TESSERA: LLM-Guided Monte Carlo Tree Search for Drug-Disease Mechanism Explanations
Researchers have introduced TESSERA, a novel neuro-symbolic framework designed to extract multi-step mechanistic explanations for drug-disease pairs from knowledge graphs. Addressing the combinatorial challenges of path exploration and credit assignment in complex biological data, TESSERA integrates Large Language Models (LLMs) with Monte Carlo Tree Search (MCTS). Unlike autonomous generative approaches, this framework employs LLMs in a circumscribed role for local discriminative judgments, serving as both a prior policy to bias exploration and a comparative state evaluator providing reward signals. The knowledge graph enforces hard structural constraints on the hypothesis space, while MCTS coordinates long-horizon search with principled backpropagation. Evaluations across two complementary knowledge graphs demonstrate that TESSERA maintains fidelity to curated biological facts while uncovering coherent alternative mechanisms. Ablation studies confirm the distinct contributions of the LLM components. This approach offers a general paradigm for compositional reasoning over structured knowledge, mitigating the degradation of performance often seen in lengthy reasoning chains generated by standalone LLMs.
Wire timeline
TESSERA: LLM-Guided Monte Carlo Tree Search for Drug-Disease Mechanism Explanations
Researchers have introduced TESSERA, a novel neuro-symbolic framework designed to extract multi-step mechanistic explanations for drug-disease pairs from knowledge graphs. Addressing the combinatorial challenges of path exploration and credit assignment in complex biological data, TESSERA integrates Large Language Models (LLMs) with Monte Carlo Tree Search (MCTS). Unlike autonomous generative approaches, this framework employs LLMs in a circumscribed role for local discriminative judgments, serving as both a prior policy to bias exploration and a comparative state evaluator providing reward signals. The knowledge graph enforces hard structural constraints on the hypothesis space, while MCTS coordinates long-horizon search with principled backpropagation. Evaluations across two complementary knowledge graphs demonstrate that TESSERA maintains fidelity to curated biological facts while uncovering coherent alternative mechanisms. Ablation studies confirm the distinct contributions of the LLM components. This approach offers a general paradigm for compositional reasoning over structured knowledge, mitigating the degradation of performance often seen in lengthy reasoning chains generated by standalone LLMs.
cs.AI updates on arXiv.org