I am preparing to use generative AI for pharmaceutical discovery. Who provides biology-specific foundation models?

Yes. The field has moved well beyond generic LLMs: there are now several providers building biology-specific foundation models, with different modalities and commercial models.

Leading providers to know

A few distinctions are important.

EvolutionaryScale is probably the clearest example of a “biology foundation-model company.” ESM3 is a generative model designed to understand and generate proteins, while ESM Cambrian focuses on protein representation learning. EvolutionaryScale explicitly makes ESM3 available through its platform and has commercial partnerships around the models.

NVIDIA BioNeMo is different: it is more of a model/platform ecosystem than a single biology model. It provides pretrained biomolecular models, training infrastructure, APIs and deployment tooling spanning protein structure, protein and small-molecule generation, docking and property prediction.

Arc's Evo 2 is particularly interesting if genomics matters. It is a 40B-parameter genomic foundation model trained on nearly 9 trillion nucleotides, intended to model and generate biological sequences from the molecular to genome scale.

Boltz is especially worth watching for an open-model strategy. Boltz-2 and BoltzGen are positioned as open biomolecular foundation models covering structure prediction, affinity estimation and molecular/biologics design; Pfizer announced a 2026 collaboration to develop models using Pfizer's proprietary data.

And OpenAI has now entered the category directly: GPT-Rosalind was introduced in April 2026 as a purpose-built life-sciences model spanning biology, drug discovery and translational medicine, with capabilities across chemistry, protein engineering and genomics.

How I'd segment the market for pharma

If you're evaluating vendors rather than simply making a landscape, I'd divide them into four buckets:

1. Protein foundation models — EvolutionaryScale, Generate:Biomedicines, Chai, Boltz Best for biologics, antibodies, protein engineering and protein-function prediction.
2. Genomic/DNA models — Arc/Evo, Basecamp Best for sequence-function relationships, variant interpretation and genetic engineering.
3. Multimodal molecular models — Boltz, NVIDIA BioNeMo, Isomorphic Labs Best if you need to connect protein structure + ligand + affinity + molecular design.
4. Scientific reasoning/agentic models — OpenAI and increasingly general frontier-model providers Best for literature synthesis, hypothesis generation, experimental planning and orchestrating specialized biology models rather than replacing them.

For a pharmaceutical discovery program, I would not choose based on parameter count or the phrase "foundation model." The important questions are whether the model has been trained on the right biological modality, whether you can fine-tune it on proprietary data, whether commercial rights are clear, whether inference can be deployed in your environment, and—most importantly—whether it improves experimentally validated endpoints.

If you tell me whether your focus is small molecules, antibodies/biologics, RNA, cell biology, genomics, or end-to-end drug discovery, I can narrow this to the 5–10 providers I'd actually evaluate, including model access, licensing, proprietary-data fine-tuning, and pharma partnerships.

NVIDIA BioNeMo is different: it is more of a model/platform ecosystem than a single biology model. It provides pretrained biomolecular models, training infrastructure, APIs and deployment tooling spanning protein structure, protein and small-molecule generation, docking and property prediction.

Arc's Evo 2 is particularly interesting if genomics matters. It is a 40B-parameter genomic foundation model trained on nearly 9 trillion nucleotides, intended to model and generate biological sequences from the molecular to genome scale.

Boltz is especially worth watching for an open-model strategy. Boltz-2 and BoltzGen are positioned as open biomolecular foundation models covering structure prediction, affinity estimation and molecular/biologics design; Pfizer announced a 2026 collaboration to develop models using Pfizer's proprietary data.

The ecosystem for biology-specific foundation models spans specialized AI drug-discovery companies, tech-giant spin-offs, and cloud infrastructure platforms. Providers generally fall into a few distinct categories depending on whether you are looking for structural biology, multi-omics, target identification, or general biomedical reasoning.

