enriched contextual and domain knowledge. However, the extent to which LLM-driven insights can improve performance on complex predictive tasks (e.g., toxicity) remains unclear. Further, the extent to which relevant knowledge can be extracted from LLMs also remains unknown. In this study, we present Molecule Caption Arena: the first comprehensive benchmark of LLM-augmented molecular property prediction. We evaluate over twenty LLMs, including both general-purpose and domain-specific molecule captioners, across diverse prediction tasks. To this goal, we introduce a novel, battle-based rating system. Our findings confirm the ability of LLM-extracted knowledge to enhance state-of-the-art molecular representations, with notable model-, prompt-, and dataset-specific variations. Code, resources, and data are available at github.com/Genentech/molcap-arena.

Deep neural networks excel in mapping genomic DNA sequences to associated readouts (e.g., protein-DNA binding). Beyond prediction, the goal of these networks is to reveal to scientists the underlying motifs (and their syntax) which drive genome regulation. Traditional methods that extract motifs from convolutional filters suffer from the uninterpretable dispersion of information across filters and layers. Other methods which rely on importance scores can be unstable and unreliable. Instead, we designed a novel mechanistically interpretable architecture for regulatory genomics, where motifs and their syntax are directly encoded and readable from the learned weights and activations. We provide theoretical and empirical evidence of our architecture's full expressivity, while still being highly interpretable. Through several experiments, we show that our architecture excels in de novo motif discovery and motif instance calling, is robust to variable sequence contexts, and enables fully interpretable generation of novel functional sequences.

Diffusion models are powerful generative models that allow for precise control over the characteristics of the generated samples. While these diffusion models trained on large datasets have achieved success, there is often a need to introduce additional controls in downstream fine-tuning processes, treating these powerful models as pre-trained diffusion models. This work presents a novel method based on reinforcement learning (RL) to add additional controls, leveraging an offline dataset comprising inputs and corresponding labels. We formulate this task as an RL problem, with the classifier learned from the offline dataset and the KL divergence against pre-trained models serving as the reward functions. We introduce our method, $\textbf{CTRL}$ ($\textbf{C}$onditioning pre-$\textbf{T}$rained diffusion models with $\textbf{R}$einforcement $\textbf{L}$earning), which produces soft-optimal policies that maximize the abovementioned reward functions. We formally demonstrate that our method enables sampling from the conditional distribution conditioned on additional controls during inference. Our RL-based approach offers several advantages over existing methods. Compared to commonly used classifier-free guidance, our approach improves sample efficiency, and can greatly simplify offline dataset construction by exploiting conditional independence between the inputs and additional controls. Furthermore, unlike classifier guidance, we avoid the need to train classifiers from intermediate states to additional controls.

AI-driven design problems, such as DNA/protein sequence design, are commonly tackled from two angles: generative modeling, which efficiently captures the feasible design space (e.g., natural images or biological sequences), and model-based optimization, which utilizes reward models for extrapolation. To combine the strengths of both approaches, we adopt a hybrid method that fine-tunes cutting-edge diffusion models by optimizing reward models through RL. Although prior work has explored similar avenues, they primarily focus on scenarios where accurate reward models are accessible. In contrast, we concentrate on an offline setting where a reward model is unknown, and we must learn from static offline datasets, a common scenario in scientific domains. In offline scenarios, existing approaches tend to suffer from overoptimization, as they may be misled by the reward model in out-of-distribution regions. To address this, we introduce a conservative fine-tuning approach, BRAID, by optimizing a conservative reward model, which includes additional penalization outside of offline data distributions. Through empirical and theoretical analysis, we demonstrate the capability of our approach to outperform the best designs in offline data, leveraging the extrapolation capabilities of reward models while avoiding the generation of invalid designs through pre-trained diffusion models.

Diffusion models excel at capturing complex data distributions, such as those of natural images and proteins. While diffusion models are trained to represent the distribution in the training dataset, we often are more concerned with other properties, such as the aesthetic quality of the generated images or the functional properties of generated proteins. Diffusion models can be finetuned in a goal-directed way by maximizing the value of some reward function (e.g., the aesthetic quality of an image). However, these approaches may lead to reduced sample diversity, significant deviations from the training data distribution, and even poor sample quality due to the exploitation of an imperfect reward function. The last issue often occurs when the reward function is a learned model meant to approximate a ground-truth "genuine" reward, as is the case in many practical applications. These challenges, collectively termed "reward collapse," pose a substantial obstacle. To address this reward collapse, we frame the finetuning problem as entropy-regularized control against the pretrained diffusion model, i.e., directly optimizing entropy-enhanced rewards with neural SDEs. We present theoretical and empirical evidence that demonstrates our framework is capable of efficiently generating diverse samples with high genuine rewards, mitigating the overoptimization of imperfect reward models.

Diffusion models excel at modeling complex data distributions, including those of images, proteins, and small molecules. However, in many cases, our goal is to model parts of the distribution that maximize certain properties: for example, we may want to generate images with high aesthetic quality, or molecules with high bioactivity. It is natural to frame this as a reinforcement learning (RL) problem, in which the objective is to fine-tune a diffusion model to maximize a reward function that corresponds to some property. Even with access to online queries of the ground-truth reward function, efficiently discovering high-reward samples can be challenging: they might have a low probability in the initial distribution, and there might be many infeasible samples that do not even have a well-defined reward (e.g., unnatural images or physically impossible molecules). In this work, we propose a novel reinforcement learning procedure that efficiently explores on the manifold of feasible samples. We present a theoretical analysis providing a regret guarantee, as well as empirical validation across three domains: images, biological sequences, and molecules.

Uncertainty estimation is critical in high-stakes machine learning applications. One effective way to estimate uncertainty is conformal prediction, which can provide predictive inference with statistical coverage guarantees. We present a new conformal regression method, Spline Prediction Intervals via Conformal Estimation (SPICE), that estimates the conditional density using neural-network-parameterized splines. We prove universal approximation and optimality results for SPICE, which are empirically validated by our experiments. SPICE is compatible with two different efficient-to-compute conformal scores, one oracle-optimal for marginal coverage (SPICE-ND) and the other asymptotically optimal for conditional coverage (SPICE-HPD). Results on benchmark datasets demonstrate SPICE-ND models achieve the smallest average prediction set sizes, including average size reductions of nearly 50% for some datasets compared to the next best baseline. SPICE-HPD models achieve the best conditional coverage compared to baselines. The SPICE implementation is made available.