---
title: 'DrugGen-2: Disease-Aware Generative Drug Design'
url: https://www.emergentmind.com/topics/druggen-2
type: topic
---

# DrugGen-2: Disease-Aware Generative Drug Design

Searching arXiv for DrugGen-2 and related conditioning/generative drug design work.
DrugGen-2 is a disease-aware generative model for small-molecule design that conditions molecular generation on both disease ontology and target protein sequence. It was developed by fine-tuning a pre-trained GPT-2 model on approved drug, disease, and target associations, using a two-stage procedure of supervised fine-tuning followed by reinforcement learning via group relative policy optimization (GRPO). In the reported evaluation, the model was assessed on five protein targets relevant to diabetic nephropathy and was found to outperform DrugGPT and DrugGen in unique molecule generation, similarity to approved drugs, and predicted binding affinity, with docking analyses providing additional support for several ACE candidates [2607.08404].

## 1. Conceptual basis and problem formulation

DrugGen-2 is motivated by a limitation of target-centric molecular generation: many computational drug design systems condition on protein targets or generic molecular properties while ignoring disease context, even though target behavior and therapeutic outcomes can be disease-dependent. The reported motivation is explicitly biological rather than purely cheminformatic. PPARγ activation, for example, is described as improving insulin sensitivity in type 2 diabetes while potentially interacting unfavorably with the Wnt/β-catenin axis in colon cancer. ACE is similarly described as pro-hypertensive via Ang II formation in the vasculature yet neuroprotective through Aβ degradation in the brain. DrugGen-2 addresses this by conditioning generation on a disease ontology representation and a target sequence jointly, so that ligand design is informed by both molecular target identity and pathological context [2607.08404].

The system extends DrugGen and DrugGPT along three axes. First, it jointly conditions on a MeSH DAG identifier and an amino-acid sequence, thereby accepting disease-target pairs rather than protein-only prompts. Second, it uses a two-stage training pipeline consisting of supervised fine-tuning (SFT) and reinforcement learning with GRPO. Third, it reports empirical gains on diabetic nephropathy targets in unique generation count, structural similarity to approved drugs, and predicted binding affinity. The central claim is therefore not merely that DrugGen-2 generates valid SMILES, but that disease ontology can act as a meaningful conditioning signal for de novo design [2607.08404].

A common misunderstanding would be to treat the disease-aware property as evidence of an explicit mechanistic disease model. The reported architecture does not introduce pathway simulators, structured biological graphs inside the generator, or cross-attention modules specialized to disease biology. Disease information enters the model as a serialized ontology token sequence in the causal language model input. The disease awareness is thus implemented through conditional language modeling rather than explicit systems-biology reasoning.

## 2. Input representation and generative architecture

DrugGen-2 fine-tunes DrugGPT, described as a GPT-2-style causal language model specialized for ligand generation. The manuscript does not report the parameter count of the base model. Molecules are represented as SMILES, proteins as amino-acid sequences in single-letter code, and disease context as MeSH DAG identifiers. Three structural conditioning tokens were added to the tokenizer vocabulary: `D` for disease, `P` for protein, and `L` for ligand. The resulting tokenizer vocabulary contains 53,086 tokens [2607.08404].

The conditioning mechanism is implemented by serial concatenation within the causal input stream:

```text
<|startoftext|> <D> <MeSH DAG> <P> <Sequence> <L> <SMILES> <|endoftext|>
```

No cross-attention modules, adapters, or prefix-tuning mechanisms are introduced. Disease and sequence information are placed before the ligand segment, and the model autoregressively generates the SMILES continuation conditioned on those preceding tokens. Protein sequences are padded or truncated to 768 tokens; the context window is 1,024 tokens, and the maximum generation length is 256. Sampling is reported with `top-p=1.0`, `top-k=None`, and `temperature=1.0` [2607.08404].