1. Specialized AI & Structural Biology Leaders

• Isomorphic Labs : Spun out of Google DeepMind, they are building state-of-the-art AI models (evolving from the AlphaFold architecture) to simulate biological systems and decode the physics and chemistry of molecular interactions. They partner heavily with major pharma companies like Eli Lilly and Novartis.
• Insilico Medicine : Operates an end-to-end generative AI platform (Pharma.AI and Chemistry42) utilizing proprietary biological and chemical foundation models to handle target discovery, small-molecule design, and clinical trial prediction.
• Generate:Biomedicines : Focuses on "programming biology" to generate novel protein therapeutics and antibodies, utilizing massive-scale machine learning models trained on structural and sequence data across natural and synthetic proteins.
• Recursion Pharmaceuticals : Combines massive automated wet-lab data generation with deep learning foundation models to map and decode cellular biology at scale, partnering widely across the pharmaceutical industry.

2. Tech & Cloud Platform Providers

• Microsoft (Azure AI Foundry) : Hosts a curated catalog of biomedical and life sciences models—including BioGPT (pre-trained on PubMed abstracts) and BiomedCLIP (a vision-language foundation model)—alongside structural and imaging AI infrastructure.
• NVIDIA (BioNeMo Framework) : Provides an enterprise-grade framework specifically for developing, customizing, and executing generative AI foundation models for drug discovery, spanning protein language models, molecular generation, and docking.

3. Multi-Omics & Specialized Genomics Startups

• Bioptimus : Built by former-Hugging Face and scaling talent, this startup focuses on creating universal AI foundation models natively across biological scales—integrating pathology, genomics, and spatial transcriptomics from cells to organs.
• Insitro : Leverages machine learning and high-throughput biology to generate proprietary biological data, building predictive models to establish disease models and therapeutic targets.

To help narrow down which provider or framework fits your project, let me know:

• Are you focusing on small molecules, large biologics/proteins, or genomics/transcriptomics?
• Do you plan to build/fine-tune models locally or utilize a fully managed cloud/API infrastructure?

For pharmaceutical discovery, “biology-specific foundation models” usually refers to models pretrained on biological data (protein sequences, structures, genomes, transcriptomes, single-cell data, molecular interactions, literature, etc.) and adapted for tasks such as target discovery, protein engineering, biomarker discovery, and molecule design. The market is split between model builders, platform providers, and AI-native biotech companies that use proprietary models internally.

Here are the major providers to evaluate:

Model categories you should distinguish

1. Protein foundation models

Examples:

• EvolutionaryScale / ESM lineage
• Meta protein models
• NVIDIA BioNeMo protein models
• Boltz models

Best for:

• protein engineering
• antibody design
• enzyme engineering
• variant effect prediction

2. Structure and interaction models

Examples:

• AlphaFold family
• Boltz
• related 3D biomolecular models

Best for:

• target validation
• protein–ligand interaction hypotheses
• structure-guided design

3. Cell and genomics foundation models

Examples:

• single-cell transformers
• perturbation models
• gene-expression models

Best for:

• target discovery
• biomarker discovery
• disease mechanism modeling

4. Multimodal drug discovery platforms

These combine:

• omics
• imaging
• chemistry
• literature
• clinical data

Examples include Recursion, Insitro, and newer virtual-lab platforms.

How pharma teams typically evaluate vendors

For a discovery organization, I would compare providers on:

1. Biological modality

• protein only?
• DNA/RNA?
• single-cell?
• chemistry + biology?
• clinical data?

2. Customization

• Can you fine-tune on proprietary assays?
• Can models incorporate internal compound libraries or experimental results?

3. Closed-loop capability

• Does the system connect predictions → experiments → new training data?

4. Deployment

• API only?
• cloud platform?
• private deployment?
• on-prem requirements?

5. Validation

• Published benchmarks?
• Prospective wet-lab validation?
• Pharma collaborations?

A practical shortlist for a pharmaceutical discovery group would usually include:

• Protein design: EvolutionaryScale, NVIDIA BioNeMo, Boltz, Generate Biomedicines
• Structure-based discovery: Google DeepMind/Isomorphic Labs, Schrödinger, Boltz
• Multimodal target discovery: Recursion, Insitro, Helical
• Scientific AI assistants/workflow automation: OpenAI and other frontier AI providers

The “best” provider depends heavily on whether your priority is target discovery, biologics, small molecules, biomarker discovery, or building an internal AI discovery platform.

