Introduction

Drug discovery is a complex and costly process, often requiring over a decade and substantial financial investment, reaching approximately 879 million dollars, to bring a single compound from discovery to market [1,2]. This makes early screening and accurate prediction of a drug candidate’s eventual success crucial for optimizing resource allocation. In recent years, computational methods, including classical machine learning and deep learning, have contributed to improve drug discovery processes from structural validation to toxicity and efficacy screening [3]. A notable example is ChemAP [4], which applies knowledge distillation from a teacher model that integrates multi-modal information into a student model capable of predicting drug approval from chemical structures alone. However, such models remain constrained by their limited interpretability, making it challenging to build trust for model-driven decisions [5].

Large language models (LLMs) have recently emerged as transformative tools in artificial intelligence (AI), demonstrating broad capabilities across a wide range of domains through training on massive and diverse datasets [6]. A central advancement that has enhanced their precision and problem-solving capacity is the incorporation of chain-of-thought (CoT) reasoning, which allows LLMs to simulate human-like reasoning and improve the interpretability of their outputs [7]. These reasoning abilities have attracted considerable interest in the field of drug discovery, where the complexity of biological systems and the need for integrative analysis present significant challenges [8]. Recent efforts, including frameworks such as DrugReAlign [9] and DrugAgent [10], have applied agentic AI approaches that extend LLMs with specialized tools for information retrieval, knowledge integration, and decision support. By combining CoT reasoning with structured access to domain-specific data, these systems enable novel applications such as drug repurposing. On the other hand, models like MolReasoner [11] and Mol-R1 [12] have focused on leveraging these reasoning capabilities to fine-tune models for de novo molecular design, enabling the generation of novel chemical structures with desirable pharmacological properties, highlighting a paradigm shift toward reasoning-augmented LLMs in drug discovery.

In this study, we introduce DrugReasoner, an interpretable LLM built on the LLaMA architecture [13], designed to predict the likelihood of a drug approval based on molecular features. The reasoning capability of DrugReasoner is central to this task, as approval decisions depend on integrating complex, multifaceted evidence into a coherent judgment. By fine-tuning the model with group relative policy optimization (GRPO) [14] on a curated dataset containing both approved and unapproved compounds, as well as structurally similar approved and unapproved molecules for each entry, DrugReasoner is able to not only produce accurate predictions but also to articulate its decision-making process. Comparative evaluation against baseline and recent approval predictor models demonstrates that DrugReasoner achieves both strong predictive performance and enhanced interpretability, underscoring the importance of reasoning-augmented LLMs for transparent and efficient AI-assisted drug discovery.

Results

In order to address the challenge of drug approval prediction, we developed DrugReasoner, an algorithm based on Llama-3.1-8B-Instruct. By fine-tuning it using GRPO with customized reward functions, this framework is designed to generate CoT reasoning and accurately predict small-molecule approval.

The model first analyzes the molecular features of the query compound, then compares them with the molecular features of the five most similar approved and unapproved molecules from the training set. Following this, it outputs a binary label (approved/unapproved) along with an explanatory rationale (Fig 1).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. Schematic representation of DrugReasoner development and assessment.

The upper panel illustrates dataset preparation and processing. The lower panel shows model training with group relative policy optimization (GRPO) using customized reward functions, followed by comparative evaluation against other models on validation, test, and external datasets.

https://doi.org/10.1371/journal.pone.0342940.g001

DrugReasoner is effectively trained to interpret small-molecule approval

DrugReasoner was fine-tuned on a dataset of 2,255 approved and 2,255 unapproved small molecules, using customized reward functions that encouraged accurate predictions and coherent reasoning. Training was performed over 14,500 optimization steps, during which the model generated four outputs per input. The reward diagram is presented in Fig 2A (S1 File). Performance was monitored every 500 steps on the evaluation set. Checkpoint 12,500 was selected as the final model based on the AUC and other evaluation metrics and complete adherence (100%) to the expected output structure (Fig 2B, S2 File). In addition, the model was required to produce a confidence score for each prediction. This score stabilized at 0.87 toward the end of training. An example of model input formatting and reasoning output is shown in S3 File.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. DrugReasoner assessment.

