Artificial Intelligence for the Wet Lab: A Reference-Grounded Toolbox for Protein Structure, Antibody Antigen Design, and mRNA Vaccine Discovery

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

---

Abstract

Artificial intelligence (AI) has moved from a peripheral analysis aid to a central engine of biomedical discovery. Deep learning now predicts protein structure at near-experimental accuracy, designs antibodies and antigens in silico, and assembles vaccine candidates directly from a pathogen genome. Yet these capabilities remain fragmented across command-line tools, web servers with submission limits, and code that assumes a bioinformatics core facility. Here we describe a reference-grounded computational toolbox integrated into an electronic lab notebook (ELN), built on a curated library of 542 primary references spanning AI in biomedical research, protein-structure prediction, antibody/epitope design, mRNA-vaccine immunoinformatics, multi-omics, drug discovery, medical imaging, and large language models in medicine. We define three first-wave tools — an interactive 3D protein-structure viewer backed by ESMFold/AlphaFold, an antigenic-peptide finder for raising antibodies in mice and rabbits, and an mRNA-vaccine candidate designer — and show that all of them, including the GPU-bound structure-prediction and protein-language-model steps, run comfortably on a single 4-GPU server with commodity CPU nodes. The result is a design-to-notebook workflow in which a bench scientist pastes a sequence and receives a reagent, its rationale, and its citations, without leaving the lab notebook.

---

1. Introduction

The last five years have reframed what computation can contribute to the life sciences. Large-scale protein-language and structure-prediction models have made the three-dimensional fold of essentially any protein a few seconds of inference away [1–4], while generative and agentic systems now propose therapeutics, design experiments, and even run portions of the discovery loop autonomously [5–8]. Reviews and news analyses in Nature, Science, and PNAS document both the breadth of this shift and its failure modes — benchmarking crises, reproducibility gaps, and a flood of low-quality AI-assisted papers [9–12]. The through-line is clear: AI is now a routine instrument of biomedical research, and the bottleneck is no longer capability but access — getting these methods into the hands of bench scientists in a form they can trust and reuse.

Three application areas are mature enough to productize for everyday wet-lab use:

• Protein structure prediction. AlphaFold2, RoseTTAFold All-Atom, and the ESM family of protein-language models predict folds, complexes, and the effects of mutations with accuracy that increasingly rivals crystallography, and can even serve as a structural prior for experimental determination [1–4, 13–15]. Protein-language models are now trainable and shareable by ordinary labs [16, 17].
• Antibody and antigen design. B-cell epitope prediction has progressed from physicochemical scales to protein-language-model and structure-aware predictors; deep learning now guides antibody affinity maturation and antigen target selection [18–23].
• Vaccine design. Reverse vaccinology and immunoinformatics assemble multi-epitope and mRNA vaccine constructs directly from genome sequence, with deep learning increasingly driving epitope selection and sequence optimization [24, 25].

In parallel, multi-omics integration and large language models are reshaping diagnostics and clinical reasoning [26–33], though these are largely analysis overlays on data an ELN already holds and form a natural second wave rather than a first one.

This manuscript describes how we translate that literature into working tools inside NextELN, an AI-native ELN, and argues that the compute required is modest enough for a single institution to own.

2. A reference-grounded approach

Every tool in the toolbox is anchored to primary literature rather than to a model's unattributed prior. We assembled a library of 542 references by programmatic search of PubMed across eight themes plus the full citation list of a recent systematic review of AI in biomedical engineering [34]. The library is stored as a structured, searchable table inside the ELN, so the AI assistant can cite it when it recommends a method, and so each tool's output can carry the papers that justify it. Composition by theme is given in Table 1.

Table 1. Reference library composition (n = 542).

Theme	References
AI + multi-omics	80
Protein structure prediction (AlphaFold/ESMFold)	70
Antibody / B-cell epitope antigen design	69
AI in biomedical research (Nature/Science/PNAS)	67
mRNA vaccine design / immunoinformatics	60
AI drug discovery	57
Cited by the Wiley IJCP 2025 biomedical-AI review	52
LLMs / foundation models in medicine	44
AI medical imaging / diagnostics	43

3. The toolbox

3.1 Tool 1 — Interactive 3D protein-structure viewer

Problem. A scientist has a sequence and wants to see its fold, its surface, and where the confident and disordered regions lie.

