Assay2Mol: LLM-Based Drug Discovery Workflow

• The paper introduces Assay2Mol, a retrieval-augmented LLM workflow that conditions molecular generation on curated BioAssay evidence rather than relying on protein structures.
• It leverages detailed assay texts and quantitative screening metrics from PubChem to generate SMILES with enhanced docking scores (e.g., average -8.064 kcal/mol) and balanced drug-likeness.
• The approach demonstrates counterscreen-aware optimization by reducing off-target predictions such as hERG activity and outperforms several structure-based baselines in key benchmarks.

Searching arXiv for the primary Assay2Mol paper and key related methods mentioned in the provided data. Assay2Mol is a retrieval-augmented LLM workflow for early-stage drug discovery that conditions molecular generation on BioAssay context rather than on protein structures. It leverages unstructured text and quantitative outcomes in public biochemical screening records, primarily PubChem BioAssays, to retrieve relevant assay evidence, summarize protocols and target attributes, and present curated experimental tables of SMILES and activity metrics such as IC50, Ki, Kd, or percent inhibition to a LLM for in-context molecular design. In the reported evaluation on CrossDocked2020, it produces molecules with better docking scores and competitive drug-likeness and synthetic accessibility relative to several structure-based generative baselines, while also demonstrating counterscreen-aware optimization such as reducing predicted hERG channel activity without sacrificing target affinity (Deng et al., 16 Jul 2025).

1. Conceptual basis and scope

Assay2Mol is motivated by the observation that scientific databases aggregate quantitative assay outputs together with descriptive text that specifies biological mechanisms, experimental protocols, target attributes, counterscreen rationale, and disease relevance. The central hypothesis is that assay text and quantitative outcomes encode target function, mechanism, and chemotype–response patterns, and that conditioning an LLM on this context enables in-context learning of actionable structure–activity relationships for molecular design. This replaces the usual dependence of structure-based drug design on a resolved protein structure with a language-mediated conditioning scheme based on previously reported assay evidence (Deng et al., 16 Jul 2025).

The system is positioned explicitly against standard structure-based drug design workflows such as cVAE, flow, diffusion, autoregressive 3D models, and GPT-like models conditioned on protein pockets. Those methods depend on a protein structure and generate 3D ligands atom-by-atom in the pocket. Assay2Mol instead retrieves text-rich BioAssays for related targets, curates assay tables containing actives, inactives, and numeric outcomes, and prompts the LLM to design SMILES directly from this evidence. It therefore requires neither a target structure nor sequence, and the reported formulation extends naturally to phenotypic and cell-based assays.

The motivating data scale is substantial. PubChem alone reports roughly 1.77 million BioAssays comprising approximately 300 million bioactivity outcomes across approximately 250,000 protein targets and thousands of cell lines. Within that setting, Assay2Mol treats assay context as a latent repository of mechanism annotations, experimental setups, and response measurements that can be converted into candidate molecules for new discovery campaigns.

2. Data model, retrieval, and assay curation

The primary source is PubChem BioAssay. These records contain rich text fields and tabular experimental results, including SMILES, categorical labels such as Active, Inactive, and Unspecified, and numeric parameters such as IC50, Ki, Kd, and percent inhibition with relations “<, =, >” and explicit units. For CrossDocked proteins, PDB IDs are mapped to UniProt IDs to query PubChem protein pages when needed, so that the target description used in prompts is accurate. For fair comparison with structure-based baselines, BioAssays directly involving the query protein, matched via UniProt IDs, are excluded from retrieval (Deng et al., 16 Jul 2025).

Assay retrieval is based on text embeddings of protein or phenotype descriptions and BioAssay records. If the query description is embedded as $p$ and each BioAssay record as $b_i$, cosine similarity is used to rank assays:

$$\mathrm{sim}(a,b)=\frac{a\cdot b}{\lVert a\rVert \lVert b\rVert}.$$

Retrieval selects the top-$k$ candidate assays by semantic similarity, after which filtering and LLM-based relevance gating are applied. FAISS is cited for scalable similarity search, and the retrieval index is pre-embedded.

The retrieval layer is deliberately conservative. Retrieved BioAssays are filtered by min_mol_num to prioritize larger assay tables for informative in-context learning, while the number of assays retained in prompt context is capped by max_assay_num to fit LLM context limits; in practice, max_assay_num=10 gave the best trade-off. Molecules drawn from assay tables are balanced by sampling up to Nmol actives and Nmol nonactives, where nonactives combine Inactive and Unspecified examples. If no actives exist, all molecules or up to $2\cdot Nmol$ are included. Context molecules above a size cap, max_mol_size, are filtered to avoid biasing the LLM toward overly large designs; max_mol_size=45 provided good control.