This design has two immediate implications. First, the model remains close to a standard causal LM training and inference stack, which simplifies reuse of DrugGPT infrastructure. Second, disease conditioning is coupled to the tokenization and prompt schema rather than to an external conditioning tower. This suggests that any benefit of disease ontology must be recoverable from the sequential token context itself, rather than from a dedicated multimodal fusion mechanism.

## 3. Data curation and construction of disease-target-ligand triplets

The training corpus was built by linking drug-target, target-disease, and drug-disease resources. Drug-target pairs came from DrugBank v5.1.10, with 1,710 small molecules annotated to human targets. Of these, 117 were labeled “withdrawn,” and 50 were excluded due to safety or adverse-effect considerations. SMILES were available for 1,634 of the 1,660 selected molecules, sourced from DrugBank, ChEMBL, and ZINC20. Target information comprised 2,116 related protein targets; 27 absent from UniProt were mapped via reviewed IDs, names, or BLAST, UniProt ID Q5JXX5 was removed, and final sequences were retrieved for 2,093 proteins via the UniProt API. Target-disease associations were obtained from DisGeNET v3.12.1, with disease CUIs updated to UMLS 2023AA and MeSH terms retrieved through the UMLS API, yielding 608 MeSH terms mapped to 2,042 protein targets. Drug-disease associations were taken from approved indications in ChEMBL33, retaining 1,299 small molecules associated with 643 MeSH terms, which were further mapped to 1,632 DAGs [2607.08404].

From these sources, the pipeline constructed 1,113,539 MeSH-sequence-SMILES strings. After filtering to entries with a direct MeSH-SMILES relationship, 13,908 strings remained for training. The paper states that sequences were padded to 768 tokens and longer sequences truncated. It does not describe stereochemistry normalization, canonicalization, deduplication, or property filters such as Lipinski, QED, or SA beyond validity checks; SMILES were used as provided by the source databases [2607.08404].

The dataset design matters for interpreting the model’s behavior. Because the corpus is built from approved drugs and curated disease-target mappings, the resulting generator is biased toward approved-drug chemical space and known therapeutic associations. This is consistent with the later finding that generated molecules exhibit greater structural similarity to approved drugs. A plausible implication is that DrugGen-2 is optimized more for disease-aware exploration near validated medicinal chemistry regions than for unconstrained novelty far from known scaffolds.

## 4. Supervised fine-tuning, GRPO, and reward specification

The supervised stage uses a standard causal language modeling objective over the 13,908 MeSH-sequence-SMILES triplets. Training was performed for 10 epochs using the Hugging Face TRL SFT trainer, AdamW with learning rate \(5 \times 10^{-4}\), \(\epsilon = 1 \times 10^{-8}\), batch size 8, linear warmup of 100 steps, and gradient accumulation of 1. The checkpoint at epoch 6 was selected as the reference policy for reinforcement learning. Regularization and early stopping are not specifically reported [2607.08404].

The second stage applies GRPO. For a prompt \(q\), the algorithm samples a group of \(G\) outputs and evaluates each completion with a scalar reward. Group-relative advantage is defined against the group mean rather than a separately learned value function:

$$
J(\theta) = \mathbb{E}_{o \sim \pi_\theta}[R(o)]
$$

$$
A(o_i) = R(o_i) - \bar{R}, \qquad \bar{R} = \frac{1}{G}\sum_{i=1}^{G} R(o_i)
$$

The policy update uses PPO-style clipping with a KL penalty to the SFT reference policy:

$$
L(\theta) = \mathbb{E}\left[\min\left(r_t(\theta)A_t,\ \operatorname{clip}(r_t(\theta), 1-\epsilon, 1+\epsilon)A_t\right)\right] - \beta D_{\mathrm{KL}}(\pi_\theta \parallel \pi_{\mathrm{ref}})
$$

where

$$
r_t(\theta)=\frac{\pi_\theta(o_{i,t}\mid q,o_{i,<t})}{\pi_{\theta_{\mathrm{old}}}(o_{i,t}\mid q,o_{i,<t})}.
$$

