Singapore

We study how to impose domain-consistent structure on large language models (LLMs) used for scientific reasoning and early-stage drug discovery. We present MedRule-KG, a compact knowledge-graph scaffold paired with a lightweight verifier that steers generation toward mathematically and biomedically valid outputs. The system injects curated symbolic facts into prompts and then enforces rule satisfaction with a deterministic checker. We formalize generation as constrained inference, introduce a soft guidance surrogate suitable for decoding, and perform a thorough statistical analysis with uncertainty quantification. Across 90 tasks spanning reaction feasibility, metabolic compatibility, and toxicity screening, MedRule-KG reduces violation counts by 83.2% relative to a strong chain-of-thought baseline while improving exact match. Results remain stable under stratification and scale with dataset size, and the verifier adds negligible latency, making the approach practical for interactive design.

AAAI 2026

MedRule-KG: A Knowledge-Graph--Steered Scaffold for Reliable Mathematical and Biomedical Reasoning

workshop paper

We are pleased to announce the Fortieth AAAI Conference on Artificial Intelligence (AAAI-26), which will be held in Singapore EXPO from January 20 to January 27, 2026.

The purpose of the AAAI conference series is to promote research in Artificial Intelligence (AI) and foster scientific exchange between researchers, practitioners, scientists, students, and engineers across the entirety of AI and its affiliated disciplines. AAAI-26 will feature technical paper presentations, special tracks, invited speakers, workshops, tutorials, poster sessions, senior member presentations, competitions, and exhibit programs, and a range of other activities to be announced.<br><br>

To access this event page, you need to log in with the **email address you registered with**. <br>Access credentials will be sent to your email from Underline -  subject line "Welcome to AAAI 2026". Please be sure to check your spam email folder if you do not see an email confirmation right away.

Please log in

To access this event page, you are required to register.
Please complete your registration to continue.

We recommend reading [**the registration information**](https://aaai.org/conference/aaai/aaai-26/registration/) first.

**Online Registration Form**: https://aaai.getregistered.net/conference-2026 

Registration Required

We are pleased to announce the Fortieth AAAI Conference on Artificial Intelligence (AAAI-26), which will be held in Singapore EXPO from January 20 to January 27, 2026.

Parkinson’s disease (PD) drug discovery is uncertainty- and constraint-laden: candidates must match disease mechanisms (e.g., GBA1, 
-synuclein), cross the BBB, and meet developability/novelty requirements amid fragmented evidence. We present PD Brain, a disease-specialized, continuously operating platform that turns natural-language intents into executable end-to-end workflows. PD Brain is built on a PD Deep Knowledge Base (DKB) comprising: (i) an Evidence KG of PD facts, (ii) an Understanding Graph that contracts facts into mechanism motifs (Drug
Target
Pathway
Cell-type/Region
Phenotype), and (iii) a Reasoning Graph that scores causal paths under CNS predicates (BBB transport, SNc expression, exposure thresholds) with calibrated uncertainty. The system \emph{plans, auto-codes, executes, and self-updates} the pipeline.

In a GBA1-focused study, PD Brain expanded an Ambroxol seed to 
1,100 analogs, used a multi-agent MoA debate to triage to 
500, applied variant-aware docking (GCase WT and PD-linked mutants) and BBB-aware ADMET to 64, and shortlisted 5 novel leads. A representative candidate Cpd-1 (SMILES: C1C(=O)CCC(CC1NCC2=C(C(=CC(=C2)Br)Br)N)O) docked to GCase at 
kcal/mol with predicted brain penetration, 
90% human intestinal absorption, 
7.7 h plasma half-life, and moderate toxicity risk; other shortlist molecules exceeded it on specific criteria.

Key contributions: a PD-vertical planner (Plan), automatic tool selection and code synthesis (Design), a provenance-first OS for robust execution (Execute), and continuous DKB-driven updates with drift/contradiction handling (Evolve). We discuss implications for disease-focused automated discovery, including oversight, IP awareness, and continual learning.

Parkinson’s on Purpose: A Disease-Vertical Agentic Pipeline from Literature to Leads

Accurate prediction of patient-specific drug responses to new therapeutic compounds is essential for advancing personalized drug discovery and development. However, limited patient-derived datasets and biological heterogeneity between preclinical and clinical data constrain predictive models generalizability. Existing methods that utilize cell-line drug screens to infer patient outcomes often struggle with distributional shifts and lack biological context.

Here, we have presented DCODE (Domain-aligned Context-aware multimODal attEntion autoencoder), a novel framework that learns domain-consistent molecular representations by aligning latent spaces between preclinical and clinical data. This alignment mitigates distributional shifts while preserving clinically relevant biological signals. A gene–drug interaction attention module further models nonlinear treatment dependencies, allowing DCODE to disentangle the molecular determinants of drug response.

