🧬 Computational Biology

💬 ACL2026 · 5 paper notes

📌 Same area in other venues: 📷 CVPR2026 (21) · 🔬 ICLR2026 (156) · 🧪 ICML2026 (52) · 🤖 AAAI2026 (20) · 🧠 NeurIPS2025 (76) · 📹 ICCV2025 (4)

🔥 Top topics: Reasoning ×2

AROMA: Augmented Reasoning Over a Multimodal Architecture for Virtual Cell Genetic Perturbation Modeling

The AROMA framework is proposed, which integrates text evidence, knowledge graph topology, and protein sequence features in a multimodal architecture. Combined with a two-stage training strategy (SFT + GRPO), it achieves interpretable and precise genetic perturbation effect prediction.

BioTool: A Comprehensive Tool-Calling Dataset for Enhancing Biomedical Capabilities of Large Language Models

BioTool constructs an instruction fine-tuning dataset consisting of 7,040 human-verified "query–API call" pairs covering 34 commonly used tools from three major biomedical databases: NCBI, Ensembl, and UniProt. After fine-tuning 4B-scale open-source LLMs with this data, the tool-calling quality exceeds commercial models such as GPT-5.1, Gemini-3 Pro, and Claude-4.5-Sonnet by over 15%.

ChemAmp: Amplified Chemistry Tools via Composable Agents

This paper proposes the "Tool Amplification" paradigm (distinct from traditional tool orchestration). Through the ChemAmp framework, chemistry-specific tools (UniMol2, Chemformer, etc.) are treated as composable building blocks to dynamically construct task-specific super-agents. It outperforms specialized models and general LLMs on four core chemistry tasks, including molecular design and reaction prediction, while reducing inference token costs by 94%.

ProtoCycle: Reflective Tool-Augmented Planning for Text-Guided Protein Design

ProtoCycle proposes a reflective agent framework that uses an LLM as a planner combined with a lightweight tool environment for text-guided protein sequence design. It replaces one-shot text-to-sequence generation with a multi-round "plan-tool-evaluate-reflect" cycle. On Mol-Instructions, it improves ProTrek to 14.681 and Retrieval to 0.936, achieving language alignment that nears or exceeds specialized protein design models using only ~2,000 SFT trajectories and online RL.

ToxReason: A Benchmark for Mechanistic Chemical Toxicity Reasoning via Adverse Outcome Pathway

This paper introduces ToxReason, a benchmark for mechanistic chemical toxicity reasoning based on the Adverse Outcome Pathway (AOP) framework. It integrates drug-target experimental data with toxicity labels, requiring models to reason from Molecular Initiating Events (MIE) to organ-level Adverse Outcomes (AO). A 4B model trained with GRPO reinforcement learning outperforms large models like GPT-5 in both toxicity prediction (F1 71.4%) and reasoning quality.