Where Rust, Go, and TS actually land in the 2026 biotech stack
Verified from GitHub + job postings · June 2026
The bio stack has two layers: a C++/CUDA simulation core that nobody is replacing, and a Python everything-else layer that dominates orchestration, ML, and lab tooling. The entry points for systems engineers are at the edges — fast I/O, trajectory analysis, and a genuine gap in LLM-friendly simulation tools.
Production engines are C++/CUDA. Analysis and tooling is Python. Rust is emerging at the I/O and trajectory analysis layer. Go is rare.
| Layer | Dominant language | Rust/Go presence | Status |
|---|---|---|---|
| MD simulation core | C++/CUDA (GROMACS, AMBER, OpenMM) | None in production | C++ entrenched |
| I/O parsing (BAM/VCF/PDB) | C/C++ historically | noodles, alevin-fry, rust-bio — production-ready | Rust viable |
| Trajectory analysis | Python (MDAnalysis, MDTraj) | molar — peer-reviewed, faster than Python | Rust challenger |
| ML inference (folding) | Python/PyTorch (AF3, ESM3) | candle (Rust ML framework — no protein models yet) | Rust aspirational |
| Workflow orchestration | Python (Nextflow, Snakemake) | None | Python wins |
| Lab automation | Python (Opentrons v2, PyLabRobot) | None | Python wins |
| Services / APIs | Python (FastAPI) | Go listed optionally (Recursion, Benchling infra) | Python default |
GPU backends (2026): GROMACS 2026.0 ships CUDA (primary), SYCL (AMD/Intel portable), and HIP (AMD native). AMBER 24 adds AMD ROCm/HIP for MI100–MI300A. OpenMM 8.2.0 adds HIP. All three major engines now support AMD GPU — not just NVIDIA.
Ranked by realistic near-term impact. Stars and activity verified June 2026.
| Project | Lang | Stars | Why | Entry effort |
|---|---|---|---|---|
| pdbtbx PDB/mmCIF structure parser |
Rust | 70 | 17 open issues with specific gaps: secondary structure (#139), bond parsing (#137), mmCIF helix (#140). Solo maintainer, welcomes help. PDB is the universal protein structure format. | Low–medium |
| alevin-fry scRNA-seq preprocessing |
Rust | 207 | 24 open issues. Academic maintainers (Rob Patro group) receptive to external contributors. Nature Methods 2022. Beat salmon (C++) on preprocessing. 98.4% Rust. | Medium |
| molar (MolAR) MD trajectory analysis |
Rust | 52 | Peer-reviewed (PMC11609497, J. Computational Chemistry Dec 2024). Benchmarked faster than MDAnalysis, MDTraj, and CPPTRAJ. Active June 2026. Low competition for contributions. | Medium |
| noodles BAM/CRAM/VCF/FASTQ/GFF I/O |
Rust | 709 | Most production-ready Rust bio library. In use at St. Jude. Solo expert maintainer — high bar. Requires reading format specs (BAM 1.6, CRAM 3.1). High credibility payoff per merged PR. | High |
| ferritin Protein structure utilities |
Rust | 33 | Early-stage, active May 2026. Open issues: dependency slim-down (#110), MolViewSpec renderer (#18), LigandMPNN tests (#58). First-mover advantage. Risk: may stall. | Low–medium |
| SeqKit FASTA/FASTQ toolkit |
Go | 1,600 | Most-starred Go bio tool. 13 open issues. Published iMeta 2024 (10-year anniversary). Solo maintainer. Go contributor in bio is rare — differentiating signal. | Medium |
| OpenFold-3 Open AlphaFold3 reproduction |
Python | 750 | Apache-2.0, active June 2026, 5 open PRs. Best open contribution target in protein folding — AF3 itself only accepts bug fixes. Python-only, no Rust angle, but high scientific visibility. | Medium–high |
Avoid near-term: AlphaFold3 (Google CLA, CC BY-NC-SA non-commercial, bug fixes only), lumol (stalled Feb 2024), biogo (abandoned), candle protein port (multi-month investment as a first project).
Current bio simulation tools are mostly Jupyter/GUI tools designed for human scientists — slow to start, complex to invoke programmatically, not designed for looping. roadrunner (C++/Python, ODE/SBML) is the fastest existing option but has a dated API nobody designed for AI agents. (Full landscape sweep below.)
What "uv for bio sim" looks like:
With: SBML input (the standard format — interop with everything), JSON/CSV output (LLMs can parse it), Python bindings via PyO3 (LLM agents using Python call it natively), sub-100ms cold start.
Why it doesn't exist yet: Scientists don't need 10,000 simulations/sec. LLMs do. The use case is new.
