19 Companies Pioneering AI Foundation Models in Pharma and Biotech
Foundation models represent a new paradigm in artificial intelligence (AI), revolutionizing how machine learning models are developed and deployed. As these models grow increasingly capable, they become useful for applications across a wide range of economic functions and industries, including biotech. Foundation models are a class of large-scale machine learning models, typically based on deep learning architectures such as transformers, that are trained on massive datasets encompassing diverse types of data. The most prominent examples of general-purpose foundation models are the GPT-3 and GPT-4 models, which form the basis of ChatGPT, and BERT, or Bidirectional Encoder Representations from Transformers. These are gigantic models trained on enormous volumes of data, often in a self-supervised or unsupervised manner (without the need for labeled data).
Their scalability in terms of both model size and data volume enables them to capture intricate patterns and dependencies within the data. The pre-training phase of foundation models imparts them with a broad knowledge base, making them highly efficient in few-shot or zero-shot learning scenarios where minimal labeled data is available for specific tasks.
This approach demonstrates their high versatility and transfer learning capabilities, adapting to the nuances of particular challenges through additional training.
Below we summarized a number of companies building domain-specific foundation models for biology research and related areas, like chemistry.
Atomic AI
Atomic AI, a biotech company focused on AI-driven RNA drug discovery, aims for atomic precision in their work. Their proprietary platform, PARSE (Platform for AI-driven RNA Structure Exploration), is based on a machine learning model trained on a limited set of RNA molecules.
This model makes accurate predictions about the structure of various RNA molecules, enhancing RNA structure prediction. Atomic AI utilizes their foundational model internally for their drug discovery program, enabling them to pursue novel targets in RNA that were previously inaccessible. This approach aligns with the pharmaceutical industry's growing interest in novel biology, facilitating new avenues in drug discovery.
BioMap
BioMap focuses on unveiling nature's rules and generating diverse proteins with high accuracy. Their primary foundation model, xTrimo (Cross-Modal Transformer Representation of Interactome and Multi-Omics), is designed to understand and predict life's behavior at various complexity levels. xTrimo is trained on extensive datasets, including over 6 billion proteins and 100 billion protein-protein interactions, making it the largest life science AI foundation model with over 100 billion parameters.
This model's scale allows it to inform multiple downstream task models even with minimal data. BioMap’s strategic collaboration with Sanofi, announced in 2023, involves co-developing AI modules for biotherapeutic drug discovery, leveraging BioMap’s AI expertise and Sanofi’s proprietary data to create advanced AI models for biologics design and optimization.
Bioptimus
In February 2024, Bioptimus, a biotech startup based in France, announced the successful closure of a $35 million seed funding round to develop an AI foundation model targeting advancements across the biological spectrum, from molecular to organismal levels.
Led by Professor Jean-Philippe Vert, the company collaborates with Owkin to leverage extensive data generation capabilities and multimodal patient data from leading academic hospitals worldwide. Owkin's initiative, MOSAIC, represents one of the largest multi-omics atlases for cancer research, showcasing the potential of combining computational and experimental research methods.
This collaboration, supported by Amazon Web Services (AWS), is crucial for developing AI models capable of capturing the diversity of biological data.
Chai Discovery
Chai Discovery, a six-month-old AI biology startup based in San Francisco, just announced the release of its first open-source model, Chai-1. The model is designed to predict the structure of biochemical molecules, a key capability in drug discovery.
The company, founded by former OpenAI and Meta researchers, recently raised nearly $30 million in a seed funding round led by Thrive Capital and OpenAI, valuing the company at $150 million. Chai Discovery is focused on using AI foundation models to transform biology from a science into an engineering discipline, with a particular emphasis on predicting and reprogramming molecular interactions.
Chai-1 is an advanced AI model that predicts the structures of various biochemical entities, such as proteins, small molecules, DNA, RNA, and even complex chemical modifications. What sets Chai-1 apart from other tools, like Google DeepMind’s AlphaFold, is its ability to achieve higher accuracy in predicting these structures, with improvements of 10% to 20% in success rates on key tasks related to drug discovery.