Extensive comparative analyses demonstrated that DCODE outperformed state-of-the-art models, including CODE-AE, in predictive accuracy and robustness. When evaluated on independent clinical datasets covering five drugs for 90808 patients, DCODE achieved over 70% accuracy in predicting responses to cisplatin, sorafenib, and temozolomide. Our results are consistent with existing clinical observations, highlighting its potential for developing personalized therapy and drug response biomarkers.

A domain-aligned multimodal attention autoencoder for predicting personalized clinical drug response from cell-line drug screens

The utility of large language models (LLMs) for reviewing clinical trial protocols—focusing on statistical analysis plans (SAPs) and pharmacokinetics–pharmacodynamics (PK–PD)—was evaluated using a testbed of 15 trials spanning small molecules, biologics, and global public-health studies. GPT-4o, prompted with a regulatory-expert persona, was tasked to extract study-design attributes, surface relevant regulatory guidance, and produce structured SAP and PK–PD evaluations. Outputs were benchmarked against source protocols/registries and graded post hoc by ChatGPT and Grok using a predefined rubric. Disease, interventions/comparators, and trial sample size were correctly identified for 15/15, 15/15, and 14/15 trials, respectively; overall protocol-based correctness was 95% (80% versus ClinicalTrials.gov actuals reflecting realized enrollments). Frequently surfaced guidance included ICH/FDA E9 for SAPs and FDA Population Pharmacokinetics for PK–PD. SAP summaries were generally coherent, with occasional context errors (e.g., outdated endpoint adjudication, sidedness of tests, incomplete multiplicity handling). PK–PD objectives/measures were correctly captured in 8/9 protocols and analysis methods in 7/8, including dose-proportionality reasoning for a Phase I example. Cross-model grading showed similar judgments on primary-outcome accuracy, statistical-method correctness, and E9 alignment, with divergence in clinical-interpretability scores. Findings support an LLM-assisted, human-in-the-loop workflow to triage and standardize protocol reviews at scale, while highlighting needs for version awareness, explicit multiplicity control, and more precise guidance retrieval in complex or adaptive designs.

Large Language Models for Clinical Trial Protocol Assessments

The Ribonucleic Acid (RNA) inverse folding problem, designing nucleotide sequences that fold into specific tertiary structures, is a fundamental computational biology problem with important applications in synthetic biology and bioengineering. The design of complex three-dimensional RNA architectures remains computationally demanding and mostly unresolved, as most existing approaches focus on secondary structures. In order to address tertiary RNA inverse folding, we present BeeRNA, a bio-inspired method that employs the Artificial Bee Colony (ABC) optimization algorithm. Our approach combines base-pair distance filtering with RMSD-based structural assessment using RhoFold for structure prediction, resulting in a two-stage fitness evaluation strategy. To guarantee biologically plausible sequences with balanced GC content, the algorithm takes thermodynamic constraints and adaptive mutation rates into consideration. In this work, we focus primarily on short and medium-length RNAs (
100 nucleotides), a biologically significant regime that includes microRNAs (miRNAs), aptamers, and ribozymes, where BeeRNA achieves high structural fidelity with practical CPU runtimes. The lightweight, training-free implementation will be publicly released for reproducibility, offering a promising bio-inspired approach for RNA design in therapeutics and biotechnology.

BeeRNA: tertiary structure-based RNA inverse folding using Artificial Bee Colony

Antibodies recognize antigens via complementary and structurally dependent mechanisms. Therefore, inclusion of antibody inputs is crucial for accurate epitope prediction. Given limited availability of antibody-antigen complex structures, any epitope prediction model will require minimal yet sufficient antibody inputs to ensure precise epitope identification. To address this need, we introduce Epi4Ab, an antibody-specific epitope prediction model that focuses on identifying unique in-contact antigen residues for a given antibody. Epi4Ab requires minimal antibody inputs, specifically VH/VL families and complementarity-determining region sequences. This work has been published in mAbs with doi:10.1080/19420862.2025.2531227.

Epi4Ab: A data-driven prediction model of conformational epitopes for specific antibody VH/VL families and CDRs sequences

Alzheimer’s disease (AD) is a widely studied neurodegenerative disorder, and researchers have endeavored to identify target genes for drug development. However, most existing approaches for target identification rely on correlation-based signals, which often fail to distinguish causal drivers from correlated markers. We present a practical causal discovery framework for gene-level target identification that learns local causal neighborhoods around key phenotypes from single-nucleus RNA-seq data. It integrates AD domain knowledge to guide structure learning and employs stability selection for robust ranking under sampling variability. Leveraging the genetically diverse AD-BXD mouse panel, our method achieves competitive performance against baselines in empirical evaluations and yields candidates consistent with established biological evidence.