Implementation used Hugging Face TRL v0.23.1 with a custom `GRPOTrainer`, AdamW with \(\beta_1=0.9\), \(\beta_2=0.999\), \(\epsilon = 1 \times 10^{-8}\), weight decay \(0.1\), learning rate \(5 \times 10^{-6}\), linear scheduling, warmup of 100 steps, 10 epochs, fp16 mixed precision, gradient clipping with max norm 1.0, gradient accumulation for effective batch factor 4, and seeds set to 42. During RL, generation used group size \(G=40\), `top-p=1.0`, `top-k=None`, `temperature=1.0`, context 1,024, and maximum generation length 256. No entropy bonus is reported, and the manuscript does not disclose the numeric value of the KL penalty weight \(\beta\) [2607.08404].

The reward design consists of four components. Chemical validity is binary and determined through an RDKit parsing and validation pipeline. Predicted binding affinity is obtained from PLAPT, using \( \mathrm{p}K_d = -\log_{10}(K_d) \), with invalid molecules receiving zero for the affinity component. Novelty against the curated approved-drug set is binary: a generated SMILES present in the approved set receives 0, otherwise 1. Diversity during training is implemented as batch-level uniqueness: duplicated SMILES within the sampled group receive 0, unique SMILES receive 1. The manuscript emphasizes that this training-time notion of “diversity” is a binary uniqueness reward rather than a continuous internal diversity metric. It also states that similarity to approved drugs, computed later with Morgan fingerprints and Tanimoto similarity, was used in evaluation rather than as an RL reward. The exact scalarization of reward components is not reported [2607.08404].

## 5. Evaluation design and quantitative performance

Evaluation focused on diabetic nephropathies, represented by four MeSH DAG descriptors: `C12.050.351.968.419.192`, `C12.200.777.419.192`, `C12.950.419.192`, and `C19.246.099.875`. Five targets were used: ACE, PPARγ, NOS3, PAI-1, and TGF-β1. Baselines were DrugGPT and DrugGen with default parameters; these baselines received protein-only prompts, whereas DrugGen-2 received MeSH DAG-protein sequence pairs. Uniqueness was assessed by requesting up to 500 unique molecules per target for DrugGPT and DrugGen, or per MeSH-target pair for DrugGen-2, stopping after 30 sampling cycles without novel molecules. Validity, similarity to approved drugs, and predicted binding affinity were assessed on sets of 100 unique molecules per input. Statistical analyses used \(\chi^2\) tests or Kruskal-Wallis tests with Benjamini-Hochberg correction; reported effect sizes were Cramér’s \(V\) and \(\epsilon^2\) [2607.08404].

| Metric | DrugGen-2 | Comparator context |
|---|---:|---|
| Unique molecules per 500 attempts | 409 [406–427] to 444 [443–448] | DrugGen: 50 [41–87]; DrugGPT: 219 [217–254] |
| Validity | median 99–100% | Comparable to DrugGPT and DrugGen |
| Similarity to approved drugs | median ≈0.70 | DrugGen: 0.64 [0.51–0.76]; DrugGPT: 0.30 [0.26–0.41] |
| Predicted binding affinity (PLAPT) | 9.26–9.97 pK\(_d\) | DrugGen: 7.15–8.49; DrugGPT: 5.86–6.22 |

The reported significance values are strong. For uniqueness, \(\chi^2 = 88.80\), \(P < 10^{-9}\), and Cramér’s \(V = 0.05\). For validity, \(\chi^2 = 0.12\), \(P = 1.0\), and Cramér’s \(V = 0.003\), indicating no meaningful difference across models. For similarity to approved drugs, the Kruskal-Wallis result is \(H = 961.88\), \(\epsilon^2 = 0.34\), \(P < 10^{-204}\). For PLAPT affinity, DrugGen-2 showed per-target significance across all five proteins: ACE \(\epsilon^2 = 0.49\), \(P < 10^{-59}\); PAI-1 \(\epsilon^2 = 0.57\), \(P < 10^{-65}\); PPARγ \(\epsilon^2 = 0.57\), \(P < 10^{-67}\); TGFB1 \(\epsilon^2 = 0.50\), \(P < 10^{-55}\); and NOS3 \(\epsilon^2 = 0.34\), \(P < 10^{-38}\) [2607.08404].

