Source-linked AI summary
DrugGen 2: A disease-aware language model for enhancing drug discovery
Ali Motahharynia, Mohammadreza Ghaffarzadeh-Esfahani, Mahsa Sheikholeslami, Navid Mazrouei, Matin Irajpour, Yousof Gheisari, Hajar Sirous
TL;DR
Existing drug-generation methods often omit disease context, despite its relevance to target behavior and therapeutic outcomes. DrugGen-2 conditions molecule generation on disease ontologies and target sequences, outperforming DrugGPT and DrugGen across diabetic-nephropathy targets in novelty, approved-drug similarity, and predicted binding affinity.
Problem
Existing computational drug-design methods often condition generation on targets or molecular properties while omitting disease context relevant to target behavior and therapeutic outcomes.
Method
DrugGen-2 combines supervised fine-tuning and GRPO on approved drug–disease–target data to generate molecules conditioned on disease ontologies and target protein sequences.
Results
DrugGen-2 outperformed DrugGPT and DrugGen across five diabetic-nephropathy targets, generating more unique molecules with greater approved-drug similarity and higher predicted binding affinities.
Takeaways & Limitations
Disease-conditioned generation shows potential for AI-assisted de novo drug design and drug repurposing that accounts for disease–target relationships.
Takeaways & Limitations
Experimental validation is needed to confirm binding affinities, functional effects, and potential off-target interactions.
Abstract
from arXiv · showhide
Current computational approaches for drug design typically focus on generating molecules conditioned on specific targets or general molecular properties, often neglecting the influence of disease context on target behavior and therapeutic outcomes. To address this gap, we introduce DrugGen-2, a novel generative model that designs small molecules conditioned on both disease ontology and target protein sequences. DrugGen-2 was developed by fine-tuning a pre-trained GPT-2 model on a curated dataset of approved drugs linked to their diseases and targets, using a two-step strategy of supervised fine-tuning followed by reinforcement learning via group relative policy optimization (GRPO). This process was guided by reward functions optimizing for chemical validity, novelty, diversity, and high predicted binding affinity. When evaluated on five protein targets relevant to diabetic nephropathy, DrugGen-2 significantly outperformed baseline models (DrugGPT and DrugGen). It demonstrated a superior capacity to generate unique molecules, exhibited greater structural similarity to approved drugs, and achieved improved predicted binding affinities across all targets. Molecular docking analyses further supported these findings, identifying candidate ligands with strong binding potential, including compounds with predicted affinities (-9.917, -9.485, and -9.367) exceeding those of reference drugs such as enalapril for angiotensin-converting enzyme (-8.283). By integrating disease-specific context into molecular generation, DrugGen-2 advances AI-assisted drug discovery, offering a powerful tool for de novo design and drug repurposing that accounts for the complex interplay between diseases and molecular targets.
1. Introduction
Drug discovery is a multifaceted, resource-intensive process spanning small-molecule design through clinical translation. Traditional phenotypic screening and target-based design remain limited in efficiently exploring the vast chemical space.
- Drug discovery challenges: Drug discovery involves multiple stages, from designing high-quality small molecules to achieving clinical translation.The passage characterizes the overall process as multifaceted and resource-intensive.
- Traditional approaches: Classical approaches include phenotypic screening for biological activity and target-based design to interact with specific molecular targets.These strategies represent two traditional routes for identifying or designing compounds.
- Drug discovery challenges: Traditional strategies remain limited in their ability to efficiently explore the vast chemical space.The passage identifies inefficient exploration as a limitation of established discovery methods.
2. Results
DrugGen-2 was trained with supervised fine-tuning followed by GRPO and evaluated against DrugGen and DrugGPT on five diabetic-nephropathy-associated protein targets. It consistently produced more unique and valid molecules, greater similarity to approved drugs, and higher predicted binding affinities.
