AIBuildAI autonomously develops a state-of-the-art model for cancer gene identification in just four hours
In just four hours, AIBuildAI autonomously discovered a new state-of-the-art AI model for cancer gene identification. Without human intervention, it designed, implemented, trained, and optimized a graph neural network-based model that achieved a mean AUPRC of 0.774, outperforming the published state-of-the-art method, TREE (Nature Biomedical Engineering), which achieved 0.716, and surpassing it on 6 of 8 biological networks. The entire solution is fully open source and reproducible.
Cancer gene identification is a supervised node classification problem that lies at the core of cancer genomics [1]. Given a biological interaction network, multi-omics features for each gene, and a small set of labeled cancer genes, the goal is to identify which of the many unlabeled genes are cancer drivers. Accurately identifying these driver genes is essential for understanding tumor development and advancing precision oncology by enabling the discovery of therapeutic targets, biomarkers, and personalized treatments. Performance is measured by the area under the precision-recall curve (AUPRC) on a held-out test set across eight biological networks.
AIBuildAI is an autonomous system for developing AI models. Given only a task and training data, it designs model architectures, writes code, trains models, analyzes experimental results, and iteratively improves its solutions without human intervention. It organizes this process as a search tree of candidate models, where each new candidate builds on insights from previously evaluated ones. By learning from both successes and failures, the system continually refines its search, enabling it to discover increasingly effective solutions over time.
For this cancer gene identification task, AIBuildAI operated fully autonomously, with complete freedom to design models, implement algorithms, evaluate competing ideas, and allocate its computational budget. Within four hours, it discovered a model that achieved a mean AUPRC of 0.774 across the eight biological networks, outperforming TREE, the published state-of-the-art method reported in Nature Biomedical Engineering [1], which achieved 0.716, and surpassing it on 6 of the 8 biological networks.
How the published state-of-the-art solves this task
The published state-of-the-art method for this task is TREE, a transformer-based graph representation model introduced by Su et al. in Nature Biomedical Engineering [1] (Figure 1). TREE jointly models multi-omics features and the topology of biological interaction networks to learn representations for cancer gene prediction. It is designed to support both homogeneous protein-protein interaction networks and heterogeneous networks that additionally incorporate regulatory relationships. The model is trained separately on each network before being fine-tuned using the labeled cancer genes. Overall, TREE is a sophisticated, purpose-built transformer architecture optimized specifically for this task.
Figure 1. The published state-of-the-art method, TREE (Su et al. [1]). A transformer-based graph representation model jointly learns from multi-omics features and biological network topology to predict cancer genes, with a separate model trained for each biological network.
What AIBuildAI built
Given the same task and training data, AIBuildAI autonomously discovered a markedly different solution from the published approach (Figure 2). Instead of using a transformer, it designed a lightweight spectral graph neural network that combines explicit structural features with multi-task learning. The discovered model consists of three key components.
- A spectral GNN encoder. Two Chebyshev polynomial graph convolution layers aggregate information from neighboring genes and encode each gene into a 100-dimensional embedding. Dropout and DropEdge are applied for regularization.
- Explicit structural features. In addition to the 64-dimensional multi-omics feature vector for each gene, the model computes five graph topology descriptors directly from the interaction network: node degree, PageRank, k-core number, local clustering coefficient, and eigenvector centrality. These features explicitly capture each gene's structural role within the network.
- A multi-task learning head. A self-supervised link prediction objective shares the same encoder as the cancer gene classification task. The two objectives are balanced using learned uncertainty weights, while an exponential moving average stabilizes training. An adaptive training schedule allocates computation across the eight biological networks so that the entire search remains within the time budget. The final model trains in approximately 15 minutes on a single NVIDIA A100 GPU.
Figure 2. The model autonomously discovered by AIBuildAI. A spectral graph neural network integrates each gene's multi-omics profile with explicit structural features derived from the biological interaction network. A classification head predicts cancer genes, while an auxiliary link prediction head shares the encoder to improve representation learning.
Results
We evaluate each biological network as a transductive node classification task, using the same experimental protocol as TREE (Su et al. [1]) for a fair comparison. Both TREE and AIBuildAI are trained on the same labeled genes and evaluated on the same held-out test split. The training, validation, and test sets are mutually exclusive, ensuring that no test labels are available during training. The structural features used by AIBuildAI are computed solely from the network topology and do not incorporate test labels. The benchmark includes six protein-protein interaction networks (CPDB, STRINGdb, PCNet, IRefIndex v15, IRefIndex v9, and Multinet) and two heterogeneous biological networks (MTG and LTG) that additionally encode regulatory relationships. Performance is measured by the area under the precision-recall curve (AUPRC) on the held-out test genes for each network.
Across the eight biological networks, the model autonomously discovered by AIBuildAI outperforms TREE on six (Figure 3), with particularly large improvements on several networks. On MTG, AUPRC increases from 0.540 to 0.773; on LTG, from 0.731 to 0.843; on PCNet, from 0.672 to 0.768; and on Multinet, from 0.686 to 0.777. On the two networks where TREE performs better—CPDB and IRefIndex v15—the differences are relatively small (0.791 vs. 0.779 and 0.815 vs. 0.754, respectively). The largest gains occur on the heterogeneous networks, where incorporating regulatory relationships makes the prediction task more challenging. Overall, AIBuildAI achieves a mean AUPRC of 0.774, compared with 0.716 for TREE.
Figure 3. Per-network AUPRC comparison between the model autonomously discovered by AIBuildAI and the published state-of-the-art method, TREE [1]. Networks are ordered by AIBuildAI's performance advantage. Green bars show AIBuildAI's performance and gray bars show TREE's performance. AIBuildAI surpasses the published state of the art on 6 of the 8 biological networks.
The components of the method discovered by AIBuildAI prove their value in the results. The structural priors derived directly from the network topology, together with a shared link prediction objective, give the lightweight spectral network a clear understanding of both a gene's molecular profile and its position in the interaction network, allowing it to outperform a transformer designed specifically for this task while requiring substantially less computation.
AIBuildAI autonomously discovered this model through iterative exploration and evaluation. Each candidate is assessed on a held-out validation set, and only designs that demonstrate improved generalization are retained for further refinement. Building on the strengths and weaknesses of previously evaluated candidates, AIBuildAI progressively improves its solutions as the search continues. Given only a task and its training data, the system can autonomously discover increasingly effective models over time.
One caveat applies to both AIBuildAI and the published baseline. Better-studied genes tend to accumulate more recorded interactions, meaning that a gene's position in the biological interaction network reflects both biological function and research attention. As a result, the reported performance should be interpreted with this limitation in mind.
What we released
We have open-sourced the complete solution discovered by AIBuildAI, including the full source code, model configurations, and per-network predictions, enabling anyone to reproduce the results or build upon them.
Code and artifacts: github.com/aibuildai/AI-Build-AI/tree/main/tasks/cancer-gene-identification
References
[1] Su, X. et al. Interpretable identification of cancer genes across biological networks via transformer-powered graph representation learning. Nature Biomedical Engineering 9, 371–389 (2025).