Chait-1 benchmark test results
For example, Chai-1 has shown a 77% success rate on the PoseBusters benchmark, which is a test that measures how well the model can predict how proteins and other molecules fit together—a crucial step in designing new drugs. It also scored 0.849 on the CASP15 protein monomer structure prediction set, which means it’s very good at accurately predicting the shape of single proteins, outperforming other top models.
The key thing about Chai-1 is that it doesn’t rely on a method called multiple sequence alignments (MSAs), which most traditional models use to find patterns in sequences of proteins or other molecules. MSAs require lots of data and computational power, which can be a bottleneck. Instead, Chai-1 can work with just a single sequence of a molecule and still make highly accurate predictions. This makes it much more versatile and efficient, especially in situations where data is scarce or incomplete—common challenges in real-world drug discovery.
In simpler terms, Chai-1 can take a simpler input and still deliver top-notch results, making it a powerful tool for researchers aiming to speed up the process of finding new medications.
The model’s practical applications are vast, as it enhances drug discovery processes by providing precise predictions of molecular structures and interactions. Chai-1 is particularly effective at predicting protein-ligand interactions and folding multimers, achieving higher accuracy than MSA-based models.
Web version available online
By making Chai-1 free and open-source, Chai Discovery encourages further research and development in the field, allowing other developers to build on their work for non-commercial purposes.
ChemCrow
ChemCrow is an AI-driven chemistry agent designed to tackle tasks in organic synthesis, drug discovery, and materials design. By integrating 13 expert-designed tools, ChemCrow enhances the performance of large language models (LLMs) in chemistry.
The platform operates by prompting an LLM, such as GPT-4, with specific instructions related to a given chemical task, guiding the model to follow a structured process of "Thought, Action, Action Input, Observation." This method allows ChemCrow to reason about the task's current state, consider its relevance to the final goal, and plan subsequent steps. ChemCrow assists expert chemists by providing advanced chemical knowledge and simplifies complex chemical problems for non-experts, broadening access to AI-driven chemistry solutions.
Cyrus Biotechnology
Cyrus Biotechnology, based in Seattle, Washington, integrates advanced AI models into their protein biochemistry processes. Utilizing foundation models like AlphaFold2, RoseTTAFold, and OpenFold, Cyrus has significantly advanced protein structure prediction, outperforming previous models and traditional systems.
An integral part of their approach is the use of Rosetta, a model developed at the University of Washington, which employs statistics and atomic-scale physics to model proteins and other molecules. In collaboration with other entities, Cyrus has also initiated the OpenFold project, aiming to create a trainable foundation model for proteins. OpenFold surpasses AlphaFold2 in terms of speed and memory usage on commercially available hardware.
Deep Genomics
In September 2023, Deep Genomics unveiled BigRNA, a pioneering AI foundation model for uncovering RNA biology and therapeutics. It is the first transformer neural network engineered specifically for transcriptomics. BigRNA is informed by nearly two billion adjustable parameters and trained on thousands of datasets, totaling over a trillion genomic signals.
This model is designed to predict tissue-specific regulatory mechanisms of RNA expression, binding sites of proteins and microRNAs, and the effects of genetic variants and therapeutic candidates. By understanding these complex RNA interactions, BigRNA facilitates the discovery of new biological mechanisms and RNA therapeutic candidates that traditional approaches might miss, exemplifying its transformative potential in RNA-based drug discovery.
Enveda Biosciences
In May 2024, Enveda Biosciences unveiled PRISM (Pretrained Representations Informed by Spectral Masking), a foundation model trained on 1.2 billion small molecule mass spectra, aiming to enhance molecular structure identification. PRISM employs self-supervised learning on a large dataset of unannotated spectra, using a masked peak modeling approach similar to masked language modeling in NLP.
The model improves the prediction of chemical properties and spectral matching tasks, aiding drug hunters in identifying new medicines from natural molecules. Enveda plans to expand the PRISM dataset to further enhance the model’s predictive capabilities, supporting the discovery of novel therapeutics.
Evo
Developed by researchers from the Arc Institute, Stanford, and TogetherAI, Evo is a biological foundation model operating across DNA, RNA, and proteins. With over 7 billion parameters, Evo facilitates predictive tasks and the generative design of biological sequences. Its architecture, derived from the StripedHyena design, enables the model to handle long genomic sequences, improving efficiency and the quality of DNA sequence generation.