A second relevance filter addresses the limitations of cosine retrieval. Because semantic search can surface pathway neighbors or literature-associated targets with antagonistic or otherwise misleading relationships, an LLM evaluates each candidate BioAssay as Relevant=True/False according to broad biological similarity criteria such as family membership, conserved domains, and functional similarity. The same step identifies counterscreens, including undesirable actives for off-targets or anti-targets and PAINS-related evidence. Manual audits across 25 targets showed that GPT‑4o matched human judgments substantially better than DeepSeek‑V3, with approximate accuracies of 86% versus 43%, respectively. This indicates that model choice affects not only generation quality but also the integrity of the assay context that reaches the prompt.

3. Prompting strategy, model backbones, and post-processing

Assay2Mol operates with multiple LLM backbones: Gemma‑3‑27B, run locally with 4-bit quantization on RTX 5090; DeepSeek‑V3, used via API; and GPT‑4o, also used via API. Molecular generation is performed in SMILES rather than SELFIES. The prompting pipeline follows a layered contextual analysis in three stages: first, each BioAssay is summarized with its purpose, method, key results, relationship to the query target, and counterscreen status; second, the curated experimental table is presented with SMILES, categorical activity labels, and numeric outcomes; third, up to Nmol actives and Nmol nonactives are selected to provide balanced and concise context (Deng et al., 16 Jul 2025).

The generation prompt instructs the model to analyze BioAssays, identify high-affinity molecules associated with low IC50 or Kd, learn scaffolds and functional groups, avoid PAINS, avoid inflating molecule size, and output 10 numbered SMILES strings, each enclosed in [BOS]…[EOS]. It also includes a section asking the model to justify its designs from assay-derived features. Candidate generation is run in batches of 10 SMILES per query target. An ablation without assay context uses only a minimal prompt conditioned on the target description, allowing the incremental contribution of retrieval and curation to be measured directly.

Generated SMILES are parsed with RDKit for validity, and invalid strings are discarded. Valid molecules may then be standardized by sanitization or canonicalization before scoring. RDKit is also used to compute QED and SA. The workflow tracks diversity by mean pairwise fingerprint distance using the TargetDiff evaluation scripts and analyzes similarity between generated molecules and context actives through Morgan fingerprints and Tanimoto similarity. The reported analysis found that most generated molecules were dissimilar to high-scoring context actives, which the authors interpret as evidence that the model is learning from context rather than simply copying examples. Counterscreen awareness is carried through the prompt construction step, where counterscreen actives are treated as negatives for generation.

4. Evaluation protocol and benchmark performance

The main evaluation is conducted on CrossDocked2020 with 100 test protein complexes, and docking scores are computed with AutoDock Vina. Inputs consist of PubChem protein descriptions or UniProt descriptions; BioAssays are retrieved by cosine similarity and LLM relevance gating; the top 10 BioAssays are selected; Nmol=8 is used per class; and max_mol_size=45 is used for context. No model fine-tuning is performed, so the reported system is zero-shot in-context (Deng et al., 16 Jul 2025).

Several metrics structure the evaluation. AutoDock Vina score is reported in kcal/mol, with lower, more negative values indicating better docking. High Affinity is defined as the fraction of generated molecules with Vina scores better than the reference molecule for the target:

$$\mathrm{HighAffinity}=\frac{\#\{g\in G:\mathrm{Vina}(g)<\mathrm{Vina}(\mathrm{reference})\}}{|G|}.$$

Validity is the fraction of valid molecules among generated molecules. QED is RDKit’s quantitative estimate of drug-likeness. The SA score is reported on a transformed scale where higher is better, even though the underlying RDKit synthetic accessibility score is ordinarily lower-is-easier. Diversity is typically the mean pairwise Tanimoto distance on ECFP4 fingerprints.

The head-to-head docking comparison across 100 proteins is summarized below.

Method	Average Vina	Median Vina
Assay2Mol (Gemma‑3‑27B)	-8.064	-8.280
TargetDiff	-7.788	-7.964
TamGen	-7.475	-7.775
Pocket2Mol	-7.200	-6.815
AR	-6.751	-6.707
CVAE	-6.114	-6.118

Assay2Mol with DeepSeek‑V3 and GPT‑4o also improved over TargetDiff, with average Vina values of -7.861 and -7.796, respectively. In terms of drug-likeness and synthesis, Assay2Mol variants had QED averages of approximately 0.585–0.616 and SA averages of approximately 0.790–0.821. TargetDiff averaged approximately 0.474 in QED and approximately 0.584 in SA; TamGen and Pocket2Mol showed higher SA, approximately 0.771 and approximately 0.754, but weaker docking in the case of TamGen.

The reported gains are not uniform across retrieval quality. Grouping targets by LLM-estimated assay relevance, the distribution was High 39%, Medium 42%, Low 7%, and No 12%. Relative to an FDA-drug baseline, Assay2Mol with Gemma‑3‑27B improved docking by approximately +1.04 kcal/mol overall, with approximately +1.28 in High, +1.04 in Medium, +0.54 in Low, and +0.55 in No relevance, outperforming TargetDiff’s normalized gain of approximately +0.76 overall. The trend is important: gains decrease with assay irrelevance, and in “No relevance” cases structure-only methods can be preferable because they are robust to poor assay context.