Design. The tool accepts a raw protein sequence or a UniProt/PDB accession. If an experimental or pre-computed structure exists (RCSB PDB or the AlphaFold Protein Structure Database) it is fetched; otherwise the sequence is folded on demand with ESMFold, which predicts a structure from a single sequence in seconds without a multiple-sequence alignment [3, 4]. The structure renders interactively in the browser (cartoon/surface/stick, spectrum and per-residue confidence coloring), with the predicted local-distance difference test (pLDDT) mapped to color so users can judge which regions to trust. The structure and its provenance are saved to a notebook page in one click.

Why it matters. Structure is the substrate for the other two tools: surface exposure defines good antibody epitopes, and folded antigens anchor vaccine design [1–4, 18].

3.2 Tool 2 — Antigenic-peptide finder for antibody production

Problem. To raise a polyclonal antibody in a mouse or rabbit, a scientist must choose a short peptide (typically 12–20 residues) that is immunogenic, surface-exposed, synthesizable, and distinct from host proteins.

Design. The tool combines classic, instantly computable propensity scales — Parker hydrophilicity, Emini surface accessibility, Kolaskar–Tongaonkar antigenicity, Chou–Fasman β-turn propensity, and Karplus–Schulz flexibility [37] — with an optional protein-language-model overlay (BepiPred-3.0, built on ESM-2) and structure-derived relative solvent accessibility from Tool 1 [35, 18]. Candidate peptides are ranked by a combined immunogenicity score that rewards surface exposure, hydrophilicity, and flexible loop/terminal regions, and penalizes synthesis-hostile features (internal cysteines, aggregation-prone runs) and similarity to the host proteome (to avoid self-tolerance). Output is a ranked table with positions, scores, and conjugation guidance (terminal Cys for KLH/BSA coupling). Recent structure-aware predictors such as AbEpiTope, which uses AlphaFold and inverse folding, motivate the structure overlay [18].

3.3 Tool 3 — mRNA-vaccine candidate designer

Problem. Given a viral genome, propose mRNA-vaccine candidates.

Design. The tool implements a reverse-vaccinology pipeline: identify open reading frames and candidate antigens (surface/structural proteins); predict CD8+ (MHC-I) and CD4+ (MHC-II) T-cell epitopes [36] and B-cell epitopes [35]; assemble a multi-epitope construct (Kozak sequence, signal peptide, epitopes joined by AAY/GPGPG/GS linkers, a molecular adjuvant, stop codon); codon-optimize for codon adaptation index and minimum free energy; and append 5′/3′ UTRs and a poly-A tail. Candidates are scored for predicted population coverage. Deep learning increasingly drives both epitope selection and sequence optimization in next-generation vaccine design [24, 25]. Output is a set of ranked constructs with the final optimized mRNA sequence, an epitope map, and a cited design rationale saved to the notebook.

4. Infrastructure: this is affordable to own

A frequent objection to self-hosting AI biology is GPU cost. Measured requirements refute it for this toolbox (Table 2). ESMFold folds most proteins in ~16 GB of VRAM in seconds and needs no sequence database; ColabFold/AlphaFold2 needs 16–24 GB and, via MMseqs2, avoids the multi-terabyte reference databases; BepiPred-3.0 runs on the same class of GPU; and NetMHCpan/NetMHCIIpan run on CPU. All of these fit on a single 4-GPU server plus CPU nodes with 160 GB RAM — the hardware assumed here.

Table 2. Compute requirements vs. available hardware.

Component	Requirement	Fits a 4-GPU + 160 GB-RAM site?
ESMFold (single-sequence folding)	~16 GB VRAM, seconds/protein, no MSA DB	Yes (1 GPU)
ColabFold / AlphaFold2	16–24 GB VRAM; MMseqs2 avoids 2.5 TB DB	Yes (1 GPU + ~32 GB RAM)
BepiPred-3.0 (B-cell epitopes)	ESM-2, GPU	Yes (1 GPU)
NetMHCpan / NetMHCIIpan (T-cell epitopes)	CPU only	Yes (CPU node)
Classic scales, ORF, codon optimization	trivial CPU	Yes (in the web app)
3D rendering	client-side WebGL	Yes (browser)

Architecturally, pure-algorithm tools (propensity scales, ORF finding, codon optimization) run in-process in the ELN's Node/TypeScript backend, while model inference runs behind a small GPU microservice that the app calls with a shared token, caches by input hash, and degrades gracefully to the CPU scales when unavailable. Because the same models can serve external users, several tools double as billable services, so the infrastructure amortizes across tenants — consistent with a deliver-at-cost operating model.

