AutoLeadDesign: LLM-Driven Lead Design

• AutoLeadDesign is a de novo lead design framework that couples a large language model with a dynamically evolving, target-conditioned fragment library.
• The framework uses an iterative closed-loop process combining docking, fragmentation, and LLM-guided molecule generation to refine chemical structures.
• Benchmark experiments and lead design campaigns show that AutoLeadDesign outperforms baselines in achieving expert-competitive binding efficacy and novel binding modes.

Searching arXiv for the specified AutoLeadDesign paper and closely related design/lead-design works to ground the article in current arXiv records. AutoLeadDesign is a property-conditioned de novo lead design framework for computer-aided drug design that couples a LLM with an evolving chemical fragment space to progressively explore chemical space for high-affinity lead compounds with valid, often novel, binding modes. It is positioned at the intersection of fragment-based drug design, de novo molecular design, and structure-based combinatorial optimization: generated molecules are docked to a protein target, decomposed into BRICS fragments, and then used to condition subsequent LLM-guided generation. In benchmark experiments and two target-specific lead design campaigns, the framework is reported to outperform baseline methods and to achieve de novo generation of lead compounds with expert-competitive design efficacy, with structural analyses showing mechanism-validated inhibitory patterns (Tuo et al., 17 Jul 2025).

1. Problem setting and conceptual position

AutoLeadDesign addresses the classical computer-aided drug design problem of navigating a vast chemical space under target-structure constraints while preserving medicinal plausibility. The framework is motivated by three limitations identified in prior approaches. First, traditional generative models such as VAEs, RNNs, and diffusion models tend to learn distributions of known actives and therefore emphasize distributional mimicry and local optimization rather than novel chemotypes or new binding modes. Second, classical combinatorial optimization methods, including genetic algorithms and reinforcement-learning-based systems, encode only weak domain priors and do not fully exploit medicinal chemistry knowledge about fragment linking, pharmacophores, privileged motifs, and structure–activity relationships. Third, LLM-only molecular design systems such as LMLF are described as prone to local optimization around initial compounds and as sensitive to initialization, especially when no strong actives are present (Tuo et al., 17 Jul 2025).

Within that landscape, AutoLeadDesign is explicitly presented as a collaborative framework integrating LLM knowledge with a target-specific fragment space. The central claim is that neither component is sufficient alone: fragments provide a structured representation of target-relevant chemistry learned from docking feedback, while the LLM contributes global chemistry knowledge that can recombine those fragments into chemically coherent molecules. This coupling is described as “mutual inspiration,” because molecules update the fragment space and the fragment space, in turn, conditions the next round of molecular generation.

The framework is also explicitly related to fragment-based drug design. Rather than relying on a fixed fragment library, it constructs an on-the-fly fragment space from all molecules generated so far. This makes the fragment space target-specific and dynamically coupled to the search trajectory. The resulting design process is iterative rather than single-shot, and its operating logic is population-based rather than trajectory-optimized in the reinforcement-learning sense.

2. Closed-loop architecture and generative workflow

At iteration $t$, AutoLeadDesign maintains a compound library $C_t$ containing all molecules generated so far. Each round consists of docking, fragmentation, fragment filtering and weighting, fragment-conditioned LLM generation, and reinsertion of newly generated molecules into the compound library. The loop can begin from a random library drawn from ZINC100K or from the same random library augmented with a known co-crystal ligand, depending on the experimental setting (Tuo et al., 17 Jul 2025).

The molecule decomposition stage uses BRICS, an extension of RECAP built around 16 rules for synthetically meaningful bond disconnections. To avoid trivial substructures, each retained fragment must contain at least two heavy atoms. Fragments are represented primarily as SMILES strings. There is no static predefined fragment inventory; the fragment library is accumulated from all compounds generated and docked so far.

The generative stage is implemented with DeepSeek-v3 through the “deepseek-chat” interface. The model is used as a frozen pretrained LLM, without additional training, and generation is performed at temperature $1.5$, explicitly to encourage diversity and exploration. The prompt template is fixed:

$\min_{s \in \mathcal{S}} Docking(s,p),$0

Here, the bracketed placeholders are replaced by three sampled fragment SMILES. The instruction requires inclusion of at least one of the provided fragments, but does not hard-code a graph assembly rule. Instead, the LLM is expected to perform chemically sensible growing, linking, or merging using its internal chemistry knowledge. Docking then closes the loop: newly generated SMILES are converted to 3D conformers, docked to the target, and used to update both the compound library and the fragment library.