Utilizing Local Causal Structure Learning in Target Gene Identification for Alzheimer’s Disease

We introduce Temporal Eigenstate Networks (TEN), a novel architecture for sequence mod eling that achieves O(T) complexity compared to the O(T2) complexity of transformer attention mechanisms, where T is the sequence length. TEN operates by decomposing temporal dynam ics into learned eigenstate superpositions that evolve through complex-valued phase rotations, enabling efficient capture of both short and long-range dependencies. Our approach combines insights from spectral theory, state-space models, and quantum mechanics to create a mathe matically principled framework that eliminates the attention bottleneck while maintaining or exceeding transformer performance. On benchmark tasks, TEN achieves 3-28× speedup over transformers on sequences of length 512-8192 while using 120× less memory, with competitive or superior accuracy on language modeling, long-range reasoning, and time-series prediction. We provide theoretical analysis proving universal approximation capabilities, stability guaran tees, and efficient gradient flow. TEN opens a new direction for scalable sequence modeling that is particularly promising for extremely long sequences, edge deployment, and AGI systems requiring efficient temporal reasoning

Temporal Eigenstate Networks: Linear-Complexity Sequence Modeling via Spectral Decomposition

Reliable, informative, and individual uncertainty quantification (UQ) remains missing in applied machine learning (ML). This hinders the effective application of AI/ML to risk-sensitive drug discovery pipeline. Most methods either fail to provide coverage on new data, inflate intervals so broadly that they are not actionable, or assign uncertainties that do not track actual error, especially under a distribution shift. Protein-ligand binding affinity prediction is especially challenging as assay noise is heterogeneous, chemical space is biased and large, and practical evaluations routinely involve distribution shift. In this work, we introduce a novel uncertainty quantification method, 
Trustworthy Expert Split-conformal with Scaled Estimation for Efficient Reliable Adaptive intervals (TESSERA), that provides per-sample uncertainty with reliable coverage guarantee, informative and adaptive prediction interval widths that track the absolute error. We evaluate on protein–ligand binding affinity prediction under both independent and identically distributed (i.i.d.) and scaffold-based out-of-distribution (OOD) splits, comparing against strong UQ baselines. TESSERA attains near-nominal coverage while maintaining strong efficiency and adaptivity. The intervals expand naturally in regions where data are sparse or uncertainty is high, and contract in well-sampled regions where predictions are reliable, keeping overall coverage consistent across the domain. This adaptivity ensures that uncertainty tracks model error closely, making the predictions actionable. TESSERA produces uncertainty estimates that are simultaneously well-calibrated, efficient, and responsive to data difficulty. These are crucial requirements for trustworthy decision-making in drug discovery.

Adaptive Individual Uncertainty Quantification with Expert-Routed Conformal Prediction for Protein-Ligand Binding Affinity Prediction

An AI agent, faced with the fundamental challenge of validating causal structures in static networks, autonomously developed a novel validation protocol, transforming a causal graph into a sequence for wavelet coherence analysis. This method provided a more direct test of mechanistic coordination than standard graph-based techniques, proving its efficacy with overwhelming statistical significance (p < 0.001). The agent’s core insight was transforming hierarchical causal graphs into analyzable sequences via depth-first search, then applying wavelet coherence analysis to quantify mechanistic coordination. This protocol was a key component of a fully autonomous pipeline that subsequently achieved state-of-the-art drug discovery results (AUPR = 0.87 ± 0.02) from 100,000 protein interactions, recovering known cancer drivers (TP53, TNF) while processing 242,843 decision tokens with zero human interventions. The agent’s methodological innovation represents autonomous identification of research gaps and development of solutions. Key contribution: A novel validation protocol for causal graph validation in static networks (p < 0.001).

Agentic Causal Graph Learning for Drug Target Discovery: A Self-Directed AI System at STRING Scale

Predicting the treatment response for a specific drug in a specific situation is a key goal in drug development and personalized medicine. A vast amount of relevant data is generally available, ranging from microscopic, such as the molecular structure of proteins, to the macroscopic, such as response data for other drugs in other situations. Neural networks are well suited for integrating these complex multimodal datasets and connecting them to the final prediction target. In this work, we adopt an existing method for predicting tumor volume progression using Neural Ordinary Differential Equations (Neural-ODEs) and Graph Neural Networks (GNNs) on patient-derived xenograft (PDX) data. We implement this method in a flexible research platform (LynxKite) and demonstrate its potential by extending the input dataset with protein embeddings from ESM-2. The addition of ESM-2 embeddings increased the accuracy of predictions from 68.1% to 75.5%.

Premium content

Next from AAAI 2026

Parkinson’s on Purpose: A Disease-Vertical Agentic Pipeline from Literature to Leads

Stay up to date with the latest Underline news!

PRESENTATIONS

CONFERENCES

COMPANY

RESOURCES