Biology-specific generative AI models

Biology-specific foundation models for pharmaceutical discovery

As of October 2026, the field includes several distinct groups of providers: companies building foundation models for proteins and other biological molecules, research institutes developing DNA and genome models, and AI platforms that combine biological models with drug-discovery workflows.

My recommendation is to evaluate providers by the biological problem you want to solve, rather than treating all “AI drug discovery” companies as equivalent. Protein design, target identification, molecular binding, gene regulation, and cell-state prediction require different capabilities.

1. Leading providers to evaluate

• 

1. EvolutionaryScale / Biohub

(Protein generation)

Models: ESM3 and ESM Cambrian.

ESM3 generates proteins using sequence, structure, and functional information; ESM Cambrian focuses on protein representations and predictive performance. A strong starting point for protein engineering, therapeutic proteins, and antibody-related research. ESM3 has API access, with a smaller open model also available.www.evolutionaryscale.ai+1

Website:evolutionaryscale.ai

---

• 

2. Arc Institute and NVIDIA

(DNA and genomes)

Models: Evo and Evo 2.

These models learn patterns across genomic sequences and support DNA sequence generation and biological-effect prediction. Consider them for genomics, genetic variants, gene-function research, and genome engineering—not as direct substitutes for small-molecule drug-design models. Evo 2 was released with open model resources and NVIDIA BioNeMo integration.arcinstitute.org+1

Website:Arc Institute — Evo 2

---

• 

3. Boltz

(Structure and binding)

Models: Boltz-2 and BoltzGen.

A relevant choice for biomolecular structure prediction, estimating small-molecule binding affinity, and designing biologics. Its collaboration with Pfizer announced in January 2026 also targets pharmaceutical deployment and models adapted to proprietary drug-discovery data.boltz.com

Website:boltz.com

---

• 

4. OpenAI

(Scientific reasoning)

Model family: GPT-Rosalind.

A specialized life-sciences reasoning model for research workflows spanning biology, chemistry, protein engineering, and genomics. It is worth evaluating for literature synthesis, experimental planning, and tool-assisted research, alongside—not instead of—molecular models. Eligible organizations can access the model series through a trusted-access program.openai.com

Website:GPT-Rosalind

2. More providers worth knowing

These organizations are also relevant, depending on how broadly you define a biology foundation model.

The last three are worth distinguishing from model-first providers: they are primarily drug-discovery companies or integrated AI-biotechnology platforms, rather than simply vendors of general-purpose foundation models. Their proprietary models and access arrangements may differ substantially.

3. Match the provider to your discovery program

Small-molecule discovery

Start with Boltz-2 and its newer molecular-design offerings for structure, binding and screening. Use AlphaFold 3 as a complementary structure-prediction option.

---

Genomics and target discovery

Evaluate Evo 2 and AlphaGenome for sequence-level biology, variant effects and regulatory mechanisms.

---

Antibodies and therapeutic proteins

Compare ESM3 with BoltzGen. Test generated candidates experimentally for binding, specificity, stability, expression and developability.

---

Cell biology and disease mechanisms

Consider Recursion, insitro and other experimental-data-driven platforms. Evaluate whether they can model your specific cell types, perturbations and disease phenotypes.

---

Research orchestration and hypothesis generation

Evaluate GPT-Rosalind alongside molecular models for literature review, experimental design and tool-assisted scientific workflows.

4. What to check before selecting a provider

For pharmaceutical R&D, model quality is only one part of the decision. I would assess each candidate against these criteria:

• Evidence: Prospective experimental validation, not just benchmark scores or retrospective results.
• Data fit: Performance on your target class, assay type, disease area and biological context.
• Access: Open weights versus API-only access, commercial licensing, fine-tuning and on-premises deployment.
• Data governance: Rights to use proprietary assay data, confidentiality, retention policies and ownership of generated IP.
• Workflow integration: Compatibility with your structure databases, screening pipelines, laboratory information systems and experimental feedback loops.
• Economics: Total cost per experimentally validated hit or lead, rather than cost per model inference.