The optimization target is mono-objective. If $Docking(s,p)$ denotes the docking score of molecule $s$ against protein $p$, then the search is effectively directed toward

$\min_{s \in \mathcal{S}} Docking(s,p),$

with more negative docking scores interpreted as better predicted affinity. Other properties, including QED, SAScore, logP, molecular weight, and Lipinski compliance, are computed for analysis rather than used as optimization objectives.

3. Fragment library construction, scoring, and LLM conditioning

The fragment space is updated by decomposing every molecule in the current compound library and aggregating the resulting BRICS fragments. Fragment quality is quantified through a docking-derived contribution score. For a fragment $f$,

$$Score(f)=\frac{\sum_{s \in C,\, f \in Decompose(s)} Docking(s,p)}{|C|}.$$

Here, $C$ is the set of all compounds generated so far, $C_t$0 is the BRICS fragment set of molecule $C_t$1, and $C_t$2 is the docking score against target protein $C_t$3 (Tuo et al., 17 Jul 2025).

The fragment library is then filtered to retain the top-$C_t$4 fragments by score, and a normalized sampling weight is assigned: $C_t$5 where $C_t$6 is the filtered fragment library. Because docking scores are negative for favorable binders, the scoring scheme is framed in terms of affinity contribution rather than positivity. Operationally, the weighting biases generation toward fragments occurring in high-affinity molecules while still preserving stochastic exploration through weighted sampling and high-temperature LLM decoding.

This scoring-and-sampling mechanism is the framework’s main exploitation channel. Exploration arises from three sources: the evolving fragment library, stochastic fragment sampling, and stochastic LLM generation. The paper emphasizes that this combination allows the system to operate effectively even under random initialization, where no strong actives are supplied. In contrast, LMLF is reported to depend strongly on favorable initialization and to cluster around seed molecules.

The role of the LLM is therefore not to rank fragments but to act as a chemistry-aware conditional recombination module. The paper’s structural analyses attribute to the LLM the ability to generate classic medicinal-chemistry motifs such as amide linkers, to preserve privileged aromatic interactions, and to propose modified heterocyclic growth patterns that retain key binding geometries while extending interactions into additional subsites.

4. Structure-based evaluation and benchmark performance

Docking is performed with smina, a fork of AutoDock Vina. Ligand conformers are generated from SMILES using RDKit v2023.9.6 with ETKDG, and protein binding sites are defined from co-crystal ligand coordinates with a $C_t$7 Å grid expansion. For mechanistic inspection rather than optimization, the study uses PyMOL v3.1 for 3D visualization and Maestro v14.2 for 2D interaction analysis. Interaction criteria are stated explicitly: hydrogen bonds require distance $C_t$8 Å with $C_t$9 and $1.5$0; halogen bonds require distance $1.5$1 Å with similar angular constraints; $1.5$2-$1.5$3 stacking is identified as parallel at $1.5$4 Å and angle $1.5$5, or T-shaped at $1.5$6 Å and angle $1.5$7; and $1.5$8-cation interactions require distance $1.5$9 Å and angle $Docking(s,p)$0 (Tuo et al., 17 Jul 2025).

The main benchmark uses 10 random proteins from CrossDocked2020, identified as PDB entries 4AAW, 14GS, 1FMC, 2AZY, 1PHK, 3U5Y, 5NGZ, 1E8H, 5TGN, and 3HY9. Each target is optimized for 20 generations, with 100 compounds designed per generation. Two initialization conditions are studied: random initialization with 100 random molecules from ZINC100K, and prior initialization with the same random set plus the co-crystal ligand. Performance is reported as mean docking scores for the top-1, top-10, and top-100 molecules per target.

Method	Random init, Top-1/10/100	Prior init, Top-1/10/100
ChemGE	$Docking(s,p)$1	$Docking(s,p)$2
RGA	$Docking(s,p)$3	$Docking(s,p)$4
REINVENT	$Docking(s,p)$5	$Docking(s,p)$6
LMLF	$Docking(s,p)$7	$Docking(s,p)$8
AutoLeadDesign	$Docking(s,p)$9	$s$0

These results are used to support two claims. First, AutoLeadDesign outperforms all baselines under both initialization regimes. Second, its relative stability under random initialization contrasts with LMLF, whose top-1 score changes from $s$1 to $s$2 when a known ligand is added, indicating stronger dependence on initialization.

5. PRMT5 and SARS-CoV-2 PLpro lead design campaigns

The framework is further evaluated in empirical lead design campaigns targeting PRMT5 and SARS-CoV-2 papain-like protease (PLpro), two clinically relevant targets emphasized in the study (Tuo et al., 17 Jul 2025).