5. Discussion

The literature and the hardware agree: the first-wave tools are buildable, useful, and affordable to run in-house. The value added by embedding them in an ELN is not the models themselves — those are published — but the workflow: input arrives as a sequence a scientist already has, output lands as a reagent in the notebook next to the experiment that needs it, and every recommendation carries its citations. This directly counters two documented pathologies of AI-in-science — opaque, unattributed outputs and unreproducible pipelines [9–12] — by making provenance a first-class feature.

Limitations. In-silico predictions are hypotheses: predicted epitopes and vaccine constructs require wet-lab validation; ESMFold and AlphaFold confidence must be read from pLDDT and not assumed uniform; codon and structure heuristics encode assumptions that vary by expression host and pathogen. The tools are decision-support instruments, not substitutes for immunization, titer, and challenge experiments. Benchmarking must be ongoing and honest [10].

Future work. Second-wave tools follow the same pattern for multi-omics analysis and an omics-aware assistant [26–33], plus molecular docking and antibody humanization. Each is gated on the same principle: ship only what the hardware can run and the literature can justify.

6. Methods (brief)

The reference library was built by PubMed E-utilities search across the eight themes of Table 1, with titles, authors, journals, years, DOIs, and abstracts retrieved by efetch and stored as a structured ELN table; the citation list of the systematic review at DOI 10.1155/ijcp/9888902 was resolved to PubMed IDs via Crossref and NCBI and merged. Tool implementations follow the ELN's existing conventions: client-only WebGL components for visualization, Node/TypeScript modules for pure algorithms, and a token-authenticated Python (FastAPI) microservice on the GPU server for model inference.