- Evaluation targets: The benchmark covered ACE, PPARγ, NOS3, PAI-1, and TGF-β1, comparing DrugGen-2 with DrugGPT and DrugGen across multiple evaluation criteria.These targets were associated with diabetic nephropathy according to DisGeNET and the DrugTar algorithm.
- Model development: DrugGen-2 used six epochs of supervised fine-tuning before GRPO refinement with rewards for binding affinity, validity, and molecular diversity.Binding affinity was evaluated with PLAPT alongside a customized invalid-structure assessor.
- Generation capacity: 409 [406–427] to 444 [443–448] unique generations were achieved by DrugGen-2, versus 50 [41–87] for DrugGen and 219 [217–254] for DrugGPT.Models were tasked with producing 500 unique candidates per selected protein target; the difference was significant (χ2 = 88.80, P < 10-9).
- Molecular validity: 99 to 100 median valid-generation values showed near-perfect structural and chemical validity for DrugGen-2 across all five targets.Validity was assessed using a customized validity assessor on 100 unique molecules per target or MeSH-target pair.
- Similarity and affinity: 0.70 similarity to approved drugs was achieved by DrugGen-2, compared with 0.30 [0.26–0.41] for DrugGPT and 0.64 [0.51–0.76] for DrugGen.The difference was highly significant (H = 961.88, ϵ2 = 0.34, P < 10-204).
- Similarity and affinity: 9.26 to 9.97 median predicted affinities for DrugGen-2 exceeded DrugGPT’s 5.86 to 6.22 and DrugGen’s 7.15 to 8.49 across all five targets.Affinities were predicted using PLAPT, and improvements were highly significant across all targets.
3. Discussion
DrugGen-2 advances disease-conditioned molecular design by generating small molecules tailored to disease ontologies and target protein sequences. The discussion also emphasizes the need for experimental validation and identifies sequence truncation and binary novelty rewards as limitations.
- DrugGen-2 generates small molecules conditioned on both disease ontologies and target protein sequences.
- Limitations and future directions: Experimental validation, including in vitro binding assays and in vivo models, is essential to confirm binding affinities, evaluate functional effects, and assess off-target interactions.
- Limitations and future directions: Truncating long protein sequences to 768 tokens may overlook distal domains critical for allosteric modulation.
- Limitations and future directions: A binary novelty reward might undervalue subtle innovations over radical departures from approved scaffolds.
- Limitations and future directions: Future studies could incorporate ADMET predictions and synthetic accessibility scores to further bridge the gap to clinical development.
4. Materials and Methods · 4.1. Dataset preparation
The study prepared a curated disease–target–drug dataset by linking approved drugs to human targets and associated diseases using DrugBank, DisGeNET, and ChEMBL. Curation produced linked molecular, target-disease, and hierarchical drug-disease resources for model development.
- 4.1. Dataset preparation: Approved drug–target pairs from DrugBank, disease–target associations from DisGeNET, and Phase IV drug–disease relationships from ChEMBL were integrated into disease–target–drug strings.The curation strategy linked approved drugs with corresponding targets and associated diseases to support clinical relevance.
- 4.1.1. Drug-target dataset: 1,710 small molecules had annotated human targets, including 117 withdrawn compounds; 50 were excluded for safety concerns or adverse effects.The selected molecules were further assessed for available SMILES representations from DrugBank, ChEMBL, and ZINC20.
- 4.1.1. Drug-target dataset: 1,634 of 1,660 selected molecules had available SMILES representations obtained from DrugBank, ChEMBL, and ZINC20.The drug–target curation covered 2,116 related protein targets, of which 27 were not present in UniProt.
- 4.1.2. Target-disease dataset: 608 MeSH terms were mapped to 2,042 protein targets after updating disease CUIs from DisGeNET 2019AA to 2023AA and retrieving MeSH terms through the UMLS terminology API.This process formed the target-disease dataset.
- 4.1.3. Drug-disease dataset: 1,299 small molecules and 643 associated MeSH terms were retained after curating approved drug indications from ChEMBL.These records formed the drug–disease dataset.
