GPU-Powered AI Tools on NextELN: A Self-Hosted Inference Tier for the Wet Lab

Author: NextELN Lab Assistant Affiliation: Excellgen, Inc. Date: July 15, 2026 Correspondence: phs.phd@gmail.com

---

Abstract

The most transformative bioinformatics methods of the last five years — protein structure prediction, protein-language-model epitope prediction, docking, and generative molecular design — share one property: their core is GPU model inference. In a screen of 2,485 published bioinformatics tools, GPU-dependent methods were a minority (~13%) but concentrated sharply in exactly the domains that matter most for reagent and therapeutic design. Here we describe how NextELN, an AI-native electronic lab notebook, delivers these GPU tools to bench scientists through a self-hosted inference tier (the bio-infer service) and an operator-brokered service-order workflow, so that a scientist uploads a sequence and receives a folded structure, an epitope map, or a vaccine construct — without a GPU, a cluster account, or a bioinformatics core. We give the model catalog, the service architecture, the graceful-degradation design that keeps the notebook working when GPUs are absent, and the economics that make a single 4-GPU server sufficient.

---

1. Introduction

Deep learning has redrawn the frontier of computational biology. AlphaFold2 and the ESM family solved single-sequence structure prediction; protein-language models now drive epitope prediction and protein engineering; diffusion and flow models generate binders and dock ligands. What these share, and what separates them from the long tail of classical bioinformatics, is a dependence on GPU inference over large learned weights. That dependence is the barrier: the science is published and open, but running it requires hardware, environments, and databases most wet labs do not maintain.

A companion analysis of 2,485 tool papers across 20 domains showed GPU need is concentrated, not diffuse — dominating machine/deep-learning methods (94 of 118 sampled), protein structure and docking (40 of 125), and clustering in multi-omics, cheminformatics, and immunoinformatics, while ~87% of bioinformatics remains CPU work. That concentration is what makes a shared GPU tier practical: a small fleet serves the few domains that need it, for an entire lab or institution.

2. The GPU tool catalog

NextELN's GPU tier targets the highest-value, best-validated model families:

• Protein structure prediction. ESMFold folds a single sequence in seconds without a multiple-sequence alignment; ColabFold/AlphaFold2 add MSA-based accuracy for hard targets. Both fit on one 16–24 GB GPU (MMseqs2 removes the multi-terabyte database requirement). These power the notebook's 3D structure viewer and the batch structure-prediction service.
• B-cell epitope prediction. BepiPred-3.0 uses ESM-2 embeddings to predict linear and conformational B-cell epitopes — the model behind the antibody-antigen and vaccine tools.
• T-cell epitope prediction. NetMHCpan and NetMHCIIpan predict peptide–MHC class I and II binding across HLA alleles; these drive the vaccine designer's CTL/HTL epitope selection and population coverage.
• Frontier design (roadmap). RFdiffusion and ProteinMPNN for de-novo binder and enzyme design, DiffDock-class models for docking and affinity, and single-cell/multi-omics foundation models for representation learning — each a natural addition to the same tier.

3. Architecture: the bio-infer service

The GPU models run in a dedicated FastAPI inference service deployed inside the lab network, not in the web application. It exposes narrow, typed endpoints — /esmfold, /bepipred, /netmhcpan, /netmhciipan, /health — each of which lazily loads its backend and returns HTTP 503 if that tool is not installed, so the service starts even with a partial install and grows as models are added. All endpoints are protected by a bearer token; requests are validated and size-capped.

The NextELN application calls this service through a thin client (bio-infer.ts) configured by a single environment variable (ELN_BIO_INFER_URL). Two design choices make the tier robust:

1. Graceful degradation. Every model-backed feature falls back to a transparent, clearly-labeled CPU heuristic when the service is unreachable. The 3D viewer still renders deposited structures; the antibody tool still ranks peptides by classical propensity scales; the vaccine designer still assembles a construct from heuristic epitopes. The notebook never breaks because a GPU is offline — it simply gets more accurate when one is connected.
2. Air-gap friendliness. Because the GPU box is typically isolated from the internet, model weights are pre-staged offline and the service is reached over a private link (WireGuard/VPN or a reverse tunnel), never the public internet.

4. The service-order workflow

For jobs rather than interactive calls, GPU work is delivered as operator-brokered service orders. A scientist selects a service (e.g. mRNA vaccine candidate design), uploads an input, sees the price and any discount, signs a research-use agreement, and submits. The input is stored in managed storage; the operator is notified, runs the job on the GPU box, and uploads the result, which the scientist downloads into the notebook. This workflow exists precisely because the GPU box is currently air-gapped: it lets the service run today, manually, with full provenance and a signed agreement, and it becomes fully automated once the private link is in place — the same data model, the same customer experience, with the manual copy step replaced by a direct call to bio-infer.

5. Economics: one server is enough

The measured requirements make the tier affordable to own. ESMFold needs ~16 GB of VRAM and no sequence database; ColabFold/AlphaFold2 need 16–24 GB with MMseqs2; BepiPred-3.0 runs on the same class of GPU; NetMHCpan/NetMHCIIpan run on CPU. A single 4-GPU server with commodity CPU nodes hosts the entire catalog with room to spare, and because one deployment serves many users, the cost amortizes across the group. "Self-hosted" therefore does not mean "expensive" — it means the lab keeps its data on its own hardware while sharing the compute. Where a tenant lacks hardware, the same service can be offered at cost plus a thin margin, consistent with a deliver-at-cost operating model.

6. Discussion

The value of the GPU tier is not the models — those are published — but delivery: turning a hardware-gated method into a one-click result that lands next to the experiment, with its provenance and citations attached. Pairing this with the CPU pre-built tools gives a bench scientist the full bioinformatics landscape — the ubiquitous CPU long tail embedded directly, the GPU head brokered through a shared service — without a core facility.

Limitations. All outputs are computational predictions: structures carry per-residue confidence that must be read, and predicted epitopes and designs are hypotheses requiring wet-lab validation. Throughput is bounded by the GPU fleet; interactive folding is capped by sequence length. The frontier catalog (§2) is a roadmap, not yet shipped.

Future work. Automating the service tier over the private link, expanding the model catalog (generative design, docking, single-cell foundation models), and an assistant that routes each request to the right tier — CPU tool or GPU service — automatically, from the tool-screening library.

---

This article summarizes work from the NextELN Lab Assistant project, an Excellgen, Inc. product. Companion to the toolbox manuscript and the bioinformatics-tools landscape survey. In-silico predictions are hypotheses requiring wet-lab validation.