This novel architecture enables the model to work at an unprecedented single-nucleotide resolution, making it adept at handling the extensive lengths of prokaryotic and phage genomic sequences.
The development team has made Evo accessible through GitHub and the Together playground, inviting the scientific community to explore its capabilities firsthand.
Ginkgo Bioworks
In August 2023, Ginkgo Bioworks and Google Cloud announced a 5-year partnership aimed at developing state-of-the-art large language models (LLMs) focused on genomics, protein function, and synthetic biology. Ginkgo’s AI foundation model will run on Google Cloud's Vertex AI platform, aiming to accelerate innovation in drug discovery, agriculture, industrial manufacturing, and biosecurity.
Furthermore, in February 2024 Ginkgo committed to building next-generation biological foundation models by acquiring key assets of Reverie Labs, a startup specializing in AI/ML tools for drug discovery. This acquisition includes Reverie's infrastructure and software for training large-scale AI models, enhancing Ginkgo's capabilities in developing comprehensive biological models.
Helical
Helical, founded in 2023 and based in Luxembourg, raised €2.2 million in seed funding in June 2024 to build the first open-source platform dedicated to bio foundation models for DNA and RNA data. Led by co-founders Rick Schneider, Mathieu Klop, and Maxime Allard, the company aims to democratize access to advanced AI tools for pharmaceutical and biotech companies, helping them integrate these models into drug discovery processes without the need for specialized AI teams.
Helical focuses on creating a user-friendly interface for biologists and data scientists, allowing them to leverage complex genomic models through simple API calls. The platform includes a library of Bio AI Agents—pre-built applications tailored for tasks such as biomarker discovery and target prediction.
Unlike other platforms, Helical specifically integrates DNA and RNA foundation models, offering researchers the ability to work directly with nucleotide data rather than general AI inputs like text or images.
Helical’s model library and tools are open-source, facilitating collaboration and continuous improvement from the scientific community. This approach enables users to not only utilize existing models but also benchmark and refine their own, supporting a wide range of applications from basic research to early-stage drug development.
Helical’s partnerships with organizations such as LuxProvide, NVIDIA, and Microsoft further enhance its capabilities by providing access to high-performance computing resources, advanced genomic analysis platforms, and essential startup tools, ensuring robust support for large-scale data processing and model training.
HyenaDNA
Developed by Stanford University scientists, HyenaDNA is a genomic foundation model designed to understand the distribution of DNA sequences, the encoding of genes, and the regulation of gene expression.
The model is designed to understand the distribution of DNA sequences, the encoding of genes, and the regulation of gene expression by the sequences located between those coding for amino acids, aiming to enhance the understanding of gene regulation and its implications for health and disease.
ImmunoPrecise Antibodies Ltd. (BioStrand)
BioStrand, a subsidiary of ImmunoPrecise Antibodies Ltd. based in the Netherlands, introduced a new AI foundation model integrating Large Language Models (LLMs) with their patented HYFT Technology. This technology discerns and utilizes universal fingerprint patterns found across biological entities, linking sequence data with structural, functional, and bibliographical information across a comprehensive knowledge graph.
This integration facilitates the decoding of complex protein languages, advancing drug discovery, protein-based therapeutics, and synthetic biology. BioStrand’s approach promises to accelerate the development of targeted treatments and revolutionize protein engineering and design.
InSilico Medicine
InSilico Medicine, a pioneer in generative AI for drug discovery founded in 2014, recently launched Precious3GPT—a cutting-edge AI model aimed at aging research and drug discovery.
This model is designed to handle a wide range of biological data, integrating information from different species: rats, monkeys, and humans, and various data types such as transcriptomics, proteomics, and methylation. With that, the model allows researchers to run virtual experiments and predict how different compounds might affect aging across tissues and species.
Model architecture and training
Precious3GPT enables the simulation of aging and drug effects across different species using simple prompts, facilitating the identification of compounds that could be effective in both mice and humans. The model is accessible to the research community via platforms like Hugging Face, positioning it as a collaborative tool for advancing aging and disease research.
