DrugMCTS: MCTS for Drug Discovery & Design

• DrugMCTS is a family of Monte Carlo Tree Search frameworks that apply sequential decision making to drug repurposing, de novo molecular generation, and retrosynthetic planning.
• It combines diverse techniques including partial SMILES token generation, graph-based edits, and AND–OR tree structures to effectively balance exploration and exploitation in chemical search spaces.
• Notable implementations demonstrate improved molecular property optimization, scalability in parallel computations, and enhanced recall in identifying protein–ligand interactions.

Searching arXiv for DrugMCTS and closely related MCTS-based molecular design papers. {"query":"DrugMCTS arXiv DrugMCTS molecular design Monte Carlo Tree Search", "max_results": 10} {"query":"all:DrugMCTS", "max_results": 5} DrugMCTS denotes Monte Carlo Tree Search-based systems for drug discovery. In the literature summarized here, the term is used in two closely related senses. First, it names a 2025 drug repurposing framework that combines retrieval-augmented generation, multi-agent collaboration, and Monte Carlo Tree Search to identify proteins likely to interact with a query small molecule (Yang et al., 10 Jul 2025). Second, it functions as a broader label for earlier MCTS-driven workflows in de novo molecular generation and synthesis planning, where search is conducted over partial SMILES strings, chemically valid graph edits, or retrosynthetic AND–OR trees rather than over free-form language outputs (Yang et al., 2020, Rajasekar et al., 2020, Hong et al., 2021). Across these variants, the unifying idea is to cast chemical design or biochemical reasoning as sequential decision making under expensive or noisy evaluation, and to use MCTS to balance exploitation of promising branches against exploration of under-sampled ones.

1. Scope and conceptual variants

In the broadest technical sense, DrugMCTS is a family of search systems in which molecules, molecular transformations, or reasoning trajectories are represented as tree nodes and evaluated through domain-specific reward functions. The search space differs substantially by application. In de novo generation, the tree may consist of partial SMILES strings whose children append one valid token at a time (Yang et al., 2020). In graph-based optimization, each node is a chemically valid molecular graph and each edge is a single atom or bond edit (Rajasekar et al., 2020). In drug repurposing, the tree can instead encode a sequence of agent actions such as retrieval, molecule analysis, molecule selection, interaction analysis, and protein selection (Yang et al., 10 Jul 2025). In retrosynthesis, the tree becomes an AND–OR structure over molecules and reaction templates (Hong et al., 2021).

A useful way to distinguish the main variants is by their state and action abstractions.

System	State and action space	Reward or objective
MP-MCTS molecular design	Partial SMILES strings; next-token additions	Penalized logP, absorption wavelength, or Zaleplon MPO (Yang et al., 2020)
unitMCTS	Chemically valid molecular graphs; unit graph modifications	QED or penalized logP (Rajasekar et al., 2020)
DrugMCTS repurposing	Partial multi-agent reasoning state; actions $A_1$ to $A_6$	Self-consistency reward over protein predictions (Yang et al., 10 Jul 2025)
EG-MCTS retrosynthesis	Sets of molecules and reaction-template applications	Success, failure, or downstream route value (Hong et al., 2021)

This taxonomy clarifies a common misconception: DrugMCTS is not restricted to a single molecular representation or a single task. The available work spans generation, optimization, repurposing, and route planning, with MCTS serving as the common control mechanism rather than the common data modality.

2. Search spaces, state transitions, and rewards

The SMILES-based formulation in "Practical Massively Parallel Monte-Carlo Tree Search Applied to Molecular Design" represents the root as a start symbol and each depth-$d$ node as the first $d$ SMILES tokens of a partial string (Yang et al., 2020). Expansion is guided by a GRU that produces a distribution over a vocabulary of size 64, and children are added in descending probability order until cumulative probability reaches 0.95. Rollout continues token sampling until EOS, after which RDKit is used to compute penalized logP or another property and map it to a reward $r \in [-1,1]$. This formulation makes MCTS a sequence generator with continuous rewards rather than a win/loss game search.

The graph-based formulation in "Goal directed molecule generation using Monte Carlo Tree Search" defines an MDP $M=(S,A,R,P)$ in which states are chemically valid molecules and actions are unit modifications: atom addition, bond addition, bond removal, and bond replacement (Rajasekar et al., 2020). The allowed atom set in the experiments is $\{ \mathrm{C}, \mathrm{N}, \mathrm{O} \}$. RDKit validates transitions, disconnected atoms are dropped automatically after bond removal, and bridge atoms are disallowed. The transition model is deterministic, and the terminal reward is the property value of the final molecule, either QED or penalized logP. In constrained optimization, similarity filtering is enforced during expansion, so the search remains within a local neighborhood of the starting molecule.

