DrugGen 2

Most target-conditioned molecular generators propose candidate drugs from a protein sequence alone, treating the target in isolation from the disease it belongs to. This ignores that the same protein can play different roles across pathologies, and that therapeutic success depends on disease context as much as on binding a single receptor. DrugGen 2 addresses this gap by conditioning molecular design jointly on disease ontology and target protein sequence, generating drug-like SMILES strings tailored to a specified disease-target pair.

DrugGen 2 extends DrugGen, a peer-reviewed GPT-2-based generator that produces valid, high-affinity molecules from a target sequence using supervised fine-tuning followed by reinforcement learning. Where the predecessor optimized purely target-conditioned generation, DrugGen 2 adds a disease channel: it encodes the disease as a Medical Subject Headings (MeSH) directed acyclic graph and combines that context with the protein sequence to steer generation. The model was developed by a research group at Isfahan University of Medical Sciences and released as a preprint.

The result is a "disease-aware" language model positioned for both de novo drug design and repurposing, where knowing the disease as well as the target lets the generator bias toward chemistry that is plausible for the intended therapeutic setting.

#Key Features

• Disease-aware conditioning: Encodes the target disease as a MeSH directed acyclic graph and conditions generation on it alongside the protein sequence, rather than on the target alone.
• Target-conditioned SMILES generation: Takes an amino acid sequence as input and emits drug-like SMILES strings intended to bind that target.
• Two-stage training: Combines supervised fine-tuning with reinforcement learning via Group Relative Policy Optimization (GRPO), building on the Transformer Reinforcement Learning (TRL) library.
• Multi-objective reward: Optimizes generated molecules for chemical validity, novelty, diversity, and predicted binding affinity simultaneously.
• Open release: Ships an inference checkpoint, training dataset, a hosted demo, and code, with Python API and CLI interfaces for immediate use.

#Technical Details

DrugGen 2 uses a fine-tuned GPT-2 backbone. Training proceeds in two stages: supervised fine-tuning on a curated dataset linking approved drugs to their associated diseases and protein targets, followed by reinforcement learning with GRPO, in which a composite reward scores each generated molecule on validity, novelty, diversity, and predicted protein-ligand binding affinity. Inputs are a disease represented as a MeSH DAG plus a target protein sequence; the output is a SMILES string. The authors evaluate the model on five protein targets relevant to diabetic nephropathy, reporting that it outperforms the DrugGPT and DrugGen baselines and that generated compounds reach predicted binding affinities exceeding those of reference drugs such as enalapril. Code is released on GitHub under GPL-3.0, with the checkpoint and training data hosted on HuggingFace; the preprint itself carries a CC BY-NC-ND 4.0 license.

#Applications

DrugGen 2 targets medicinal chemists and computational drug discovery teams who want candidate molecules generated for a specific disease-target combination rather than a target in the abstract. Because it accepts a disease and a protein sequence and returns synthesizable-looking SMILES, it fits early-stage hit generation, target-based de novo design, and drug repurposing workflows, where the disease-aware channel can bias output toward chemistry appropriate to the intended indication. The bundled demo space and CLI lower the barrier for researchers to query the model on their own targets without retraining.

#Impact

DrugGen 2 continues a line of open, reproducible drug-generation models from the same group, whose peer-reviewed predecessor demonstrated that reinforcement learning against binding-affinity and validity rewards can meaningfully improve GPT-2-based molecular generation. Its contribution is the explicit incorporation of disease context into conditioning, a direction that reframes target-based generation as disease-target generation. As a preprint awaiting peer review, its benchmarks are in-silico and demonstrated on a focused set of diabetic nephropathy targets; predicted binding affinities and generated compounds have not yet been experimentally validated. The open checkpoint, dataset, demo, and GPL-3.0 code make the approach straightforward for others to test and extend.