Let Large Models ’Reason Before They Edit’: DrugR Tackles the Balance Challenge in Multi-Objective Drug Optimization

2026-04-20

Today, large language models (LLMs) have shown strong capabilities across many general tasks and are pushing drug discovery toward a "data- and intelligence-driven" paradigm. Yet whether LLMs truly understand the relationship between molecular structures and complex pharmacological properties remains an open question. To address this, researchers from the THUNLP lab, a community member, propose in their latest paper DrugR, an explicit reasoning model for drug molecule optimization. Composed of three modules — knowledge injection, explicit reasoning supervised fine-tuning, and self-balancing multi-objective reinforcement learning — it targets 23 ADMET optimization objectives and systematically reveals the capability boundaries of current LLMs across multiple real-world drug scenarios.

DrugR at a glance

Explicit reasoning-driven optimization: The first framework to introduce structured pharmacological reasoning chains into LLM-based molecule optimization; the model must first generate interpretable optimization rationales, realizing "think it through before making the change";

Reverse data engineering: Rather than costly manual annotation, it anchors on validated drugs from DrugBank and uses an innovative "generate → filter → reverse-derive reasoning" pipeline to automatically build high-quality explicit reasoning training data;

Self-balancing multi-granularity reinforcement learning: To address convergence imbalance in multi-objective optimization, it proposes a two-level adaptive mechanism combining sample-level Pareto-aware reweighting and batch-level dynamic scaling, driving 23 pharmacological properties upward together;

Precise pain-point identification: It reveals key bottlenecks in existing models across implicit structure-property reasoning, multi-objective imbalanced convergence, cross-drug generalization, and the interpretability of autonomous planning.

Results paper

➤ Paper link

🔗https://arxiv.org/abs/2602.08213

➤ Hugging Face

🔗 https://github.com/Haoranliu-lab/DrugR-main

Broader optimization dimensions: property, structure, and decision at multiple granularities

DrugR does not just care about which molecule the model changed; it cares about why it made the change and whether the change is justified. The framework spans three levels, fully testing the limits of LLM pharmacological reasoning:

Domain pre-training (Continuous Pre-training): Built on LLaMA-3-8B-Instruct and trained on specialized chemistry and pharmacy corpora such as ChemicalQA (~150K) and MoleculeNet (~160K), it strengthens the model's fundamental understanding of molecular structures and pharmacological properties;

Supervised fine-tuning: Through reverse data engineering, it builds 4,855 high-quality supervised pairs of reasoning chains + optimized SMILES, and uses a curriculum learning strategy so the model progresses from generating correct molecules to generating pharmacologically sound reasoning processes;

Reinforcement learning: With GRPO as the backbone, it introduces the self-balancing multi-granularity mechanism, focusing on the model's multi-objective decision-making across 23 ADMET metrics and requiring genuine joint improvement among mutually constraining objectives such as molecular validity, similarity, and drug-likeness.

DrugR's three-stage training process

Truer capability boundaries: exposing LLMs' "pharmacology blind spots"

After systematic evaluation of mainstream open-source and closed-source models including GPT-4o, DeepSeek-R1, and Qwen2.5, the paper shows that although LLMs excel in general reasoning, they still face severe challenges in multi-objective drug optimization:

DrugR reaches an overall optimization score of 0.27, far above strong baselines such as directly calling DeepSeek-R1 (0.18), with notable advantages in target property identification (F1: 0.44 vs 0.19) and reasoning quality (LMS: 0.77 vs 0.70). This confirms both the complexity of drug molecule design and the capability gaps of general-purpose LLMs on vertical tasks.

Ablation experiments show that removing the reasoning process entirely and having the model generate optimized molecules directly makes the overall score plummet from 0.27 to 0.015. This strongly proves that the reasoning chain is not a nice-to-have but the core support that makes the drug optimization task work.

Comparison of overall optimization performance between baselines and our method

More behavioral insight: what is the model reasoning about?

Through fine-grained qualitative and quantitative analysis, DrugR breaks down the vague understanding of LLM drug optimization capability and uncovers the core pain points blocking AI drug design in practice:

Pain point 1: Multi-objective imbalance leads to robbing Peter to pay Paul. When optimizing 23 ADMET properties at once, naive reinforcement learning strategies suffer severe convergence imbalance — some properties are over-optimized while others are neglected for long periods. Experiments show that with Pareto-aware reweighting, the joint improvement curves of the properties become markedly synchronized, proving the necessity of the self-balancing mechanism for multi-objective optimization.

Pain point 2: Implicit structure-property mapping is a black box. Without reasoning constraints, existing LLMs tend to output optimization results based on empirical intuition and cannot explain why a particular modification improves the corresponding property. After DrugR forcibly introduces a reasoning chain, the pharmacological rationales it outputs align closely with domain expert judgment, and reasoning accuracy (target property F1) improves by more than 128%.

Pain point 3: The dilemma between structural similarity and functional improvement. Without constraining structural similarity, models tend to generate molecules with excellent pharmacological properties but a scaffold unrecognizable from the original drug, losing practical medicinal value. DrugR's reasoning chain mechanism naturally constrains the optimization direction, so the final molecules meet practical standards on both fingerprint similarity (0.64) and binding affinity (96% satisfying the ≤ -6 kcal/mol constraint).

Pain point 4: The transfer barrier of cross-domain generalization. Existing models are often trained for a single disease area and must start over for a different drug class. On an unseen anticancer molecule dataset, DrugR reaches an overall score of 0.21 after fine-tuning with only 50 samples, demonstrating strong cross-domain transfer capability and offering a viable path for drug design in low-resource settings.

To illustrate these capabilities intuitively, the paper presents representative optimization cases of DrugR in antihypertensive and anti-inflammatory drugs, with blue annotations marking valid reasoning that matches the structural modification and has a clear mechanism (e.g., "introducing a polar group to improve membrane permeability").

Case study

Toward interpretable, auditable intelligent drug design: This work aims to give the community a new drug optimization paradigm that is driven by explicit reasoning, supports multi-objective joint optimization, and offers cross-domain transfer. By integrating knowledge injection, reasoning supervision, and self-balancing reinforcement learning, DrugR not only shows the potential of LLMs in drug design but also reveals clear remaining challenges in multi-objective optimization, structure-property reasoning, cross-drug generalization, and interpretability. These findings further map the capability gap between general-purpose LLMs and real-world drug optimization needs. The author team says: "We hope DrugR can become a bridge between the two, so models not only make good changes but also explain them clearly, and push AI-assisted drug design toward a more interpretable, auditable, and trustworthy real-world R&D workflow."