The repurposing-oriented DrugMCTS framework shifts the notion of state from chemistry itself to structured reasoning over biochemical evidence (Yang et al., 10 Jul 2025). Its input is a query molecule $M_{qm}$ given as SMILES, and its output is a ranked set of proteins likely to interact with that molecule. The retrieval agent constructs a candidate molecule pool $M_{cm}$ from the union of the top 10 Tanimoto-similar and top 10 ChemBERTa-similar molecules, after deduplication; the candidate protein pool $P_{cp}$ is then assembled from proteins interacting with those molecules. Subsequent agents produce a query report $A_6$0, a subset of reference molecules $A_6$1, interaction reports $A_6$2, and finally a selected protein set $A_6$3. The rollout reward is derived from LLM self-consistency:

$A_6$4

$A_6$5

$A_6$6

In retrosynthetic planning, EG-MCTS models a state $A_6$7 as a set of molecules, with initial state $A_6$8, and an action $A_6$9 applying a reaction template $d$0 to molecule $d$1 (Hong et al., 2021). The deterministic transition replaces $d$2 by its reactants. Reaction nodes are rewarded with $d$3 if proven successful, $d$4 if proven unsuccessful, and otherwise with the average of downstream molecule-node values. This reward is route-centric rather than property-centric. A plausible implication is that DrugMCTS should be viewed not as one reward design, but as a search template into which different oracle structures can be plugged.

3. Core MCTS algorithms and architectural adaptations

All of the systems rely on the canonical MCTS phases of selection, expansion, simulation or evaluation, and backpropagation, but their tree policies differ materially. The sequential molecular design baseline uses UCB1,

$d$5

with $d$6 the cumulative reward, $d$7 the child visit count, $d$8 the parent visit count, and $d$9 the exploration constant (Yang et al., 2020). To support distributed search, MP-MCTS introduces virtual loss:

$d$0

where $d$1 counts workers currently exploring child $d$2, and $d$3. Each worker temporarily adds virtual visits with zero reward to discourage others from entering the same subtree. The decisive modification beyond standard TDS-df-UCT is that each node stores its own history table of sibling statistics, allowing backpropagation to stop before the root when the local subtree remains best. This produces an effectively depth-first search in promising regions and reduces root communication.

unitMCTS applies a more conventional UCT-style tree policy over molecular graphs:

$d$4

where $d$5 is cumulative reward, $d$6 is visit count, and $d$7 is the exploration constant (Rajasekar et al., 2020). Its domain-specific deviations occur primarily in expansion and backup. Expansion adds the top $d$8 chemically valid children, initializes each with $d$9, and assigns $r \in [-1,1]$0 using an $r \in [-1,1]$1-greedy policy. Backpropagation uses exponentially scaled rewards with factor $r \in [-1,1]$2. The paper emphasizes that there is no learning involved, so tree search itself bears the full burden of value estimation.

The 2025 DrugMCTS framework uses UCT over reasoning trajectories rather than chemical edits (Yang et al., 10 Jul 2025). Each non-retrieval agent call can generate multiple candidate outputs, which become child nodes; analysis and selection actions can generate up to 4 children, whereas Protein Selection and End generate 1 child. Leaf evaluation is not a docking score or a deep DTI predictor but an internal self-consistency computation by the same LLM that powers the agents. This is methodologically notable because it converts LLM agreement patterns into a reinforcement signal without domain-specific fine-tuning.

EG-MCTS departs still further from vanilla MCTS by removing rollout altogether (Hong et al., 2021). At molecule nodes, it uses a PUCT-style selection rule that combines an average action value $r \in [-1,1]$3, the single-step retrosynthesis prior $r \in [-1,1]$4, and an exploration constant $r \in [-1,1]$5. The initial value $r \in [-1,1]$6 comes from an Experience Guidance Network rather than from random simulation. During update, molecule-node values are set to the maximum of child reaction-node values. This design is closer to AlphaZero-style prior-plus-value search than to ChemTS-style rollout-based MCTS.