---

References

1. Jumper J, et al. (2021). Highly accurate protein structure prediction with AlphaFold. Nature.
2. Krishna R, et al. (2024). Generalized biomolecular modeling and design with RoseTTAFold All-Atom. Science. doi:10.1126/science.adl2528
3. Lin Z, et al. (2023). Evolutionary-scale prediction of atomic-level protein structure with a language model (ESMFold). Science.
4. Chen L, et al. (2024). AI-Driven Deep Learning Techniques in Protein Structure Prediction. Int J Mol Sci. doi:10.3390/ijms25158426
5. Xing J, et al. (2026). Deep-learning-based de novo discovery and design of therapeutics that reverse disease-associated transcriptional phenotypes. Cell. doi:10.1016/j.cell.2026.02.016
6. Huang K, et al. (2026). Autonomous biomedical research with an artificial intelligence agent. Science. doi:10.1126/science.adz4351
7. Gottweis J, et al. (2026). Accelerating scientific discovery with Co-Scientist. Nature. doi:10.1038/s41586-026-10644-y
8. Li B, et al. (2026). Agentic AI and the rise of in silico team science in biomedical research. Nat Biotechnol. doi:10.1038/s41587-026-03035-1
9. Teo ZL, et al. (2025). Generative artificial intelligence in medicine. Nat Med. doi:10.1038/s41591-025-03983-2
10. Mahmood F (2025). A benchmarking crisis in biomedical machine learning. Nat Med. doi:10.1038/s41591-025-03637-3
11. Naddaf M (2025). AI linked to explosion of low-quality biomedical research papers. Nature. doi:10.1038/d41586-025-01592-0
12. Gallifant J, et al. (2025). The TRIPOD-LLM reporting guideline for studies using large language models. Nat Med. doi:10.1038/s41591-024-03425-5
13. Fadini A, et al. (2026). AlphaFold as a prior: experimental structure determination conditioned on a pretrained neural network. Nat Methods. doi:10.1038/s41592-026-03047-4
14. Jiang K, et al. (2025). Rapid in silico directed evolution by a protein language model with EVOLVEpro. Science. doi:10.1126/science.adr6006
15. Drake ZC, et al. (2025). Deep-learning structure elucidation from single-mutant deep mutational scanning. Nat Commun. doi:10.1038/s41467-025-62261-4
16. Su J, et al. (2025). Democratizing protein language model training, sharing and collaboration. Nat Biotechnol. doi:10.1038/s41587-025-02859-7
17. Ma H, Ji Z (2025). Machine Learning for Protein Science and Engineering. Cold Spring Harb Perspect.
18. Clifford JN, et al. (2025). AbEpiTope-1.0: Improved antibody target prediction by use of AlphaFold and inverse folding. Sci Adv. doi:10.1126/sciadv.adu1823
19. Joubbi S, et al. (2024). Antibody design using deep learning: from sequence and structure design to affinity maturation. Brief Bioinform. doi:10.1093/bib/bbae307
20. Liang T, et al. (2024). AI-based IsAb2.0 for antibody design. Brief Bioinform. doi:10.1093/bib/bbae445
21. Li J, et al. (2025). Significantly enhancing human antibody affinity via deep learning and computational biology-guided single-point mutations. Brief Bioinform. doi:10.1093/bib/bbaf445
22. Drost F, et al. (2025). Benchmarking of T cell receptor-epitope predictors with ePytope-TCR. Cell Genom. doi:10.1016/j.xgen.2025.100946
23. Bhattacharya M, et al. (2025). Deep learning in next-generation vaccine development for infectious diseases. Mol Ther Nucleic Acids. doi:10.1016/j.omtn.2025.102586
24. Cernuto F, et al. (2025). In-silico epitope-based vaccines design: progress, challenges and the road ahead. Expert Opin Drug Discov. doi:10.1080/17460441.2025.2599178
25. Tombari W, et al. (2025). A novel mRNA-based multi-epitope vaccine for rabies virus computationally designed via reverse vaccinology and immunoinformatics. Sci Rep. doi:10.1038/s41598-025-16143-w
26. Zhang J, et al. (2025). Deep learning-driven multi-omics analysis: enhancing cancer diagnostics and therapeutics. Brief Bioinform. doi:10.1093/bib/bbaf440
27. Huang X, et al. (2025). Single-cell multi-omics and machine learning for dissecting stemness in cancer. Brief Bioinform. doi:10.1093/bib/bbaf566
28. Morello G, et al. (2025). Artificial Intelligence-Driven Multi-Omics Approaches in Glioblastoma. Int J Mol Sci. doi:10.3390/ijms26199362
29. Alhamrani SQ, et al. (2025). Machine Learning for Multi-Omics Characterization of Blood Cancers: A Systematic Review. Cells. doi:10.3390/cells14171385
30. Tordjman M, et al. (2025). Comparative benchmarking of the DeepSeek large language model on medical tasks and clinical reasoning. Nat Med. doi:10.1038/s41591-025-03726-3
31. Iqbal U, et al. (2025). Impact of large language model (ChatGPT) in healthcare: an umbrella review and evidence synthesis. J Biomed Sci. doi:10.1186/s12929-025-01131-z
32. Danning R, et al. (2025). LACE-UP: An ensemble machine-learning method for health subtype classification on multidimensional binary data. PNAS. doi:10.1073/pnas.2423341122
33. Yılmaz S, et al. (2025). Bias-aware training and evaluation of link prediction algorithms in network biology. PNAS. doi:10.1073/pnas.2416646122
34. Systematic review of AI and ML in biomedical engineering (2025). International Journal of Clinical Practice. doi:10.1155/ijcp/9888902
35. Clifford JN, Høie MH, Deleuran S, Peters B, Nielsen M, Marcatili P (2022). BepiPred-3.0: Improved B-cell epitope prediction using protein language models. Protein Science. doi:10.1002/pro.4497
36. Reynisson B, Alvarez B, Paul S, Peters B, Nielsen M (2020). NetMHCpan-4.1 and NetMHCIIpan-4.0: improved predictions of MHC antigen presentation by concurrent motif deconvolution and integration of MS MHC eluted ligand data. Nucleic Acids Research. doi:10.1093/nar/gkaa379
37. Vita R, Mahajan S, Overton JA, Dhanda SK, Martini S, Cantrell JR, Wheeler DK, Sette A, Peters B (2019). The Immune Epitope Database (IEDB): 2018 update — the analysis resource implementing the classic antigenicity, hydrophilicity, surface-accessibility and flexibility scales. Nucleic Acids Research. doi:10.1093/nar/gky1006

This article summarizes work from the NextELN Lab Assistant project, an Excellgen, Inc. product. In-silico predictions are hypotheses requiring wet-lab validation. The 542-reference library underlying it is available in the NextELN workspace as the ai_biomed_refs table.