A sweep of the open-source landscape — none of these hits clean JSON + sub-100ms cold start.
| Tool | Lang | Stars | Status | LLM-loop fitness |
|---|---|---|---|---|
| roadrunner | C++ + Python | 62 | Active (3 releases in 2026) | Best existing. SBML in, numpy out. 200–400ms startup (LLVM JIT). Not JSON-native, not sub-100ms. The engine behind Tellurium. |
| NFsim | C++ | 15 | Active (Jun 2026) | Fast compiled binary. Recently added .nfevent.json output — the only tool moving toward JSON. Rule-based stochastic. |
| BioNetGen | C++/Perl | 68 | Active (Jun 2026) | CLI scriptable, but tab-separated .gdat output (not JSON). 82 open issues. |
| COPASI / basico | C++ (SWIG) | 128 | Active (Jan 2026) | Most feature-complete. CopasiSE CLI is headless and SBML-aware — decent for orchestration. Heavy, dated API. |
| Smoldyn | C++ | 30 | Slowing (Jan 2025) | Spatial stochastic, pip-installable, simple text model format. |
| Tellurium | Python | 142 | Slowing (no release since Dec 2024) | Wraps roadrunner. ~1–2s import overhead (matplotlib, libsbml). 130 open issues. |
| GillesPy2 | Python | 84 | Abandoned (features stalled 2021) | Returns UserDict/UserList, not JSON. Don't build on this. |
| StochSS | JS/Python | 25 | Stalled (GUI-first) | Requires JupyterHub deployment. Not a library. |
Read of the field: roadrunner is the reference implementation worth studying — fast C++ core, SBML-native. NFsim's JSON output is the signal that someone else sees the gap. But nothing is Rust-native, JSON-first, and sub-100ms. The space is open.
Career fit: Demonstrates Rust + bio intersection, is benchmarkable (vs roadrunner/COPASI), publishable as a short Methods paper or biorxiv preprint, and has a clear user (any lab doing AI-assisted circuit design). Also the natural foundation for a game simulation core.
The bio simulation stack needs what uv gave Python: a Rust-native tool fast enough to run in an LLM loop without Python overhead.
A cell is a program. But asking whether a cell "understands itself" is like asking if a neuron has a model — wrong level. The useful question is: what are the thinking tools that let us reason about cells, the way language and logic let us reason with neurons?
| Level | Model | What it captures | Speed |
|---|---|---|---|
| 1 | Boolean networks | Gene ON/OFF logic, regulatory circuits, cell fate attractors (Kauffman NK model) | Laptop, milliseconds |
| 2 | Gillespie / ODE | Stochastic gene expression noise, protein concentration dynamics, repressilator circuits | Laptop, seconds |
| 3 | Flux balance analysis | Metabolic flow — linear programming over reaction constraints. What can the cell produce? | Laptop, seconds |
| 4 | Coarse-grained MD (MARTINI) | Membrane dynamics, lipid bilayers, large protein complexes — atoms grouped into beads | GPU, hours |
| 5 | All-atom MD | Every atom explicit. Nanosecond timescales. GROMACS/AMBER/OpenMM. | GPU cluster, days |
| 6 | Whole-cell model | Karr 2012 (Cell): complete Mycoplasma genitalium, 525 genes, 128 researchers. Still the state of the art. | HPC cluster, hours |
The LLM-friendly zone is levels 1–3. Boolean networks and Gillespie simulation are fast enough to run in an agent loop, rich enough to capture real biology, and the tools are Python/Java GUI — not ergonomic for programmatic use.
None of these experimental tools benefit from Rust. Their speed limits are wet-lab biology, not compute.
Verified from public job postings and GitHub org analysis, June 2026.
| Company | Stack (verified) | Rust/Go? |
|---|---|---|
| Recursion | Python primary. gflownet (289 stars) is most active OSS project. Go in 3 infra forks (logrus-stack, go-testutils), last active 2020. | Go optional in full-stack JDs alongside Python/Java/Ruby |
| Benchling | Python/Flask core, TypeScript frontend | Go optional in infra JDs; no Rust |
| Ginkgo Bioworks | Python + C# (Driver Developer JD). One internal Rust tool: gen (sequence VCS, 9 stars, last active Apr 2025) | Rust exists but not in hiring signals |
| Generate:Biomedicines | Python, ML Ops | No Rust/Go in JDs |
| Unnamed top-10 pharma | HN "Who is Hiring" May 2026: Rust SDK + TypeScript + workflow DSL. Hiring 4–5 Rust engineers. Company not named. | Rust explicitly required — only confirmed case |
Bottom line: Zero named companies list Rust in job postings as of June 2026. The most likely employers for Rust/C++ simulation engineering — D.E. Shaw Research, Schrödinger, Relay Therapeutics — were not surveyed but are the correct targets if simulation performance is the goal.
The HFT → bio transfer works, but the bottleneck shifts: from network/syscall latency to GPU memory bandwidth and Python interop. The mental model adapts.
For a Rust/Go/TS senior engineer with no prior bio background, one merged PR in a peer-reviewed or production bio project by day 90 is achievable.