4. Data interfaces, models, and systems engineering

The de novo molecular-design instantiation couples MCTS to a simple GRU inherited from ChemTS (Yang et al., 2020). SMILES tokens are encoded as 64-dimensional one-hot vectors; the model contains two GRU layers with output dimension 256 and a dense 64-way softmax output layer. It is pre-trained on approximately 250K drug-like molecules from ZINC and is used both as the expansion policy and as the rollout policy. The distributed implementation uses MPI via mpi4py, Keras with TensorFlow backend for GRU inference, and RDKit for parsing and property computation. Nodes are distributed across workers by hash-driven ownership, with

$r \in [-1,1]$7

and an overload factor $r \in [-1,1]$8 maintains the number of jobs in the system at $r \in [-1,1]$9.

unitMCTS eliminates the learned policy entirely and delegates validity to explicit cheminformatics rules (Rajasekar et al., 2020). RDKit mediates graph manipulation, valency checking, and invalid-action rejection, yielding 100% validity by construction. The resulting system is dataset-free at inference and does not depend on latent-space optimization, policy learning, or a value network. This makes the graph-edit MDP directly interpretable: every tree edge corresponds to one primitive chemical operation.

DrugMCTS, in its 2025 sense, relies on a more elaborate data stack (Yang et al., 10 Jul 2025). Qwen2.5-7B-Instruct is used for the Molecule-Analysis, Molecule-Selection, Interaction-Analysis, Decision, and reward-calculation roles, while retrieval is tool-based. The framework distinguishes scientific data, hybrid scientific–general data, and general data. RDKit and PubChemPy transform SMILES into structural and physicochemical descriptors; PLIP extracts binding pocket data from PDB files and converts raw tables into descriptive paragraphs; PubMed retrieval injects literature context. The authors explicitly frame this as a scientific $M=(S,A,R,P)$0 hybrid $M=(S,A,R,P)$1 general data pipeline. A common misconception is that RAG alone is the central innovation. The results suggest instead that the decisive factor is structured transformation of scientific data plus MCTS-mediated reasoning, because GM+RAG can perform worse than GM alone.

EG-MCTS integrates MCTS with a template-based single-step retrosynthesis model and an Experience Guidance Network (Hong et al., 2021). The EGN receives a 4096-dimensional concatenation of a 2048-bit Morgan fingerprint for the molecule and a 2048-bit reaction-template fingerprint, passes it through a 256-unit hidden layer with dropout 0.1, and outputs a scalar $M=(S,A,R,P)$2. It is trained with Adam for 20 epochs per iteration on experiences $M=(S,A,R,P)$3 collected from prior searches. This makes the guidance signal endogenous to the search process itself.

5. Empirical performance across tasks

In massively parallel de novo generation, MP-MCTS was reported to work efficiently at 1,000 worker scale and to maintain search quality at larger scale (Yang et al., 2020). On penalized logP, with 10 minutes of search averaged over 10 runs, MP-MCTS achieved $M=(S,A,R,P)$4 at 64 cores, $M=(S,A,R,P)$5 at 256 cores, and $M=(S,A,R,P)$6 at 1024 cores. The corresponding TDS-UCT results were $M=(S,A,R,P)$7, $M=(S,A,R,P)$8, and $M=(S,A,R,P)$9, while TDS-df-UCT achieved $\{ \mathrm{C}, \mathrm{N}, \mathrm{O} \}$0, $\{ \mathrm{C}, \mathrm{N}, \mathrm{O} \}$1, and $\{ \mathrm{C}, \mathrm{N}, \mathrm{O} \}$2. The paper further states that running MP-MCTS on 256 CPU cores for only 10 minutes produced candidate molecules having similar score to non-parallel MCTS running for 42 hours, and that tree depth at 256 cores for 10 minutes was approximately equal to the depth of non-parallel MCTS run for $\{ \mathrm{C}, \mathrm{N}, \mathrm{O} \}$3 minutes. On the same benchmark, MP-MCTS using the same GRU as ChemTS reported top-3 penalized logP scores of 15.13, 14.77, and 14.48, compared with 6.47, 5.65, and 5.01 for ChemTS GRU-based MCTS.

In graph-based goal-directed generation, unitMCTS reported strong benchmark performance without learning (Rajasekar et al., 2020). For penalized logP, the top score was 12.63 for unitMCTS-38 and 29.20 for unitMCTS-84; the reported baselines in the same table included GCPN at 7.98, MolDQN-bootstrap at 11.84, MSO at 26.10, and PGFS at 27.22. For QED, unitMCTS-38 and unitMCTS-84 both reported 0.948, matching the best reported values. In constrained optimization over 800 ZINC molecules with low penalized logP, unitMCTS-20 improved penalized logP by $\{ \mathrm{C}, \mathrm{N}, \mathrm{O} \}$4 at similarity threshold $\{ \mathrm{C}, \mathrm{N}, \mathrm{O} \}$5, $\{ \mathrm{C}, \mathrm{N}, \mathrm{O} \}$6 at $\{ \mathrm{C}, \mathrm{N}, \mathrm{O} \}$7, $\{ \mathrm{C}, \mathrm{N}, \mathrm{O} \}$8 at $\{ \mathrm{C}, \mathrm{N}, \mathrm{O} \}$9, and $M_{qm}$0 at $M_{qm}$1, outperforming JT-VAE, GCPN, and MolDQN-bootstrap at every threshold.