One important caveat: a predicted structure, high affinity score or generated protein sequence is not evidence that a drug candidate will work in a biological system, much less in a clinical trial.

5. Let's narrow this to the right providers for you

The most useful next step is to distinguish the kind of foundation model you want to procure or build around.

What is your primary objective?

Small-molecule discovery

Antibodies and protein therapeutics

Genomics and target discovery

Cell biology and disease modeling

A broad, multimodal biology foundation model

---

How do you want to access the models?

Open weights and self-hosting

Commercial API or partnership

Build or fine-tune a proprietary model

Evaluate all options

---

Where are you in the discovery process?

Target identification

Hit identification and lead optimization

Preclinical development

---

What data can you bring?

Public datasets only

Proprietary experimental data

Both public and proprietary data

[Compare providers for my program]

The landscape of biology-specific foundation models spans several layers—from open-source sequence models and specialized biotech startups to enterprise cloud ecosystems and major tech infrastructure providers.

The primary providers of biology-specific foundation models are organized below by their core biological modality and ecosystem:

1. Proteins, Structures, and Molecular Design

• AlphaFold Lineage ( Google DeepMind ): AlphaFold 3 expands past protein folding to predict structures and interactions for proteins, DNA, RNA, chemical modifications, and ligands.
• ESM Models ( Meta AI* / Evolutionary Scale *): Meta’s ESM-2 and the newer ESM3 models are state-of-the-art generative models trained on massive protein sequence spaces to reason over structure, function, and evolutionary scaling.lifesciencesaihandbook.com
• Profluent ( Profluent Bio*)* : Creator of Evo and open-source generative protein models (like ProGen) used for de novo design of gene-editing tools and functional proteins.codeocean.com

2. Genomics and Transcriptomics (DNA, RNA, and Single-Cell)

• Evo & Evo 2 ( Arc Institute / Together AI ) : Genomic foundation models trained across kingdoms of life capable of predicting molecular consequences, designing DNA sequences, and understanding genomic grammar at single-nucleotide resolution.ardigen.com+1
• Nucleotide Transformer ( Google Research* / Hugging Face *): Large transformer models trained on multi-species genomes to predict chromatin effects, variant impacts, and gene expression.lifesciencesaihandbook.com+1
• Single-Cell Models (Geneformer, scGPT, scFoundation) : Academic and open-source ecosystems designed for single-cell transcriptomics, predicting cellular responses to gene perturbations, drug treatments, and disease states.research.ibm.com+2

3. Multimodal & Enterprise Biology Platforms

• BioNeMo ( NVIDIA ) : An enterprise framework and cloud service providing optimized biomolecular foundation models for small-molecule generation, docking, quantum chemistry, and protein property prediction.
• Bioptimus : A startup founded by alumni from Google DeepMind and Meta that builds comprehensive multimodal AI "world models" spanning cellular to tissue scales (e.g., H-Optimus and M-Optimus for pathology and genomics).
• IBM Research Biomedical Foundation Models : Open-source and enterprise models (BMFM ) targeting target discovery, biologics, and small-molecule generation using multi-omic representations.bioptimus.com+3

4. Cloud Infrastructure & Life Science Ecosystems

• AWS (Amazon Bio Discovery) : Offers integrated lab-in-the-loop workflows featuring dozens of pre-integrated biological foundation models alongside experimental design agents.
• Illumina (BioInsight AI Lab) : Focuses on purpose-built genomic models for variant effect prediction and multiomic sequence interpretation.illumina.com+1

If you'd like to narrow this down, let me know:

• What specific modality you are focusing on (small molecules, proteins, or genomics)?
• Are you looking for open-source weights to fine-tune locally or managed cloud APIs?

Leading biological foundation model providers for pharmaceutical discovery

As of October 2026, the market includes several distinct types of providers: companies that build models directly from biological data, AI biotech companies that use those models to design medicines, and infrastructure platforms that let pharmaceutical R&D teams train and deploy their own models.