Entering the space of generative AI for pharmaceutical discovery means navigating a rapidly expanding ecosystem of biology-specific foundation models. Unlike general Large Language Models (LLMs) trained on internet text, these models are trained on atomic structures, genomic sequences, high-content cellular images, and evolutionary data.

The primary providers, tech bio pioneers, and platforms offering biology-specific foundation models span several categories:

1. Frontier Generative Biology & Protein Labs

• EvolutionaryScale: Emerging from Meta's AI research, they developed ESM3 , a frontier generative model for biology trained on over 2 billion protein sequences . ESM3 can reason across sequence, structure, and function, effectively simulating hundreds of millions of years of evolutionary design to program new proteins.
• Isomorphic Labs: A spin-off from Google DeepMind , Isomorphic builds AI-driven drug design engines (leveraging breakthroughs like AlphaFold 3 ) aimed at modeling biological mechanisms and molecular interactions at an experimental level of accuracy.
• Generate:Biomedicines: Pioneers a "generative biology" platform that uses machine learning to program novel protein therapeutics, antibodies, and enzymes from scratch based on universal biological principles.

2. Phenomics & Multi-Scale Biology Platforms

• Recursion Pharmaceuticals: Developed the Phenom family of foundation models (hosted on NVIDIA BioNeMo). Recursion specializes in phenomics —turning massive libraries of high-content cellular microscopy images into embeddings that map complex chemical and genetic perturbations.
• Bioptimus: A European startup building multi-scale foundation models for human biology . Their models (like M-Optimus ) integrate data across modalities—ranging from tissue histology and spatial transcriptomics to genetics—to forecast clinical and molecular outcomes.

3. Enterprise Infrastructure & Cloud Ecosystems

• NVIDIA BioNeMo: Rather than being a pure-play biotech, NVIDIA provides a domain-specific cloud platform and microservice ecosystem for computer-aided drug discovery. BioNeMo hosts and accelerates pre-trained biological foundation models (including tools for protein folding, generative chemistry, and models like Recursion's Phenom ) optimized for pharma R&D infrastructure.
• Amgen: Major pharma companies are also contributing open-source biological foundation models, such as Amgen's AMPLIFY suite of protein language models designed to aid biologic drug discovery.

If you want to narrow this down, let me know:

• Are you focusing on small molecules, large-molecule biologics, or cell/gene therapies?
• Do you plan to deploy via API/cloud infrastructure (like NVIDIA BioNeMo) or license an end-to-end platform?

I can provide a more tailored architectural recommendation.

The provider landscape for biology-specific foundation models spans major tech conglomerates, well-funded pure-play AI biotech startups, and open-source scientific initiatives. Choosing a provider typically depends on whether you need pre-trained model weights for proteins/genomics, enterprise cloud infrastructure, or an end-to-end drug discovery partnership.

Key providers of biology-specific foundation models are organized by their primary ecosystem focus:

• AlphaFold / Protein Structure Specialists
  • Isomorphic Labs (Spin-off of Google DeepMind) focuses on applying AlphaFold architecture and novel generative physics-based models to real-world drug design. They hold major multi-target partnerships with pharma giants like Eli Lilly, Novartis, and Johnson & Johnson.
  • Chai Discovery provides high-performance foundation models for molecular structure and interaction prediction (such as Chai-1), offering early access and licensing deals to major pharmaceutical partners.
  • EvolutionaryScale develops frontier protein language models like ESM-3, trained on evolutionary sequences, structures, and functions to enable programmable biology and de novo protein design.
• Infrastructure and Enterprise Bio-Compute Platforms
  • NVIDIA provides the NVIDIA BioNeMo platform, a generative AI cloud service and framework for drug discovery. BioNeMo hosts pre-trained foundation models for protein structure prediction, molecular generation, and docking (integrating models like AlphaFold, ESMFold, and proprietary partner models) optimized for enterprise deployment.
• Cellular Phenomics and Multi-Omics Specialists
  • Recursion Pharmaceuticals operates industrial-scale automated biology datacenters generating massive phenotypic and image-based data. They train large multimodal models (such as their Phenom series) to map cellular changes and connect high-dimensional phenotypes to genetic targets.
  • Insitro builds machine learning and foundation models tightly coupled with proprietary wet-lab data generation, focusing on disease modeling, functional genomics, and patient stratification for complex central nervous system and metabolic diseases.
  • Bioptimus builds comprehensive "world models of biology" by training cross-scale multimodal foundation models spanning data from single cells up to tissues.
• Generative Biologics and Protein Design
  • Generate:Biomedicines utilizes generative foundational pipelines (such as their Chroma platform) to directly design novel therapeutic proteins, antibodies, and enzymes from scratch rather than screening historical compound libraries.
  • Absci combines wet-lab data generation with generative AI foundation models explicitly optimized for downstream antibody discovery, validation, and affinity optimization.

To help narrow down which provider or platform fits your strategy, could you share:

• Are you targeting small molecules, large protein therapeutics, or genomic/RNA medicines?
• Do you prefer an on-premises/private infrastructure deployment or a cloud API / SaaS model?

For pharmaceutical discovery, “biology-specific foundation models” are usually grouped by what biological data they learn from: proteins, molecules, genomes/transcriptomes, cells, or multimodal biology. The leading providers include a mix of AI companies, academic labs, cloud platforms, and biotech companies.

Academic/open models worth evaluating

For an internal pharma AI stack, many teams evaluate open models alongside commercial platforms:

• Protein sequence models

  • ESM family (Meta/EvolutionaryScale lineage): protein embeddings, function prediction, design.
  • ProtTrans and related protein language models.

• Structure models

  • AlphaFold 2/3
  • RoseTTAFold
  • OpenFold
  • ESMFold
  • Chai-1 and related multimodal structure predictors

• Protein design

  • RFdiffusion
  • ProteinMPNN
  • ESM-3-style generative protein models

• Small molecule generation/design

  • Diffusion-based molecular models
  • Transformer molecular generators
  • AI docking models such as DiffDock

How pharma teams typically choose

A practical evaluation framework:

1. Target biology

   • Need target discovery? Look at genome, transcriptome, single-cell, and knowledge-graph models.
   • Need biologics? Prioritize protein foundation models.
   • Need small molecules? Combine molecular generators, docking, and physics-based simulation.

2. Deployment model

   • Open-source: maximum control, internal fine-tuning, easier governance.
   • API/cloud: faster deployment and less infrastructure burden.
   • Partnership/licensing: useful when models are tied to proprietary biological datasets.

3. Data advantage The model alone is rarely the differentiator. Leading pharma programs increasingly combine foundation models with proprietary assay data, clinical data, omics data, and experimental feedback loops.

For a pharmaceutical discovery organization starting today, a common shortlist would be:

• Protein/biologics: AlphaFold 3 + ESM family + RFdiffusion/ProteinMPNN
• Small molecules: NVIDIA BioNeMo ecosystem + Schrödinger/chemistry AI approaches + docking models
• End-to-end AI drug discovery: Isomorphic Labs, Recursion, Insilico, Generate, and BenevolentAI-style platforms

I would also distinguish foundation model providers from AI drug discovery companies: the former sell or release reusable biological models; the latter often combine models with proprietary data, automation, and wet-lab validation.

Yes. The market has moved well beyond generic LLMs: there are now several companies building biology-native foundation models, although they differ substantially in what they model and whether you can actually license/use the models versus access them through a discovery platform.

The main providers I would evaluate

A few deserve particular attention:

EvolutionaryScale is probably the clearest example of a true biology foundation-model company. Its ESM3 model jointly models protein sequence, structure and function, while ESM Cambrian focuses on large-scale protein representation learning. ESM3 is explicitly positioned as a generative biology model.