Technically, Precious3GPT is a transformer-based model that utilizes a novel tokenization logic, allowing it to process over 2 million data points from public omics datasets, biomedical text, and knowledge graphs. The model is trained to handle queries in natural language, making it highly adaptable for research purposes. For example, it can predict biological age in specific tissues, simulate the effects of drugs on different species, and even generate synthetic biomedical data for experimental simulations. By using aligned multimodal data from various tumor regions and species, the model decodes this into high-dimensional, high-resolution images that capture therapeutically relevant features of tumor and immune biology.
A standout feature of Precious3GPT is its use of structured multimodal masking during training, where most of the data is hidden, forcing the model to learn deep connections across different biological modalities. This method allows it to predict complex interactions, like how specific genes might influence protein levels in various tissues, as if asking “What if we change this gene?” and having the model predict the ripple effects throughout the body.
InSilico developed Precious3GPT in collaboration with researchers at Harvard and other institutions, aiming to create an open, community-driven resource. The model’s ability to handle complex, multi-species data sets it apart, offering a new approach to exploring aging and drug discovery that’s both scalable and precise.
Noetik
Noetik, founded in 2022 by Jacob Rinaldi and Ron Alfa in the San Francisco Bay Area, is using AI to tackle one of the toughest challenges in cancer treatment: finding the right targets and understanding how drugs will work on different patients. They just secured $40 million in Series A funding to expand their work, including growing one of the world’s largest cancer biology datasets and enhancing their in vivo CRISPR Perturb-Map platform, which helps them test and refine potential therapies.
At the heart of Noetik’s tech is OCTO, a powerful AI model that acts like a virtual lab for cancer research. While many AI models focus on predicting molecular structures, OCTO goes further—it predicts how different cancer treatments might play out in real patients. This model can be thought of as a simulator that can test “what if” scenarios, helping scientists see which treatments could work best for which patients, without the long trial-and-error of traditional methods.
OCTO is trained on a large mix of data from thousands of tumor samples, including gene expression, protein data, and images of the cancer cells. By learning from these varied inputs, OCTO can predict how tweaking a single gene could change protein levels across a tumor.
Noetik’s approach could cut down on the guesswork in cancer therapy, making it quicker and cheaper to find treatments that really work. They’re also forming partnerships with top research institutions and companies to push their innovations into the clinic, aiming to make a difference for patients.
Paige
Paige, a provider of digital pathology solutions, has launched a multi-modal AI model in pathology and oncology. Their foundation model, also called PRISM (A Multi-Modal Generative Foundation Model for Slide-Level Histopathology), enhances reporting and generative capabilities for cancer detection, biomarker identification, cellular subtyping, spatial biology, and therapy response prediction. PRISM is built on Virchow tile embeddings, aggregating these into a single slide embedding for image perception tasks.
PRISM reduces the time and resources needed for developing advanced AI systems, contributing to advancements in precision oncology and improving cancer diagnosis and treatment.
Piramidal
Piramidal, founded in 2024 by Dimitris Sakellariou and Kris Pahuja, is developing a foundational AI model specifically for analyzing brainwave data from electroencephalography (EEG). EEGs are widely used in hospitals to monitor brain activity, but interpreting these signals can be complex and varies between different machines and setups. Piramidal aims to simplify and standardize this process with a model that can consistently detect critical brainwave patterns, regardless of the equipment used or the specific patient characteristics. This could improve monitoring for conditions like seizures or strokes, especially in high-stakes environments like neural ICUs, where continuous observation is crucial but often stretched thin by staff limitations.
Piramidal’s approach is to treat EEG signals like a language of the brain, similar to how large language models like ChatGPT handle human text. By training on a massive collection of EEG data from various sources, the model learns to recognize patterns in brain activity that might indicate medical concerns. This is akin to a highly trained assistant that can spot subtle signs of brain distress, which might otherwise go unnoticed by even experienced medical professionals, especially when the data comes from different EEG machines with varying configurations. This uniform approach not only saves time but also reduces the chance of human error in diagnosing serious conditions.