(A) Reward trajectory during group relative policy optimization (GRPO) training with exponential moving average (EMA). (B) Evaluation metrics of model on the validation set. Comparative performance of DrugReasoner and baseline models on the (C) validation and (D) test sets. (E) Comparative performance of DrugReasoner with baseline models and ChemAP on an independent external dataset.

https://doi.org/10.1371/journal.pone.0342940.g002

Discussion

In this work, we presented DrugReasoner, a Llama-3.1-8B-Instruct algorithm, fine-tuned using GRPO and customized reward functions to predict drug approval outcomes while providing interpretable reasoning. The model achieved superior performance on the validation set compared to baseline models (logistic regression, KNN, SVM, and XGBoost) and classical drug-likeness metrics (Ghose filter and QED). On the test set, it surpassed or showed competitive results compared to these models. On an external dataset, however, DrugReasoner significantly outperformed both baselines and the recently developed model, ChemAP [4]. These results highlight DrugReasoner’s enhanced predictive accuracy, generalizability, and promise as a decision-support model in the drug discovery process.

The integration of reasoning capabilities into LLMs has emerged as a game-changing tool in computational drug discovery, enabling more interpretable and context-aware predictions across molecular tasks. For instance, models like CoTox [16] apply CoT reasoning to predict molecular toxicity by incorporating biological pathways and gene ontology, while MolReasoner [11] enhances interpretability through synthetic CoT samples and reinforcement learning to link chemical structures with linguistic descriptions. Similarly, frameworks such as DrugPilot [17] and DrugAgent [10] employ multi-agent collaboration and reasoning to automate workflows. They integrate domain-specific tools, addressing challenges in data processing and tasks like drug-target interaction modeling. Mol-R1 [12] further advances explicit long-CoT reasoning for molecule generation via iterative adaptation and distillation strategies, improving performance in knowledge-intensive domains. In this context, DrugReasoner extends reasoning-augmented LLMs to drug approval prediction by applying GRPO-fine-tuned CoT reasoning using molecular feature comparisons with structurally similar approved and unapproved compounds to generate binary outcomes, confidence scores, and explanatory rationales. By doing so, our model bridges predictive performance with interpretability in a domain where transparent decision-making is crucial.

Traditional machine learning approaches for drug approval prediction like DrugApp [18] have relied on diverse features, including molecular, physicochemical, clinical trial, and patent data, often achieving robust performance via ensemble methods. However, features like patent or clinical trial data are typically available only post-trial, limiting their use in early discovery. ChemAP addresses this gap by predicting approval from chemical structures via knowledge distillation, enriching semantic knowledge to distinguish approved from unapproved drugs. Furthermore, DrugReasoner introduces an LLM-based paradigm that operates on molecular features (derived solely from structures) to enhance prediction with interpretable reasoning. This enables superior results on external datasets compared to ChemAP and conventional baselines. By emphasizing coherent rationales and confidence scoring, our model not only improves predictive accuracy but also enhances trustworthiness, offering a new paradigm for early-stage drug discovery.

Despite promising results, this study has limitations that should be addressed in future studies. To mitigate the risk of potential data leakage from LLMs pretrained on vast internet datasets, we excluded simplified molecular-input line-entry system (SMILES) representations, relying instead on molecular features for improved interpretability. While this reduced the risk of bias, it may also have constrained the model’s ability to capture fine-grained structural information. Computational constraints further limited our exploration of larger models. Extending the context window, which could have allowed the inclusion of SMILES for similar approved and unapproved molecules, and systematic hyperparameter searches during GRPO fine-tuning could further enhance the reasoning performance. Future work should address these constraints by integrating structural data in a controlled manner, optimizing hyperparameters, scaling model size, and context length to improve reasoning and predictive accuracy.

Overall, DrugReasoner represents a significant advancement toward interpretable reasoning-augmented LLMs for drug approval prediction. By combining CoT reasoning with GRPO fine-tuning and molecular feature comparisons, it achieves strong performance and generates explanatory rationales across validation, test, and external datasets. Its ability to operate solely on molecular features while maintaining generalizability, highlights its potential to accelerate pharmaceutical decision-making. As such, DrugReasoner establishes a foundation for transparent and trustworthy AI tools, capable of guiding critical investment and research decisions from the earliest stages of drug discovery.