In drug repurposing, the 2025 DrugMCTS framework reports substantially higher recall than both general-purpose LLMs and deep learning DTI baselines (Yang et al., 10 Jul 2025). On DrugBank, GPT-4o-mini and Deepseek-R1 in the GM setting achieved recall 0.1552 and 0.1619, while DrugMCTS achieved 0.4466 in TopK and 0.5534 in TopK+3. On KIBA, GPT-4o-mini and Deepseek-R1 in the GM setting achieved 0.2580 and 0.2645, while DrugMCTS achieved 0.4224 in TopK and 0.4924 in TopK+3. The deep learning ensemble baseline achieved 0.2364 on DrugBank and 0.3216 on KIBA. Ablation results on DrugBank show that a baseline without MCTS scored 0.1285, adding detailed pocket information and literature without MCTS scored 0.1586, and full DrugMCTS reached 0.4466. The authors also report a case study in which the system proposed an interaction between Equol and CXCR3, later examined by molecular docking with AutoDock Vina; the docking score was $M_{qm}$2 kcal/mol.

In retrosynthesis, EG-MCTS improves both planning efficiency and route quality (Hong et al., 2021). On a 180-molecule test set with a 500-iteration limit, EG-MCTS achieved a 94.44% success rate and 60.75 average iterations for successful solves, compared with 90.56% and 85.97 for Retro$M_{qm}$3+, 81.11% and 128.96 for EG-MCTS-0, and 78.89% for MCTS-rollout. On the Retro$M_{qm}$4-190 benchmark at 500 iterations, EG-MCTS reached 96.84% success, compared with 91.05% for Retro$M_{qm}$5+ and 62.63% for MCTS-rollout. For route quality on the subset solved by all methods, EG-MCTS reported average route length 5.85 on one test set and 5.07 on Retro$M_{qm}$6-190, whereas Retro$M_{qm}$7+ reported 8.37 and 6.03.

6. Limitations, misconceptions, and extensions

The limitations differ sharply by instantiation. In the GRU-guided molecular generator, the search is constrained by the support of the fixed generative model; if the GRU under-represents a chemistry, MCTS cannot discover it (Yang et al., 2020). The main experiments also focus on penalized logP, which is explicitly noted to incentivize unrealistic, heavy, and large molecules, and the authors acknowledge that very high-scoring molecules may not be practically useful. Infrastructure is another constraint: MP-MCTS is designed for large CPU clusters up to 1024 workers.

unitMCTS avoids model-support limitations but has its own scope restrictions (Rajasekar et al., 2020). The experiments use only the atom types $M_{qm}$8, and the optimization is single-objective. Synthetic feasibility is not directly enforced beyond the synthetic accessibility component in penalized logP, and branching factor can still grow substantially as the action vocabulary expands. The paper proposes MolMCTS, in which non-unit modifications or fragment additions would replace single-atom or single-bond edits.

The repurposing version of DrugMCTS demonstrates that MCTS can orchestrate tool-using LLM agents effectively, but it does not eliminate dependence on incomplete biological knowledge bases (Yang et al., 10 Jul 2025). The authors note that absolute recall remains around 55% at best, that higher-level biological context such as pathways and tissue expression is not integrated, and that self-consistency may amplify consistent but wrong beliefs. They also report that increasing rollouts from 12 to 24 yields diminishing or negative returns except in some KIBA TopK+3 settings, indicating a performance plateau for the current search configuration.

EG-MCTS, finally, inherits well-known retrosynthesis assumptions (Hong et al., 2021). Reaction conditions, reagents, catalysts, stoichiometry, and solvents are ignored; some proposed intermediates may be unstable; and route quality is evaluated primarily through route length, which can favor shorter but less realistic plans. Template fingerprints are coarse and may incur collisions. The authors suggest forward reaction prediction, graph neural network representations, and richer multi-objective rewards as natural improvements.

Taken together, these results suggest that DrugMCTS is best understood not as a monolithic algorithm but as a reusable search paradigm. A plausible implication is that future systems will hybridize the strongest components already demonstrated separately: the scalable distributed tree management of MP-MCTS, the chemically valid graph-edit semantics of unitMCTS, the structured scientific-data transformation and multi-agent reasoning of the 2025 DrugMCTS framework, and the learned action-value guidance of EG-MCTS.