Are you ready to dive into the fascinating world of cutting-edge AI models designed especially for biology? We’ve gathered the crème de la crème of LLMs (Language Models) that are revolutionizing the way we explore and understand the biological realm. From unlocking the secrets of proteins to speeding up drug discovery, these top LLMs are here to rock the science scene.
LLM For Biology
Med-Flamingo-9B is a cutting-edge medical vision-language model that boasts remarkable multimodal in-context learning capabilities. Leveraging the power of the OpenFlamingo-9B V1 model, Med-Flamingo combines the prowess of the CLIP ViT-L/14 vision encoder and the Llama-7B language model, creating a fusion of state-of-the-art technologies.
Trained on paired and interleaved image-text data sourced from the vast expanse of medical literature, Med-Flamingo is uniquely equipped to revolutionize medical diagnostics. By analyzing medical reports and images, this model showcases the potential to go beyond our current expectations, offering the possibility of accurate and efficient medical issue identification. With its advanced capabilities, Med-Flamingo-9B promises to be a transformative tool in the field of healthcare, bridging the gap between AI and medicine.
HyenaDNA is a groundbreaking long-range genomic foundation model that revolutionizes the field of biology. With its impressive capacity for context lengths of up to 1 million tokens, it operates at single nucleotide resolution using the innovative Hyena operators. These operators, a remarkable substitute for traditional attention mechanisms, employ modified input projections, implicit convolutions, and gating to achieve unparalleled performance in language modeling.
Compared to previous genomic Transformer models with dense attention, HyenaDNA can reach context lengths up to 500 times longer and trains at a remarkable speed of 160 times faster at sequence length 1M (in comparison to Flash Attention). The model’s single character tokenizer with a primary vocab of 4 nucleotides allows for single nucleotide resolution, a significant advancement in genomic foundation models. Moreover, its implicit long convolution facilitates a global receptive field at each layer.
By pretraining on the human reference genome (HG38) and using next token (nucleotide) prediction, HyenaDNA achieves state-of-the-art results on 23 downstream tasks, including regulatory element prediction, chromatin profiles, and species classification. Additionally, it opens new possibilities in genomics, introducing in-context learning with soft prompt tuneable tokens and instruction fine-tuning, a groundbreaking advancement in the field.
MedAlpaca is a specialized language model tailored for the medical domain, featuring 7 billion parameters to effectively handle various tasks. What sets MedAlpaca apart is its relatively smaller size, making it ideal for deployment on smaller devices without compromising performance.
To fine-tune the model, diverse data sources were utilized. Anki flashcards were automatically transformed into question-answer pairs, while medical question-answer sets were generated from Wikidoc. Leveraging Chat-GPT 3.5, questions were extracted from headings, with the corresponding paragraphs serving as answers.
MedAlpaca aims to excel in medical question-answering and foster enhanced medical dialogues, proving to be a practical and efficient language model for specialized medical applications.
NVIDIA BioNeMo is a powerful cloud-based service specifically designed for generative AI in the field of drug discovery. Offering a unique edge, this non-open-source platform provides researchers with easy access to state-of-the-art generative and predictive biomolecular AI models on a large scale. By utilizing NVIDIA’s cloud APIs, users can swiftly tailor and deploy domain-specific AI models to efficiently generate the structures and functionalities of proteins and biomolecules. The significance of BioNeMo lies in its ability to expedite the process of developing new drug candidates, making it a valuable asset for researchers and developers in the biology domain.
Unlike other models in the list, BioNeMo is not open source, making it a proprietary yet effective solution for generating new drug candidates.