The key distinction is what the model learns from and what it can do. Protein sequence models, DNA/RNA models, molecular structure models, and cellular foundation models address different parts of drug discovery. They aren't interchangeable.www.williamblair.com+2

1. Main providers to evaluate

Representative providers, grouped by their primary strengths rather than ranked by market share.

• 

EvolutionaryScale

(Protein foundation models)

Builds the ESM family of protein language models, including ESM3, which works across protein sequence, structure and function. A strong starting point for protein engineering, enzyme design and therapeutic protein research.

Access: open-weight research models and commercial offerings.www.williamblair.com+4

---

• 

Chai Discovery

(Biomolecular design)

Develops foundation models for biomolecular structure prediction and design, including the Chai model family. Particularly relevant to antibody and biologics discovery, molecular interactions and designing candidate binders.

Access: commercial platform and pharma partnerships.qai.io+6

---

• 

Boltz

(Structure and affinity)

Develops Boltz-1/2 and BoltzGen models for biomolecular structure prediction, molecular interaction and generative design. A particularly relevant option for teams exploring both small molecules and biologics.

Access: open-source models and commercial platform. Pfizer announced a strategic collaboration in January 2026.qai.io+6

---

• 

Basecamp Research

(Evolutionary biology)

Its EDEN foundation model is trained on evolutionary biological data and targets therapeutic design across biological modalities. Worth evaluating for gene editing, DNA/RNA engineering and protein-based therapeutics.

Access: model and therapeutic-design partnerships.qai.io+6

---

• 

BioMap

(Multimodal biology)

Develops the xTrimo family, spanning biological data such as DNA, RNA, proteins, cells and broader biological systems. Interesting for organizations seeking a more general-purpose biological foundation model rather than a single-protein-task solution.

Access: enterprise and collaborative development.www.williamblair.com+4

---

• 

NVIDIA — BioNeMo

(Model development platform)

Provides infrastructure, pretrained models, training tools and deployment capabilities for biomolecular AI. Relevant when a pharma company wants to adapt models to proprietary assay, protein, molecular or other research data.

Access: developer frameworks, model services and enterprise infrastructure.nvidianews.nvidia.com+1

These six are a useful initial shortlist, but they don't cover the entire market. The following providers are also important, especially if your goal is an integrated discovery platform rather than a standalone foundation model.

2. Other major players, depending on the discovery problem

Two further names worth tracking are Profluent for generative protein and gene-editor design, and Generate:Biomedicines for generative biologics discovery. Both are relevant if the objective is to create new therapeutic proteins rather than simply predict biological properties.www.bvp.com+2

3. Match the provider to your use case

Protein engineering and biologics

Start with EvolutionaryScale, Chai Discovery, Boltz, Basecamp Research and Profluent. Evaluate sequence generation, binding, developability and experimental validation.

---

Small-molecule drug discovery

Start with Boltz, NVIDIA BioNeMo, Insilico Medicine and Isomorphic Labs. Prioritize target structure, binding affinity, molecular generation, synthesis feasibility and ADMET.

---

Target discovery and cellular biology

Start with Recursion, Owkin and BioMap. Assess their ability to link molecular and cellular measurements to disease mechanisms, target validation and perturbation responses.

---

Proprietary pharma data and custom models

Start with NVIDIA BioNeMo and Apheris; consider AWS for managed model workflows. Focus on data rights, privacy, fine-tuning options, deployment and integration with existing R&D infrastructure.

4. What separates a foundation model from a drug-discovery platform?

This distinction matters when evaluating vendors.

• Biological foundation model: pretrained on large biological datasets, such as protein sequences, structures, genomic sequences or cellular images, and adaptable to multiple downstream tasks.
• Drug-discovery platform: combines models with proprietary data, chemistry tools, experimental workflows and sometimes an internal therapeutic pipeline.
• Scientific reasoning model: helps researchers interpret evidence, formulate hypotheses and plan experiments, but is not necessarily a model pretrained directly on raw biological data.