- 4.1.3. Drug-disease dataset: 1,632 DAG representations resulted from mapping drug-associated MeSH terms to hierarchical disease structures.The DAG mapping captured hierarchical relationships among diseases.
4.2. Data processing
The data-processing pipeline extended the DrugGPT tokenizer and filtered MeSH-sequence-SMILES strings to retain those with direct MeSH–SMILES relationships.
- Tokenizer extension: 53086 tokens comprised the vocabulary after adding “D”, “P”, and “L” to the existing special tokens.The pre-existing special tokens were “startoftext”, “endoftext”, and “PAD”.
- String generation: 1,113,539 MeSH-sequence-SMILES strings were created before relationship-based filtering.These strings linked MeSH sequences with SMILES representations.
- Relationship filtering: 13,908 strings remained after retaining only direct MeSH–SMILES relationships.The filtered strings formed the processed dataset used for subsequent tokenization.
4.3. Model Development
DrugGen-2 was developed through supervised fine-tuning of a pretrained DrugGPT model followed by GRPO refinement. Training optimized molecular validity, predicted binding affinity, novelty relative to approved drugs, and within-batch diversity.
- Supervised fine-tuning: 10 epochs on 13,908 tokenized MeSH–sequence–SMILES strings produced the supervised fine-tuned model, with epoch 6 selected for subsequent processing.The SFT learning rate was 5 × 10−4 and batch size was 8.
- GRPO refinement: GRPO refined the supervised model using group-level statistics to compute advantages without a separate value function, improving memory and computation efficiency.The method optimizes the policy using rewards assigned to generated outputs.
- Generation configuration: 40 completions per prompt were generated during GRPO training within a 1,024-token context window, with maximum input and generation lengths of 768 and 256 tokens.Sampling used top p = 1.0, top k = None, and temperature = 1.0.
- Validity control: RDKit 2023.9.5 filtered syntactically or chemically invalid SMILES during GRPO training through rule-based molecular representation checks.Strings that could not be interpreted or did not produce a valid molecular graph were flagged as invalid.
- Reward design: Three reward objectives guided optimization: predicted binding affinity with invalid-structure penalties, novelty relative to approved drugs, and diversity through preventing batch repetition.Binding affinity was evaluated with PLAPT, a deep learning model measuring affinity between proteins and SMILES.
4.4. Assessment
DrugGen-2 was assessed on diabetic nephropathy using disease ontology–target pairs across molecule generation, validity, approved-drug similarity, binding affinity, and docking analyses. The evaluation compared DrugGen-2 with DrugGen and DrugGPT and examined docking against ACE and PPARγ.
- Experimental setup: Five highly druggable diabetic-nephropathy targets—ACE, PPARγ, NOS3, PAI-1, and TGF-β1—were selected using DisGeNET and DrugTar.DrugGen-2 paired these target proteins with disease information represented by MeSH DAG structures.
- Comparative assessment: DrugGen-2 was compared with DrugGen and DrugGPT across unique-molecule generation, chemical validity, approved-drug similarity, and target binding affinity.DrugGen-2 used MeSH DAG–target protein pairs, whereas the baseline models used target proteins alone.
- Unique-molecule generation: 500 unique small molecules per input were requested, with generation terminated after 30 sampling cycles if novel molecules were not produced.The number of successfully generated unique molecules was used to assess generation capability.
- Chemical validity: 100 unique molecules per input were filtered through a validity-checker pipeline, and the percentage of valid structures measured chemical validity.Statistical analysis used the χ2 test.
- Molecular docking: Docking evaluated newly designed ligands against ACE and PPARγ using GLIDE XP scores, where more negative values indicated stronger predicted binding affinities.ACE and PPARγ structures were selected from the PDB because co-crystallized ligands enabled docking-protocol validation; proteins were rigid and ligands flexible.
Funding
The study and its publication received no funding.
- No funding was received for the study or its publication.