Methods

DrugReasoner development

We fine-tuned Llama-3.1-8B-Instruct [23] using GRPO trainer from hugging face transformer reinforcement learning (TRL) library (version 0.15.2) [24] to perform CoT reasoning on molecular features.

Training utilized the unsloth library (version 2025.3.19) [25] with:

• Maximum sequence length: 4,096 tokens (prompt + completion, 2,048 each)
• 4-bit quantization [26]
• Low-rank adaptation (LoRA) [27] with rank 32, applied to key attention projection layers (q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj) to reduce memory overhead.

Reinforcement learning with GRPO.

The model was trained using GRPO trainer for 14,500 steps, generating four outputs per prompt during training, which were scored using multi-objective rewards. Training was performed with a learning rate of 5 × 10 ⁻ ⁶, Adam optimizer parameters β₁ = 0.9 and β₂ = 0.99, a weight decay of 0.1, 100 warmup steps, and a cosine learning rate scheduler. All other variables set to the default of the library. Optimization employed the paged_adamw_8bit optimizer [28] with a per-device batch size of 4 and gradient clipping at a maximum norm of 0.1.

Group relative policy optimization works by generating multiple possible responses or a group of possible responses from the current policy for each input . Each response in the group is evaluated and assigned a reward score based on the defined reward function. The average reward across the group serves as a baseline (Eq. 1), and the advantage of each response is calculated as the difference between its reward and this group mean (Eq. 2).

(1)(2)

Using this advantage metric, the model updates its parameters to reinforce responses with above-average rewards while discouraging those with below-average performance. To ensure training stability, a clipped surrogate objective is applied, where the clipping function ( prevents updates from deviating excessively from the reference policy. Here, is the policy before the update and is a small hyperparameter (Eq. 3 and 4).

(3)(4)

This optimization is guided by a specialized loss function () that also incorporates a KL-divergence penalty () to prevent abrupt policy changes and ensure stable learning (Eq. 5).

(5)

Reward functions.

We implemented a multi-faceted reward function strategy composed of five distinct objectives to perform the scientific reasoning task.

• Correctness: Compares predicted label to ground truth (2.0 for correct, 0.0 otherwise). This function also logs intermediate outputs for qualitative analysis.
• XML format: Assesses whether the response follows the specified format (presence and singularity of <think > , < label > , and <score> tags). Each correctly placed start (<>) and end (</>) tag contributes 0.125, summing to a maximum of 0.75 (6*0.125).
• Soft format compliance: A regular expression-based reward function that returns 0.5 if the response approximately matches the expected format, meaning it includes the required structural elements (e.g., specific tags) in the correct order with allowances for minor whitespace variations, extra text, or small ordering changes, and 0.0 if key elements are missing or the structure is significantly incorrect.
• Interpretability: Assigns 0.5 if the extracted label is semantically valid (i.e., approved or unapproved).
• Confidence-alignment: Evaluates consistency between the predicted label and its confidence score:
  • Correct with high confidence (≥ 0.7): 2.0
  • Correct with moderate confidence (0.4 - < 0.7): 1.0
  • Correct with low confidence (< 0.4): 0.0
  • Incorrect with low confidence (< 0.4): 1.0
  • Incorrect with moderate confidence (0.4 - < 0.7): 0.5
  • Incorrect with high confidence (≥ 0.7): 0.0

Each reward function returned a scalar value, which was used to score the model’s generated completions during GRPO training. Reward signals were recorded at each step for inspection and debugging (Fig 2A and S1 File).

Checkpoint selection.

Training was performed for a total of 14,500 steps. To identify the optimal checkpoint for inference, we tracked the trajectory of reward scores throughout training. In addition, every 500 steps, we performed inference using top_p = 0.9, top_k = 9, and temperature = 1, evaluating the model on the validation set with metrics including, AUC, F1, precision, recall, specificity, and accuracy. The checkpoint at 12,500 steps achieved optimal performance and was selected as the final model (Fig 2B and S2 File).

Supporting information