For PRMT5, the paper considers both a prior-initialization setting, using the known inhibitor LLY-283 together with 100 random ZINC compounds, and a purely random initialization setting. The analysis compares AutoLeadDesign with LMLF. Under prior initialization, t-SNE projection shows that LMLF-generated molecules cluster closely around LLY-283, whereas AutoLeadDesign explores both the local region around LLY and more distant parts of chemical space. The paper contrasts PRL001, an LMLF design that largely preserves the known scaffold and binding pattern, with ADD001, an AutoLeadDesign design exhibiting a novel scaffold and binding mode. ADD001 retains hydrogen bonds with ASP419 and MET420 in the SAM site and additionally forms a hydrogen bond with GLU444 in the protein substrate binding site, leading the authors to characterize it as a dual-site inhibitor. Under purely random initialization, LMLF is described as remaining close to low-affinity initial structures, while AutoLeadDesign generates ADR001, which forms a hydrogen bond with GLU444 despite the absence of any known active seed.

For PLpro, the campaign is conducted under five purely random initializations, each starting from 100 random ZINC compounds. Known inhibitor references include GRL0617 and Jun12682, whose critical interactions with TYR268 and ASP164 are used as structural landmarks. AutoLeadDesign consistently identifies compounds whose docking scores are significantly better than that of Jun1268, reported as $s$3 kcal/mol, and the standard deviation of top docking scores across the five runs is $s$4 kcal/mol, indicating low sensitivity to the initial library. The physicochemical analysis further reports that most generated molecules have QED $s$5, SAScore $s$6, logP in a pharmaceutically acceptable range, molecular weight between 300 and 500 Da, and that more than 78% fully satisfy Lipinski’s Rule of Five.

Two PLpro exemplars are singled out. PLP001 resembles GRL0617 in its binding pattern, preserving $s$7-$s$8 stacking with TYR268 and a hydrogen bond with ASP164 while achieving a better docking score than GRL0617. PLP002 forms a hydrogen bond with ASP164, a $s$9-$p$0 stacking interaction with TYR264, additional hydrogen bonds with TYR273, and a $p$1-cation interaction with ARG166, producing a denser interaction network than the reference inhibitor.

6. Mechanistic interpretation, relation to fragment-based drug design, and limitations

A distinctive feature of AutoLeadDesign is that the paper does not treat the system merely as a black-box optimizer. Instead, it analyzes design trajectories and argues that the method behaves analogously to fragment-based drug design, specifically through fragment linking, fragment merging, and fragment growing (Tuo et al., 17 Jul 2025).

In the fragment-linking example for PLpro, two high-affinity fragments derived from Mol 1 and Mol 2 are provided to the LLM. The resulting Mol 3 joins them with a classic amide linker. The original aromatic fragments preserve their $p$2-$p$3 stacking with TYR268, and the linker introduces a new hydrogen bond with ASP164. The docking score improves by $p$4 and $p$5 kcal/mol relative to the two parent molecules. In the fragment-merging example for PRMT5, two fragments occupy opposite ends of a narrow pocket and cannot be linked directly because of overlap. The LLM generates a merged scaffold in which the overlapping region is replaced with an amino group while maintaining the relative binding orientations associated with LYS393 and PHE327. In the fragment-growing analysis, the evolution of the fragment library is traced across iterations. For PLpro, a pyridine ring emerges in later iterations, preserves geometry with an RMSD of $p$6 Å relative to an earlier fragment, forms $p$7-$p$8 stacking with TYR268, and improves binding free energy by $p$9 kcal/mol.

These analyses are used to explain why the framework works: the fragment library supplies target-conditioned local chemical information, while the LLM performs higher-order medicinal-chemistry operations that resemble expert fragment-based reasoning. This suggests that AutoLeadDesign is not simply sampling around known actives, but is progressively reorganizing a target-specific fragment space under docking feedback.

The paper does not present an explicit limitations section, but several boundaries are evident from the reported setup. The optimization objective is docking affinity alone, and validation is computational rather than biochemical; this suggests that docking dependence and the absence of experimental confirmation remain important constraints on interpretation. Synthetic accessibility is analyzed through SAScore but is not enforced as an optimization target, and no retrosynthetic planning loop is incorporated. Multi-objective criteria such as ADME, selectivity, or toxicity are likewise not part of the search objective. These points do not contradict the reported results; rather, they define the current scope of the framework and indicate plausible directions for extension toward multi-objective, synthesis-aware, and experimentally coupled lead design.