Foundation models for biological systems—highlighted in Y Combinator’s Summer 2024 Request for Startups—remain a significant opportunity for founders building at the intersection of AI, life sciences, and healthcare. The request was not simply for another chatbot layered onto medical data. It pointed toward models that could learn the structure and behaviour of biological systems, then support research, diagnostics, therapeutics, and laboratory workflows.
The Summer 2024 programme is closed, but its thesis remains relevant in 2026. Biology generates rich, multimodal data, while many high-value workflows are still slow, fragmented, and expensive. Startups that combine strong biological insight with defensible data, rigorous evaluation, and a clear route to deployment can still build compelling companies—particularly in India, where research institutions, hospitals, pharmaceutical manufacturers, and clinical networks offer varied sources of real-world problems.
What are foundation models for biological systems?
A foundation model is a broadly trained model that can be adapted to multiple downstream tasks. In biology, the training material may include:
- Genomic sequences, variants, and annotations
- Protein sequences and structures
- Single-cell and spatial transcriptomics
- Molecular graphs, compounds, and assay results
- Microscopy, pathology, and radiology images
- Electronic health records and clinical notes
- Laboratory protocols, papers, and scientific knowledge graphs
Unlike a general language model, a biological foundation model must represent domain-specific constraints. A useful system should understand that a molecule’s chemical structure affects binding, that cell states vary by tissue and context, and that clinical observations are shaped by incomplete data, treatment history, and measurement bias.
The strongest products will often be multimodal rather than purely text-based. A drug discovery model may combine molecular structure, protein sequence, assay data, and scientific literature. A pathology system may link tissue images to clinical history and molecular findings. The model is only one part of the product: data governance, laboratory integration, validation, and expert review are equally important.
Why the opportunity matters
Drug discovery and protein engineering
Models can help researchers prioritise compounds, predict molecular properties, design proteins, and identify experiments that reduce uncertainty. The commercial value comes from improving a measurable workflow—such as hit discovery, lead optimisation, or assay selection—not from producing plausible scientific language.
Genomics and precision medicine
Models can interpret variants, analyse gene regulation, and identify patterns across cohorts. In India, this could support rare-disease research, population-specific genomics, and more affordable clinical interpretation. However, models trained mostly on European datasets may perform poorly on Indian populations, making local data and careful validation a genuine advantage.
Biomedical imaging
Large models can assist with image segmentation, triage, quality control, and research discovery across pathology, radiology, microscopy, and ultrasound. Teams working in this area should study how to build computer vision models on GitHub with reproducible datasets, versioned experiments, and transparent benchmarks.
Laboratory automation
A model becomes substantially more valuable when connected to instruments, laboratory information systems, and experiment planning. It can recommend the next experiment, flag anomalous results, or translate protocols into machine-executable steps. This is a systems problem: founders may need the principles used in building distributed systems with AI agents, adapted to safety-critical laboratory environments.
What YC’s request means for founders
Y Combinator’s RFS was a directional signal, not a grant scheme or a guarantee of funding. It encouraged founders to pursue ambitious problems where advances in AI could unlock new biological capabilities. A strong application or company thesis would need more than access to a large dataset and a generic transformer.
Founders should be able to answer four questions clearly:
1. Which biological decision or workflow is being improved?
2. Why is a foundation model necessary rather than a smaller specialised model?
3. What proprietary data, feedback loop, or experimental capability creates defensibility?
4. How will performance be measured in the laboratory or clinic?
The best answers connect model performance to outcomes: fewer failed experiments, faster review, better sensitivity at a fixed specificity, lower interpretation cost, or improved patient recruitment. “Accuracy” alone is rarely sufficient.
A practical build and validation plan
Start with a narrow, high-value workflow
Do not begin by attempting to model all of biology. Choose one problem with a defined user, data source, and decision point. Examples include prioritising variants for expert review, predicting assay outcomes, or extracting structured information from pathology reports.
Establish data rights and quality
Biological data is difficult to collect, standardise, and legally reuse. Document consent, provenance, licensing, de-identification, and permitted commercial use. Measure missingness, label consistency, batch effects, and population coverage before training. In India, partnerships with hospitals and laboratories must address ethics approval, data-sharing agreements, and local operational constraints from the outset.
Build a credible baseline
Compare the foundation model with simple statistical methods, domain heuristics, existing open models, and expert workflows. Use patient-level or study-level splits to prevent leakage. For longitudinal clinical data, ensure that information available after the prediction time is not accidentally included in training.
Evaluate scientific usefulness
Use external validation, prospective testing, calibration, subgroup analysis, and uncertainty estimates. For discovery systems, measure whether model-guided experiments outperform random or conventional selection. For clinical tools, evaluate workflow impact and safety, not just retrospective scores. A model that knows when it is uncertain is more useful than one that confidently fabricates an answer.
Design for deployment
Plan early for audit logs, access controls, model monitoring, human review, and rollback. Sensitive health and genomic information requires strong security and clear governance. Where infrastructure or privacy makes cloud deployment difficult, teams can consider deploying large language models locally, while recognising that local inference does not remove the need for data controls.
India-specific opportunities and constraints
India offers a large and diverse healthcare market, strong pharmaceutical and contract research capabilities, and growing public investment in digital infrastructure. Potential wedges include affordable clinical decision support, Indian-population genomics, multilingual research interfaces, laboratory workflow software, and tools for pharmaceutical process development.
But market size is not a substitute for evidence. Hospitals may have inconsistent data formats, limited interoperability, and long procurement cycles. Clinical validation can be slow, and regulatory expectations vary by intended use. Founders should secure a design partner, define a paid or measurable pilot, and identify the regulatory classification before making clinical claims.
Language and accessibility also matter. A research platform serving Indian clinicians or field workers may need multilingual interfaces and speech capabilities; teams can learn from work on open-source vision-language models for Indian languages. The underlying biological model still needs rigorous English-language scientific evaluation, but the product experience should match its users.
Common mistakes to avoid
- Training a large model before proving a valuable workflow
- Treating synthetic data as a replacement for real biological labels
- Reporting random-split performance that hides leakage
- Ignoring population shift and underrepresented Indian cohorts
- Presenting generated hypotheses as experimentally validated findings
- Building a research demo without a laboratory, hospital, or pharmaceutical buyer
- Assuming an accelerator application replaces regulatory and ethics planning
What a strong 2026 startup looks like
A credible company in this category usually combines three assets: specialised data, deep biological or clinical expertise, and a workflow that generates proprietary feedback. The model should improve as users run experiments, review cases, or generate validated outcomes. This creates a compounding advantage that is harder to copy than model architecture alone.
For founders revisiting the YC RFS thesis in 2026, the practical goal is not to recreate a 2024 application. It is to show that a narrowly defined biological problem can be solved better, faster, or more affordably with modern foundation-model techniques—and that the result can survive expert scrutiny, real-world validation, and India’s operational complexity.