IBM Research this week unveiled MAMMAL, a multi-modal AI model that achieved state-of-the-art performance on 9 of 11 biological benchmarks, outperforming Google’s AlphaFold 3 on several key tasks including antibody-antigen binding prediction. According to research published in Nature, the model combines protein, molecular, and gene expression data in a unified architecture designed for drug discovery applications.
The breakthrough comes as AI systems increasingly tackle complex biological problems that require understanding multiple types of molecular data simultaneously. While AlphaFold 3 and MAMMAL serve different primary purposes, the head-to-head comparison highlights rapid progress in AI-driven biological research.
MAMMAL’s Multi-Modal Architecture
MAMMAL distinguishes itself by integrating three critical data types that previous models handled separately: protein structures, small molecules, and cellular gene expression patterns. This unified approach allows the model to predict how drugs interact with biological systems at multiple levels.
The model excelled particularly in interaction-based tasks that require understanding biological context rather than just molecular structure. Key performance areas include:
- Drug-target interaction prediction: Determining whether specific molecules will bind to target proteins
- Ligand binding affinity: Quantifying how strongly therapeutic compounds attach to their targets
- Antibody-antigen binding: Critical for vaccine development and immunotherapy design
- Gene expression prediction: Modeling how cells respond to drug treatments or genetic changes
Performance Comparison with AlphaFold 3
While both models tackle protein-related problems, they approach biological prediction from different angles. AlphaFold 3 focuses primarily on predicting protein structures and their interactions, while MAMMAL emphasizes functional predictions across multiple biological scales.
MAMMAL’s most significant advantage appeared in antibody-antigen binding prediction, where it substantially outperformed AlphaFold 3. This capability proves particularly valuable for vaccine development and cancer immunotherapy, where understanding how immune system proteins recognize disease targets determines treatment effectiveness.
The model also demonstrated superior performance in molecular property prediction tasks, including toxicity assessment, drug solubility, and chemical stability analysis. These capabilities directly support pharmaceutical development pipelines where early identification of problematic compounds saves significant time and resources.
Broader AI Research Developments
MAMMAL’s release coincides with several other notable developments in AI research. Miami-based startup Subquadratic emerged from stealth this week claiming its SubQ model achieves 1,000x efficiency improvements over existing large language models through a fully subquadratic architecture.
According to VentureBeat, Subquadratic raised $29 million in seed funding at a $500 million valuation, though the AI research community has demanded independent verification of the company’s extraordinary efficiency claims. The startup’s SubQ 1M-Preview model allegedly processes 12 million tokens while reducing attention computation by nearly 1,000 times compared to frontier models.
Meanwhile, researchers demonstrated autonomous scientific discovery capabilities with the Qiushi Discovery Engine. Published on arXiv, the system autonomously identified and experimentally validated “optical bilinear interaction,” a previously unreported physical mechanism analogous to Transformer attention operations.
Time Series Foundation Models Advance
The foundation model approach continues expanding beyond language and vision into specialized domains. Timer-XL, developed by Tsinghua University’s THUML lab, represents the latest advancement in time series forecasting with support for variable context lengths and long-range predictions.
According to Towards Data Science, Timer-XL introduces “TimeAttention,” an attention mechanism specifically designed for temporal data that handles non-stationary univariate series, complex multivariate dynamics, and exogenous variables in a unified framework.
The model’s decoder-only Transformer architecture allows single-model deployment across different forecasting scenarios, eliminating the need for multiple specialized versions that previous approaches required.
What This Means
MAMMAL’s performance suggests multi-modal biological AI models may soon become standard tools in drug discovery pipelines. The ability to simultaneously model protein interactions, molecular properties, and cellular responses could accelerate pharmaceutical development by providing more comprehensive predictions earlier in the research process.
The convergence of advances across multiple AI research areas — from biological modeling to autonomous discovery systems — indicates 2025 may mark a inflection point where AI transitions from assisting human researchers to conducting independent scientific investigations. However, the mixed reception of Subquadratic’s claims underscores the importance of rigorous peer review and independent validation in evaluating breakthrough announcements.
For pharmaceutical companies and biotechnology firms, these developments signal both opportunity and competitive pressure to integrate advanced AI capabilities into their research workflows.
FAQ
How does MAMMAL differ from AlphaFold 3?
MAMMAL focuses on functional predictions across proteins, molecules, and gene expression data, while AlphaFold 3 primarily predicts protein structures. MAMMAL excels at drug-target interactions and binding predictions, making it more directly applicable to pharmaceutical development.
What makes Subquadratic’s efficiency claims controversial?
The startup claims 1,000x efficiency improvements through subquadratic scaling, which would represent a fundamental breakthrough in AI architecture. Researchers are demanding independent verification because such gains would dramatically change the economics of large language model deployment.
What is autonomous scientific discovery?
Systems like Qiushi Discovery Engine can formulate hypotheses, design experiments, collect data, and draw conclusions without human intervention. The recent demonstration marks the first time an AI system independently discovered and validated a previously unknown physical mechanism.
