IBM Research on Tuesday released MAMMAL, a multi-modal AI model that combines protein, molecular, and gene data to achieve state-of-the-art performance on 9 out of 11 biological benchmarks — surpassing Google’s AlphaFold 3 on several key tasks. According to IBM’s Nature publication, MAMMAL demonstrates particular strength in antibody-antigen binding prediction and drug-target interaction tasks.
The model represents a significant shift toward multi-modal biological AI, integrating diverse data types rather than focusing solely on protein structure prediction like AlphaFold 3. IBM researchers designed MAMMAL to handle interaction-based and contextual biology tasks that require understanding relationships between different biological components.
MAMMAL’s Benchmark Performance
MAMMAL achieved superior results across nine critical biological prediction tasks. The model excelled in drug-target interaction prediction, determining whether specific molecules will bind to target proteins — a fundamental challenge in pharmaceutical development.
In ligand binding affinity prediction, MAMMAL accurately estimates how strongly drugs bind to their targets, providing quantitative measurements essential for drug dosing and efficacy studies. The model also demonstrated breakthrough performance in antibody-antigen binding prediction, a capability crucial for vaccine development and immunotherapy design.
Gene expression prediction represents another area where MAMMAL outperformed existing models. The system can forecast how cells respond to drugs or environmental changes by analyzing genetic activity patterns. This capability enables researchers to predict therapeutic effects and potential side effects before conducting expensive clinical trials.
Multi-Modal Architecture Advantages
Unlike AlphaFold 3’s protein structure focus, MAMMAL processes multiple biological data types simultaneously. The model integrates protein sequences, molecular structures, and cellular context to make predictions about complex biological interactions.
This multi-modal approach enables MAMMAL to perform cross-domain generalization, applying knowledge learned from one biological system to different contexts. For example, insights from bacterial protein interactions can inform predictions about human drug responses.
The model’s molecular property prediction capabilities extend beyond binding affinity to include toxicity assessment, solubility analysis, and stability measurements. These properties determine whether potential drugs can be safely administered and remain effective in biological systems.
Autonomous Discovery Breakthrough
Separately, researchers demonstrated the first end-to-end autonomous scientific discovery system on a real physical platform. The Qiushi Discovery Engine conducted independent optical physics research, consuming 145.9 million tokens across 3,242 language model calls and 1,242 tool interactions.
The system autonomously identified and experimentally validated optical bilinear interaction — a previously unreported physical mechanism analogous to Transformer attention operations. This discovery suggests potential pathways for high-speed, energy-efficient optical computing hardware designed for pairwise mathematical operations.
Qiushi Engine reproduced published transmission-matrix experiments and converted abstract coherence-order theory into measurable experimental observables. The system maintained research trajectories across long-horizon investigations through its Meta-Trace memory system and dual-layer architecture.
DeepSeek V4 Advances Long-Context Processing
Chinese AI firm DeepSeek released V4, its flagship model optimized for Huawei’s Ascend chips rather than NVIDIA hardware. According to MIT Technology Review, V4 processes significantly longer prompts than previous generations through improved text handling efficiency.
The open-source model matches performance of leading closed-source systems from Anthropic, OpenAI, and Google while demonstrating China’s reduced dependence on NVIDIA semiconductors. V4’s optimization for Huawei chips represents a critical test of alternative AI hardware ecosystems.
DeepSeek’s continued open-source approach contrasts with the proprietary strategies of Western AI companies. The model’s competitive performance suggests that open development can match or exceed closed commercial systems.
OpenClaw Reaches 250,000 GitHub Stars
The open-source OpenClaw project surpassed 250,000 GitHub stars within 60 days, becoming the most-starred software project on the platform. According to NVIDIA’s analysis, the self-hosted AI assistant attracted over 2 million weekly visitors by March 2026.
Created by Peter Steinberger, OpenClaw enables users to deploy persistent AI assistants locally without cloud dependencies or external APIs. The project’s rapid adoption reflects growing demand for self-hosted AI systems that maintain data privacy and operational independence.
OpenClaw’s architecture supports unbounded autonomy, allowing AI agents to operate continuously without external service limitations. This capability appeals to organizations requiring persistent automation while maintaining control over their AI infrastructure.
What This Means
These developments signal a maturation of AI research beyond language models toward specialized scientific applications. MAMMAL’s multi-modal approach demonstrates that combining diverse data types can outperform single-domain models, even sophisticated ones like AlphaFold 3.
The autonomous discovery capabilities of Qiushi Engine represent a milestone for AI-driven scientific research. Systems that can independently formulate hypotheses, design experiments, and validate findings could accelerate discovery across multiple scientific disciplines.
DeepSeek V4’s performance on alternative hardware suggests the AI ecosystem is becoming less dependent on single suppliers. This diversification could reduce costs and increase access to high-performance AI systems globally.
FAQ
How does MAMMAL differ from AlphaFold 3?
MAMMAL processes multiple biological data types (proteins, molecules, genes) simultaneously, while AlphaFold 3 focuses primarily on protein structure prediction. This multi-modal approach enables MAMMAL to excel at interaction-based tasks like drug-target binding.
What makes autonomous scientific discovery significant?
The Qiushi Discovery Engine represents the first AI system to independently identify and experimentally validate a new physical mechanism. This demonstrates AI’s potential to accelerate scientific research by conducting experiments and making discoveries without human intervention.
Why does DeepSeek V4’s hardware optimization matter?
V4’s optimization for Huawei chips rather than NVIDIA hardware demonstrates that high-performance AI models can run on alternative semiconductors. This reduces dependence on single suppliers and could lower AI deployment costs globally.