An important interpretive point is that no significant differences were observed among the MeSH-specific DrugGen-2 variants. This means the reported gains should not be read as sensitivity to one particular diabetic nephropathy DAG encoding. Rather, the result supports robustness across multiple MeSH paths for the same disease family.

## 6. Docking results, limitations, and reproducibility

Docking was carried out with GLIDE XP in Schrödinger Maestro using ACE (PDB 1UFZ) and PPARγ (PDB 4EMA), with protein preparation through the Protein Preparation Wizard, protonation at \( \mathrm{pH}\ 7.4 \pm 0.5 \), LigPrep under OPLS4, rigid-receptor and flexible-ligand docking, and blind docking grids of \(40 \times 40 \times 40\ \text{\AA}^3\). The docking library size was 125 molecules per MeSH-target pair. Redocking RMSD was used for protocol validation [2607.08404].

For ACE, several generated compounds scored more favorably than the reported reference drug. P12821-293, P12821-10, and P12821-269 scored \(-9.917\), \(-9.485\), and \(-9.367\), respectively; P12821-225 scored \(-9.29\). The reference Enalapril scored \(-8.283\), and redocked Captopril scored \(-6.168\). The paper states that binding poses aligned spatially with reference ligands in the ACE active site and that visualizations showed comparable anchoring. For PPARγ, the reference Rosiglitazone redocked at \(-7.422\), Balsalazide scored \(-7.399\), and generated molecules P37231-165, P37231-117, P37231-39, and P37231-247 scored \(-6.749\), \(-6.461\), \(-5.464\), and \(-5.394\), respectively. The manuscript notes that some PPARγ candidates showed novel pharmacophoric interactions despite lower scores, and that P37231-165 contained substructures absent in Rosiglitazone. Redocking fidelity differed substantially between the two systems: Rosiglitazone→PPARγ achieved RMSD \(=1.44\ \text{\AA}\), whereas Captopril→ACE achieved RMSD \(=4.87\ \text{\AA}\), which the paper attributes to a challenging pocket and metalloprotein scoring limitations [2607.08404].

The limitations are explicit. The dataset contains only approved drugs and mapped targets and diseases, which may bias generation toward known scaffolds. Protein sequences are truncated or padded to 768 tokens, potentially omitting distal or allosteric information. PLAPT affinity predictions and docking scores are in silico proxies, and synthetic feasibility or ADMET properties were neither optimized nor evaluated; the paper specifically notes the absence of Lipinski, QED, SA, and ADMET objectives. Off-target liabilities also remain unknown. The recommended safety workflow includes in vitro target engagement assays, selectivity panels, cytotoxicity screens, ADMET profiling, and cheminformatics filters for PAINS, reactive motifs, and toxic substructures before wet-lab testing [2607.08404].

Reproducibility is comparatively strong at the software and artifact level. The paper reports use of Hugging Face Transformers and TRL, RDKit v2023.9.5, Schrödinger Maestro GLIDE XP (2021-2), the OPLS4 force field, and public APIs for UMLS, UniProt, DrugBank, and ChEMBL. Seeds were fixed at 42, and the dataset, model checkpoint, code repository, and interactive interface were released. Hardware and wall-clock training time, however, are not specified. The reported future directions include multi-target and multi-disease conditioning, addition of ADMET, QED, SA, Lipinski, and toxicity objectives, integration of continuous diversity metrics, more structural awareness through 3D or pocket-based methods, active-learning loops with wet-lab feedback, and richer sequence handling to avoid the limitations of 768-token truncation [2607.08404].

Source: https://www.emergentmind.com/topics/druggen-2