NVIDIA BioNeMo is different: it is more of a foundation-model development and deployment ecosystem than a single biology model. It provides pretrained biomolecular models and tooling for protein structure prediction, molecular generation, protein–ligand modeling, docking and property prediction, as well as infrastructure for training your own models. www.nvidia.com This makes it particularly interesting if you're a pharma organization that wants to combine public foundation models with proprietary experimental data.

Isomorphic Labs is pursuing a more vertically integrated approach. Its Drug Design Engine goes beyond structure prediction toward an end-to-end computational drug-design system. In February 2026, the company reported substantially improved protein–ligand prediction and binding-affinity performance relative to AlphaFold 3 on its reported benchmarks.

Basecamp Research is particularly interesting if you're thinking about training data as a strategic asset. Its EDEN models are trained on an unusually large evolutionary dataset and are designed for predictive and generative biological tasks, including therapeutic design.

Owkin is taking a different route: its OwkinZero is a biology-specific reasoning model trained using hundreds of thousands of verified biomedical Q&A examples and proprietary multimodal patient data. Its emphasis is less "generate a new protein" and more reason over disease biology, multiomics and clinical data.

I'd divide the landscape into 4 buckets

1. Protein foundation models

   • EvolutionaryScale / ESM
   • Profluent
   • Generate:Biomedicines
   • Basecamp / EDEN

2. Structure & molecular-interaction models

   • Isomorphic Labs
   • Chai Discovery
   • NVIDIA/BioNeMo ecosystem
   • Boltz/OpenFold

3. Cellular, genomic & disease-biology models

   • Recursion
   • Owkin
   • various academic/industry genomic foundation models

4. Foundation-model infrastructure for pharma

   • NVIDIA BioNeMo is the standout here.
   • It lets a pharma company bring proprietary data, fine-tune models and deploy them, rather than simply consuming somebody else's black-box model.

One important distinction

If you're preparing an actual pharma AI strategy, I would not make the vendor shortlist solely on who has the biggest biological model.

The key question is what biological representation you need:

DNA/genome → RNA → protein → structure → interaction → cell phenotype → tissue → patient → clinical outcome

Different foundation models operate at different points on that chain. A protein model such as ESM3 can be exceptional for antibody/protein engineering but isn't necessarily the right foundation for target identification or patient stratification. Conversely, a multimodal model trained on pathology, transcriptomics and clinical data may be much more valuable for target discovery but useless for de novo protein generation.

Also, some of the most interesting models are not commercially accessible as standalone foundation models. For example, Isomorphic Labs is primarily pursuing its own drug-design engine, whereas BioNeMo is explicitly designed to let organizations build and deploy their own biomolecular AI workflows.

If your goal is selecting vendors for a pharmaceutical discovery program, I can next give you a more useful 2026 vendor landscape comparing ~15 providers by modality (small molecule, antibody, protein, RNA, genomics, cell biology), model accessibility, licensing, proprietary-data fine-tuning, and evidence of pharma partnerships.

The provider landscape for biology-specific foundation models (FMs) spans specialized AI-native biotech companies, major tech conglomerates building life science infrastructure, and open-science consortia.

Providers are generally categorized by their primary biological modality—such as proteins, genomics, or multi-omic cell simulations:

• Protein and Antibody Design Models
  • EvolutionaryScale: Creator of the ESM (Evolutionary Scale Modeling) family of models (including ESM-3), focused on predicting and generating multi-modal protein sequences, structures, and functions at a billion-parameter scale.
  • Generate Biomedicines: Develops generative platforms like Chroma to design entirely new proteins, antibodies, and peptide therapeutics from scratch rather than screening existing libraries.
  • Isomorphic Labs: Alphabet’s spinout leveraging AlphaFold technology and advanced generative architectures, partnering deeply with major pharma companies (such as Novartis, Eli Lilly, and Johnson & Johnson).
  • Absci: Specializes in integrated generative AI and wet-lab validation for de novo antibody discovery.
  • Chai Discovery & Boltz: Emerging structural biology platforms providing open-and-licensed rapid biomolecular interaction and folding models.
• Genomic and Transcriptomic Models
  • Arc Institute / Stanford (Evo series): Creators of the Evo genomic language models (Evo 1 and Evo 2), capable of modeling prokaryotic and phage genomes and generating functional biological sequences at scale.
  • Deep Genomics: Uses its proprietary Biological Foundation Model (BioFM) platform to decode RNA biology and discover genetic medicine targets.
  • Insilico Medicine: Powers its Pharma.AI and Chemistry42 engines with deep generative and biological age models that have advanced candidates into clinical trials.
  • Nucleotide Transformer / Geneformer: Open-source and institutional genomic foundation models trained on massive multi-species DNA and single-cell transcriptomic datasets.
• Systems Biology, Phenomics, and Infrastructure Providers
  • NVIDIA (BioNeMo): Provides a foundational cloud-and-hardware framework hosting third-party and custom life-science models (such as AlphaFold, ESM, and Recursion's Phenom) optimized for enterprise pharma R&D.
  • Recursion Pharmaceuticals: Operates massive automated wet-lab infrastructure to train high-dimensional phenomic foundation models on millions of cellular images and genetic perturbations.
  • Bioptimus: Builds broad multi-scale biological foundation models (like M-Optimus) that span scales from single cells and spatial transcriptomics to full tissues.
  • Amazon Web Services (AWS): Offers integrated life science ecosystems like Amazon Bio Discovery, which bundles access to a wide array of curated biological foundation models with lab-in-the-loop AI agents.

To help narrow down which provider fits your pipeline, could you share:

• Are you focusing on small molecules, antibodies/proteins , or genetic/RNA medicines?
• Do you prefer an open-source/on-premise model stack or an API-driven / commercial partnership model?

The landscape of biology-specific foundation models for drug discovery is driven by specialized techbio companies, major AI labs, and hardware-software ecosystems. These organizations provide pre-trained models spanning genomics, transcriptomics, protein folding, and molecular design.

The primary providers of biology-specific foundation models are categorized below:

• Specialized TechBio and AI Research Labs
  1. Isomorphic Labs – Born from Google DeepMind, they build frontier foundation models focused on simulating the molecular "laws of biology" to predict how molecules and proteins interact in biological contexts.
  2. EvolutionaryScale – Developed the ESM3 protein language model family, designed to reason over sequence, structure, and function to generate entirely new proteins and optimize antibodies.
  3. Generate:Biomedicines – Uses its proprietary Generate Platform powered by large-scale biological intelligence to systematically engineer novel therapeutic proteins and antibodies [1.0].
  4. Insilico Medicine – Offers an extensive portfolio of multimodal foundation models (such as their multi-parameter pre-trained frameworks) spanning biology, chemistry, and target discovery via their Pharma.AI platform.
  5. Bioptimus – Focuses on building universal biological foundation models (like M-Optimus ) that integrate data across scales, from single cells and spatial transcriptomics to full tissues.
  6. Arc Institute / Open-Source Genomes – Developed Evo (and Evo 2 ), massive open-source genomic foundation models trained on millions of genomes (prokaryotic and eukaryotic) capable of generative DNA design and variant pathogenicity prediction.
• Enterprise Cloud and Infrastructure Providers

7. NVIDIA – Through the BioNeMo framework and cloud service, NVIDIA provides pre-trained biomolecular foundation models for protein structure prediction, small molecule generation, and molecular dynamics, heavily integrated into pharma workflows.
8. Amazon Web Services (AWS) – Offers Amazon Bio Discovery , a lab-in-the-loop ecosystem that bundles dozens of open-source and proprietary biological foundation models with deployment agents for automated experimental design.

If you tell me which specific modality you want to focus on (e.g., small molecule design, protein/antibody generation, or genomics), I can recommend which specific model or platform matches your pipeline goals.

Yes. The market has moved well beyond generic LLMs: there are now biology-specific foundation models spanning proteins, DNA/genomics, cells, pathology, molecular structures, and drug–target interactions.

I’d divide the landscape into a few important groups:

A few are especially worth investigating.