For example, EvolutionaryScale is a useful reference point for protein foundation models; Recursion for cellular data and experimental feedback; and NVIDIA BioNeMo for building and deploying biomolecular models. GPT-Rosalind is a scientific reasoning model rather than a direct substitute for a protein-generation model.www.williamblair.com+3

5. My recommended initial shortlist

If I were conducting a market scan for a pharmaceutical R&D organization, I would start with these eight:

1. EvolutionaryScale — protein foundation models.
2. Chai Discovery — biomolecular structure and design.
3. Boltz — structure prediction, affinity and molecular design.
4. Basecamp Research — evolutionary data and therapeutic design.
5. BioMap — multimodal biological foundation models.
6. Recursion — cellular phenomics and data-driven discovery.
7. NVIDIA BioNeMo — model training, customization and deployment.
8. OpenAI — scientific reasoning and integrated life-sciences research workflows.

This is a practical shortlist, not a definitive ranking. For each provider, verify commercial access, licensing, model availability, training-data provenance and prospective experimental performance.

One important caveat: a strong benchmark result does not establish that a model will produce better drug candidates in your program. Ask vendors to demonstrate improvements on your own target class, assays and experimental constraints.

To narrow this down further: are you primarily looking to license pretrained models, build proprietary models using your own pharma data, or partner with a provider to discover new drugs? The leading candidates differ considerably across those three objectives.

The landscape of biological and biomedical foundation models for pharmaceutical discovery has shifted rapidly from narrow, task-specific machine learning to general-purpose biological intelligence.pubmed.ncbi.nlm.nih.gov+1

The main providers and frontier AI labs building these foundational systems span techbio startups, Alphabet spin-offs, and major computational biology platforms:

• Isomorphic Labs – An Alphabet/Google DeepMind spin-off building on the breakthrough legacy of AlphaFold. They focus on creating an AI-first, end-to-end foundation model platform for small-molecule and complex molecular design, partnering directly with major pharma companies like Novartis and Eli Lilly.isomorphiclabs.com+4
• EvolutionaryScale – Emerging out of Meta’s FAIR (Fundamental AI Research) team, this frontier AI lab developed ESM3 , a powerful multimodal generative language model for protein biology that reasons simultaneously across the sequence, structure, and function of proteins.evolutionaryscale.ai+2
• Recursion Pharmaceuticals – A clinical-stage TechBio leader operating massive automated wet-lab infrastructure. Recursion combines petabytes of proprietary biological, chemical, and phenomic image data with open-source/collaborative model development (such as co-developing advanced predictive tools like the Boltz model family with MIT).techlifesci.com
• Generate:Biomedicines – Specializes in de novo protein design and generative biology. Their platform integrates proprietary machine learning with high-throughput wet-lab validation to generate entirely new clinical-stage biologic drugs and therapeutics on demand.
• Insilico Medicine – An end-to-end, generative AI-driven pharma technology company utilizing proprietary platforms like PandaOmics and Chemistry4Netix. They build multimodal foundation models trained on extensive biological, multi-omic, and clinical data to discover novel targets and small-molecule candidates.
• Chai Discovery – A prominent AI biology startup focusing on foundational models for molecular interaction and structure prediction, noted for high-profile early-access licensing agreements with major pharma players like Pfizer.genengnews.com
• Genesis Molecular AI – Known for its GEMS (Genesis Exploration of Molecular Space) platform, which uses advanced deep learning for protein-ligand structure and property prediction, scaling via massive data-sharing partnerships with top-tier pharmaceutical companies.genengnews.com

To help narrow down which provider aligns with your pipeline, let me know:

• Are you focusing on small molecules or biologics/proteins?
• Are you looking for an off-the-shelf/API solution , an open-source model , or a custom co-development partnership with a pharma-focused TechBio?