Ablations reinforce that point. LLM-only generation without assay context still produced molecules with good QED and SA, but docking was weaker; for example, GPT‑4o averaged -7.198 without assay context versus -7.796 when assay-conditioned. Validity also varied substantially by backbone, with GPT‑4o at approximately 94.33%, Gemma‑3‑27B at approximately 84.32%, and DeepSeek‑V3 at approximately 75.84%. Diversity for Assay2Mol, approximately 0.542–0.742, was lower than for some structure-based baselines, reflecting topical clustering induced by context conditioning; batching only 10 molecules was used to mitigate mode collapse. The Gemma‑3‑27B variant also generated larger molecules on average, approximately 26.6 atoms, and the study notes that docking correlates with size, motivating both max_mol_size and prompt-level size constraints.

5. Case studies and counterscreen-aware design

A proof-of-concept case study on GRK4, using PDB 4YHJ and UniProt P32298, illustrates the effect of assay conditioning directly. When GRK family BioAssay context, including GRK2 and GRK5 assays and associated tables, was added to the prompt, average AutoDock Vina docking improved from -7.44 to -8.48 kcal/mol. The top-scoring pose generated with context formed three hydrogen bonds to MET-267 in the GRK4 pocket, whereas generation from target description alone did not report the same interaction pattern (Deng et al., 16 Jul 2025).

A second case study demonstrates sequential counterscreen optimization against hERG. For GRK4 (4YHJ), HD8 (4RN0), and CD38 (3DZH), the workflow first generated candidate binders and then optimized them in a second prompt using hERG BioAssay context. In the 4YHJ case, predicted hERG blocker scores from ADMETlab 3.0 decreased from 0.503 to 0.393 after optimization, while target docking scores were largely retained; the reported FDA-drug average was approximately 0.284. This is presented as an example of context-driven, multi-objective generation guided by counterscreen evidence.

These case studies clarify the practical meaning of “assay context” within the system. It is not limited to affinity readouts. It includes protocol summaries, target annotations, family-level biological relationships, active and inactive exemplars, and explicit anti-target information. The result is a workflow that can use retrieved assay evidence not only to bias molecules toward target affinity but also to steer them away from liabilities when counterscreen records are available.

6. Limitations, reproducibility, and prospective extensions

The reported limitations are principally about retrieval quality, biological ambiguity, and the status of docking as a proxy. Cosine-based text retrieval can return pathway neighbors with opposing functions, exemplified by TF versus TFPI, and even with LLM gating false positives remain. The current pipeline also does not fully support complex logical constraints such as simultaneously inhibiting several targets while avoiding several others in a single pass; hERG avoidance was handled sequentially rather than through one-shot multi-objective generation. Assay text itself is noisy and heterogeneous, so LLM summarization can miss assay-type distinctions or cell-context nuances. In underrepresented “No relevance” settings, assay context helps little and structure-based TargetDiff can be superior. Docking energies are proxy measurements affected by molecule size, and hERG predictions are computational, so wet-lab validation remains necessary. The study also notes dual-use concerns and the absence of a formal risk assessment (Deng et al., 16 Jul 2025).

Reproducibility resources are unusually explicit. Code is released under MIT at https://github.com/gitter-lab/Assay2Mol and archived at DOI 10.5281/zenodo.15871304. Datasets are available at DOI 10.5281/zenodo.15867214 under CC BY 4.0. Reported hyperparameters include max_assay_num=10, Nmol=8, and max_mol_size=45. The listed model versions are GPT‑4o chatgpt-4o-latest and gpt-4o-2024-11-20, DeepSeek‑V3 DeepSeek-V3-0325 and DeepSeek-V3-1226, and Gemma‑3‑27B quantized to 4-bit with unsloth. Docking used AutoDock Vina v1.5.7, and hERG predictions used ADMETlab 3.0. GPT‑4o and DeepSeek‑V3 incurred API costs, whereas Gemma‑3‑27B incurred local compute cost but no API fees.

The future directions named for the framework center on improved assay normalization and retrieval, richer biological context, reaction-aware synthesizability guidance, human-in-the-loop steering, open biomedical embeddings, and conditional multi-target design. A plausible implication is that the core “assay-to-molecule” idea could be extended beyond biochemical screening text to other assay modalities in which assay outputs can be converted into explicit structural constraints. One candidate direction is symbolic structure enumeration from mass spectrometry-derived constraints using Answer Set Programming, where elemental composition, fragment evidence, valence, unsaturation, and canonical graph generation are handled in a declarative search procedure (Küchenmeister et al., 18 Dec 2025). In the form actually reported, however, Assay2Mol remains a language-driven retrieval-and-generation system grounded in PubChem BioAssay records rather than a general symbolic chemistry engine.