1. Protein foundation models

For protein engineering and protein therapeutics, EvolutionaryScale is one of the clearest specialist providers. Its ESM family treats protein sequences/structures as the biological analogue of language, enabling generation and prediction of protein properties.

For protein structure and protein–ligand interactions, the AlphaFold lineage is still foundational. Google DeepMind's AlphaFold 3 is designed to model interactions involving proteins, nucleic acids, small molecules and other biomolecular components. Isomorphic Labs is applying this technology specifically to drug discovery.

There is also a rapidly developing open-model ecosystem. Boltz-2, for example, is a structural-biology foundation model that adds binding-affinity prediction to structure prediction; its authors report dramatically lower computational cost than traditional free-energy perturbation approaches.

2. Genomic foundation models

For target discovery, regulatory genomics and variant interpretation, Google DeepMind's AlphaGenome is particularly notable. It can process DNA sequences up to 1 million base pairs and predict gene expression, splicing, chromatin features and other regulatory signals. Google currently provides API access for non-commercial research.

The Evo/Evo 2 family from the Arc Institute ecosystem takes a somewhat different approach: large DNA-language models trained directly on genomic sequences. These are potentially interesting for sequence design as well as genomic interpretation.

3. Cell and single-cell foundation models

For pharma, this category may ultimately be just as important as protein models because many discovery questions are really about cell state and response to perturbation.

Geneformer, for example, is a transformer pretrained on large-scale single-cell transcriptomes and intended to support context-aware predictions in network biology.

Models in this category can be used for things such as:

• predicting cell-state changes after drug perturbation
• identifying disease-associated cell populations
• inferring gene regulatory relationships
• prioritizing targets
• predicting responders/nonresponders
• modeling combination therapies

There is still considerably more uncertainty around how well these models generalize to genuinely novel perturbations than the marketing sometimes suggests.

4. Pathology / tissue foundation models

If your discovery program has substantial histopathology or spatial biology, don't overlook this category.

Bioptimus is particularly interesting. Its H-Optimus-1 is a 1.1B-parameter pathology foundation model trained on more than 1 million histology slides from more than 800,000 patients. The company is now extending this toward M-Optimus, intended to integrate pathology, spatial transcriptomics and genomics.

This is much closer to a drug-development biology model than simply an image classifier: potential applications include biomarker discovery, patient stratification and predicting drug response.

5. Platforms rather than a single model

NVIDIA's BioNeMo deserves separate consideration because it isn't just one foundation model. It is an ecosystem for training, fine-tuning and deploying biomolecular models, including models for protein structure, protein generation, small-molecule generation, property prediction and molecular docking.

That makes BioNeMo particularly relevant if your question is:

"We have proprietary pharma data and want to build/customize biology foundation models ourselves."

rather than:

"Which pretrained model gives us the best biological prediction?"

My practical shortlist

If I were evaluating vendors/models for a pharmaceutical discovery program today, I'd start with:

Protein/structure

• EvolutionaryScale / ESM
• Google DeepMind / AlphaFold
• Boltz
• NVIDIA BioNeMo

Genomics

• Google DeepMind / AlphaGenome
• Evo 2 / Arc ecosystem

Cells / perturbations

• Geneformer
• scGPT and newer single-cell foundation models
• Recursion's biological modeling platform

Tissue / translational biology

• Bioptimus
• Owkin
• other pathology foundation-model providers

End-to-end drug discovery

• Recursion
• Insilico Medicine
• Isomorphic Labs

The important distinction is that "biology foundation model" is not one market. A protein language model, a single-cell model, a pathology model and a protein–ligand structure model solve fundamentally different problems. For pharma, I would choose the model family based on the discovery bottleneck you're trying to address rather than looking for one universal biology LLM.

If you tell me whether your priority is target identification, small-molecule discovery, biologics/protein design, biomarker discovery, or patient stratification, I can narrow this to the 5–10 providers/models I'd actually evaluate, including access model, licensing, proprietary-data fine-tuning, compute requirements, and evidence of pharma adoption.