The company's model is designed to work out-of-the-box with any EEG setup, unlike traditional models that require retraining for each machine or scenario. It acts like a universal translator for brainwaves, simplifying implementation across different hospital settings without customization. By using diverse, harmonized data from various open-source sources, the model starts with a strong foundation, making it more effective from the outset compared to models that begin from scratch.
They are currently scaling the model to billions of parameters, making the model more detailed and capable of understanding even finer nuances in brainwave data. The first production version of this model is set to be tested in hospitals through four pilot programs starting in early 2025, focusing on ICU settings. These pilots will evaluate whether the model can reliably interpret EEGs across different real-world conditions, acting, once again, as an additional layer of monitoring to support medical staff.
Piramidal raised $6 million in seed funding from investors like Adverb Ventures, Lionheart Ventures, and Y Combinator to support computing costs and team expansion.
Recursion
Recursion, a biopharma company known for integrating AI and massive datasets into drug discovery, recently released Phenom-Beta, the first in a series of foundational models available on NVIDIA’s BioNeMo platform. This model is part of a broader initiative to harness phenomics—the study of cell phenotypes in response to various chemical and genetic perturbations. Phenom-Beta is designed to process cellular microscopy images into general-purpose embeddings, making it a flexible tool for analyzing and understanding underlying biological systems.
Phenom-Beta uses a vision transformer (ViT) architecture, a type of neural network initially developed for image recognition tasks that can also handle various complex data inputs. It employs self-supervised learning through masked autoencoders (MAEs), which involves hiding about 75% of the pixels in an image and training the model to fill in the gaps. This approach allows the model to learn patterns and relationships without needing labeled data, making it more adaptable and robust compared to traditional supervised learning methods that rely on extensive labeling.
The model was trained using the RxRx3 dataset, a publicly available collection of approximately 2.2 million images of HUVEC cells—a type of human cell often used in research—featuring ~17,000 genetic knockouts (where specific genes are intentionally disrupted) and 1,674 chemical entities (various compounds tested for their biological effects). Despite being trained on Cell Painting—a specialized imaging assay that uses multiple fluorescent dyes to highlight different cellular components—Phenom-Beta can generalize to other types of microscopy, such as brightfield imaging, which uses simple light to visualize cells and is typically lower in detail.
Phenom-Beta, training and inference (source)
One of Phenom-Beta’s key capabilities is its ability to extract biologically meaningful features from these images, capturing subtle changes in cell structure that might be missed by the human eye. By converting these images into high-dimensional embeddings, which are numerical representations of the data, the model can map out complex relationships between different genetic and chemical interventions. For example, it can help researchers see how knocking out a particular gene might affect cell function, or how different drugs might produce similar or distinct effects on cells.
Phenom-Beta is particularly versatile, able to work with both brightfield and more complex fluorescent imaging techniques. It can also perform in-silico (computer-based) cellular organelle fluorescent staining on brightfield images, predicting high-contrast fluorescent images from simpler, less detailed brightfield inputs. This makes it a valuable tool for scaling imaging techniques in drug discovery, as brightfield microscopy is less costly and more accessible than specialized fluorescent imaging, yet Phenom-Beta can still extract a similar depth of biological information.
Researchers can access Phenom-Beta on the NVIDIA BioNeMo platform via a Cloud API, allowing them to leverage its capabilities at a supercomputing scale. This availability broadens the model's impact by providing a powerful tool for the scientific community to explore cellular phenotypes and drug effects without the need for extensive computational resources. Recursion continues to develop more advanced versions of Phenom, like Phenom-1, for its internal teams and select partners, but Phenom-Beta represents a significant step in making these tools available for broader research purposes.
Terray Therapeutics
Terray Therapeutics integrates large-scale experimentation with generative AI to improve small molecule drug discovery. Their platform combines ultra-high throughput experimentation, generative AI, biology, medicinal chemistry, automation, and nanotechnology.
In November 2023, Terray announced a collaboration with NVIDIA to harness NVIDIA's technologies for training foundation models for chemistry. Utilizing NVIDIA DGX Cloud, Terray aims to develop the world’s most comprehensive chemistry foundation models for small molecules. This collaboration enhances their ability to explore broad molecular spaces, solving complex problems in drug discovery and advancing their preclinical pipeline.
Topic: AI in Bio
