cheminformatics tooling

Large language models (LLMs) exhibit poor chemical validity and low efficiency (only 3%) in molecular design, generating numerous chemically invalid or unsynthesizable structures. Method: This work proposes a scientific-constraint-guided LLM framework for molecular design, integrating domain-adapted fine-tuned models, chemistry-aware prompt engineering, and automated structural validity verification to ensure generated molecules satisfy chemical validity, synthetic accessibility, and target property optimization. Contribution/Results: The approach elevates the rate of chemically valid molecule generation to 83% and achieves up to a 17-fold improvement in computationally predicted target protein binding affinity. To our knowledge, this is the first systematic, generalizable, and reproducible interdisciplinary generative paradigm—termed “chemical-rule embedding → LLM generation → closed-loop validation”—establishing a methodological foundation and practical blueprint for AI-driven drug discovery.

Current evaluation metrics for molecular generation models often fail to reflect the practical viability of compounds in drug discovery, leading to a high rate of false positives. To address this limitation, this work proposes HEDGEHOG—a six-stage cascaded filtering benchmark that emulates industrial hit identification workflows by incorporating multidimensional constraints, including physicochemical properties, structural plausibility, synthetic accessibility, molecular docking, and three-dimensional conformational stability. HEDGEHOG establishes the first unified, hierarchical evaluation framework closely aligned with real-world drug discovery practices. When applied to 230,000 generated molecules, only 0.65% passed all screening stages, exposing a fundamental shortcoming of existing models in simultaneously satisfying multiple realistic constraints. This benchmark provides a more rigorous and reliable standard for assessing molecular generation methods.

This work addresses the challenge in structure-based drug design where existing large language model (LLM) agents struggle to simultaneously optimize ligand binding affinity and drug-likeness within a single editing step. To overcome this limitation, the authors propose the PROBE framework, which introduces a novel “edit-response probing” mechanism. By conducting controlled probe experiments on ligand–pocket complexes, PROBE constructs pocket-specific site maps and an EditManual that guide a multi-agent system—comprising affinity-, drug-likeness-, and co-optimization-focused agents—to collaboratively explore the feasible edit space. This approach effectively identifies regions amenable to joint optimization of both objectives, thereby breaking through the single-step editing bottleneck. Evaluated on the CrossDocked2020 benchmark, PROBE achieves state-of-the-art performance and substantially mitigates the optimization failure commonly observed in current methods.

Existing generative models for drug design frequently produce molecules that are chemically unsynthesizable. Method: We propose a GFlowNet-based generative framework grounded in forward-synthesis pathway modeling, which—uniquely—explicitly embeds chemical reaction templates and a purchasable reagent library into the action space to enable end-to-end learning of synthetic feasibility constraints. To mitigate reaction-encoding bias and support multi-constraint MDP formulation, we introduce an inverse-policy learning mechanism. Our approach integrates reaction graph encoding, SA Score-guided optimization, and independent retrosynthetic validation. Results: Experiments demonstrate substantial improvements over baselines: generated molecules exhibit significantly higher structural diversity, average SA Score decreases by 12.3%, retrosynthetic success rate increases by 18.7%, and the model reliably infers feasible synthesis pathways for novel molecules.

To address the critical bottleneck in drug discovery—where molecular generation models neglect synthetic feasibility, hindering experimental validation—this work proposes a novel molecular generation framework projectable onto synthetically accessible chemical space. Methodologically, it introduces synthesis path expressions (SPEs) as a novel molecular representation that intrinsically encodes retrosynthetic logic, and designs a graph-based Transformer architecture for end-to-end translation from molecular graphs to SPEs. This formulation inherently guarantees synthetic feasibility of generated molecules and enables structure-preserving, synthetically constrained analog generation for initially infeasible candidates. Experiments demonstrate substantial improvements in retrosynthetic planning accuracy and successful re-mapping of multiple state-of-the-art generative model outputs—previously deemed synthetically intractable—into property-preserved, experimentally viable analogs. The approach effectively bridges the gap between de novo molecular generation and practical synthesis.

Molecular editing and optimization require iterative property improvement while preserving chemical validity and structural similarity. This work formalizes the task as an embodied agent reinforcement learning problem—the first such formulation—and introduces a two-stage RL training framework. Within this framework, a large language model (LLM) agent collaborates with domain-specific chemical tools—including validity checkers, LogP/solubility predictors, and Tanimoto similarity calculators—to close a “reasoning–tool invocation–optimization” loop. The approach enables interpretable, iterative exploration of chemical space. Our MolEditAgent-7B achieves 100%, 95%, and 98% chemical validity for atom addition, deletion, and substitution edits, respectively. MolOptAgent-7B outperforms Claude 3.7 in LogP optimization and demonstrates balanced performance across multi-objective tasks (e.g., aqueous solubility). These results establish a new paradigm for LLM-driven, tool-augmented molecular design.

Traditional lead compound optimization often relies on single-step strategies that struggle to balance structural constraints with long-term optimization objectives, frequently disrupting critical pharmacophores and yielding limited improvements in ADMET properties. This work reframes molecular optimization as a trajectory-aware sequential decision-making problem and introduces TRACE, an intelligent agent powered by large language models that plans multi-step tool invocation sequences to achieve forward-looking optimization while preserving essential molecular scaffolds. By integrating molecular generation and editing tools, ADMET prediction, and structural similarity constraints into a unified framework, TRACE substantially outperforms existing baselines across multiple tasks, significantly enhancing optimization success rates, the magnitude of property improvement, molecular validity, and retention of original structural features.

This work addresses the challenge of accurately translating natural language instructions into chemically valid molecular structures under stringent constraints in AI-driven drug discovery. The authors propose Mol-Debate, a novel framework that introduces a multi-agent debate mechanism to simulate the multi-perspective critique and iterative refinement characteristic of real-world drug design. Through a generate–debate–optimize loop, Mol-Debate uniquely coordinates developer and evaluator roles, integrating global and local structural reasoning with both static and dynamic chemical knowledge. Experimental results demonstrate that Mol-Debate achieves a 59.82% exact match rate on ChEBI-20 and a 50.52% weighted success rate on S²-Bench, substantially outperforming existing baselines.

This work addresses the instability in oracle-in-the-loop search for conditional molecular optimization, which arises from a mismatch between supervisory signals and action-decision hierarchies. The authors propose a discrete optimization framework centered on edit-level response modeling. By decomposing property differences between molecular pairs into minimal editing units, they construct transferable action primitives and introduce a single-step molecular edit response predictor that, in conjunction with guided tree search, translates endpoint property discrepancies into process-level supervision. Chemical feasibility constraints and weakly associated molecular pair mining are integrated to substantially reduce reliance on external oracles. Experimental results demonstrate that the method achieves higher target property success rates and more chemically plausible editing pathways while requiring fewer oracle queries.

This work addresses the challenge in lead compound optimization of simultaneously ensuring molecular validity and synthetic feasibility. The authors formulate the problem as a Markov decision process, leveraging a tool-augmented large language model (LLM) agent that operates within an action space constrained by validated reaction templates to dynamically identify reactive sites and generate chemically plausible transformation pathways. Their approach innovatively integrates trajectory-level Group Relative Policy Optimization (GRPO), a SMILES caching mechanism, and chemical analysis tools to guarantee that each modification step is both synthesizable and efficient. Experimental results demonstrate that the method achieves an average Top-10 score of 0.563 across 14 tasks, representing a 10.4% improvement over the strongest synthesizability-aware baseline, attains the highest sample efficiency in 10 tasks, and reduces end-to-end optimization time by approximately 43%.