Bioinformatics Research

🪥Single-Cell Biology

Making sense of single-cell atlases, and how well what's learned from one transfers to another.

scGPT hyperparameter transfer across tissues

A 2026 Bioinformatics paper found that Bayesian-optimized fine-tuning settings meaningfully improve scGPT — a single-cell foundation model — at integrating multi-omic data, but only tested this on bone-marrow blood cells. Does that same already-tuned configuration still work when reused, unmodified, on a different tissue, or does it need its own expensive re-optimization? Surfaced by cross-checking three independent AI “deep research” runs against each other for genuinely open questions.

Read the full write-up → — all three regret tables, the UMAP figures, and the data behind them. Notebooks and results are public at github.com/kaysagit/scgpt-hyperparameter-transfer.

⏳Epigenetics & Aging

How well a molecular signal — DNA methylation — tracks the number that supposedly matters least in biology and most everywhere else: age.

Replicating Horvath's epigenetic age clock

In 2013, Steve Horvath showed that DNA methylation levels at a specific set of 353 CpG sites predict a person's chronological age to within about 3.6 years — one of the most independently-replicated results in aging biology, and the foundation the whole later family of "epigenetic clock" biomarkers (Hannum, PhenoAge, GrimAge) built on. This project checks that result directly: take Horvath's own published coefficients, unmodified, and score them against a real cohort that was never used to build the clock, to see how close the prediction actually lands.

Predicted vs. actual, n=656

Scored zero-shot (no retraining) against 656 real whole-blood samples from an independent cohort (Hannum et al. 2013): MAE 4.77 years, r=0.918, comfortably clearing the sanity floor of 11.91 years (always predicting the mean). The gap versus Horvath's own reported 3.6-year MAE is expected, not a concern — this cohort was used to build a different clock (Hannum's), never Horvath's, so this is genuine cross-cohort external validation rather than an in-sample number.

🎯Cancer Immunology & Neoantigens

When a tumor's DNA mutates, the resulting altered proteins can get displayed on the cell surface for the immune system to recognize — a displayed mutant fragment is a neoantigen, and it's the basis of most modern cancer immunotherapy. Which mutations actually produce ones the immune system notices is still one of the field's open questions.

Immunogenicity-aware neoantigen landscape of chromophobe kidney cancer

Chromophobe renal cell carcinoma (KICH) is a rare, understudied kidney cancer with a distinctive genomic quirk: relatively few point mutations, but a disproportionate share of small insertions/deletions (indels). An indel that shifts the DNA reading frame produces a stretch of protein sequence the immune system has never seen — and frameshift-derived neoantigens are reported in the literature as unusually immunogenic. Nobody had pointed a modern immunogenicity-aware scorer (not just a binding-affinity predictor) at this specific cancer type before. This project builds that catalog from scratch and tests whether that expectation holds up here.

The original angle — benchmarking immunogenicity predictors against each other — turned out to be oversaturated by well-resourced 2023–2026 papers already doing exactly that; this project instead points an established scorer (PRIME2.1) at real, previously-unanalyzed patient data. One real, disclosed limitation runs through every result below: true per-patient HLA genotypes turned out to be gated behind restricted-access archives everywhere checked, so every candidate is scored against a fixed population-representative panel of common HLA-I types, not each patient's real genotype.

Correction, 2026-08-18: the finding originally published here (frameshift candidates score less immunogenic than missense, p≈3×10⁻¹⁸) was retracted after external review caught a pseudoreplication bug: the significance test ran on all sliding-window candidate peptides as independent observations, but frameshift mutations mechanically generate ~2.7× more overlapping peptide-windows per event than missense mutations, inflating the p-value asymmetrically. Re-run correctly at the mutation level (one representative peptide per mutation, n=66 frameshift vs. 865 missense): no significant difference, p=0.54. Fixing the same bug in the burden metric used for survival analysis had the opposite effect — it turned a previously-dismissed result into the project's one real finding, below.

Headline result (corrected): immunogenicity-weighted neoantigen burden associates with worse survival in this cohort (Cox HR=2.14, p=0.00073; log-rank p=0.0092 — the two tests now agree, and the result survives removing the single largest-burden patient). The originally-published mutation-type comparison and the immune-infiltration correlation are both honest null results; only the survival association held up under a corrected, non-inflated analysis. Caveat carried forward plainly: this rests on only 65 patients and 9 observed deaths (8 without the outlier) — a real, worth-following-up signal, not a definitively established one.

Frameshift vs. missense immunogenicity in MSI-High gastric cancer

A direct follow-on to the KICH project: after that project's original mutation-type finding turned out pseudoreplicated and null once corrected — at only 66 frameshift mutations, too few to detect anything but a large effect — this project reruns the identical corrected pipeline (PRIME2.1 + MixMHCpred3.0, same fixed 10-allele HLA-I panel, mutation-level aggregation built in from the start) on a cohort chosen specifically to have the statistical power KICH lacked: the MSI-High subset of TCGA-STAD (gastric adenocarcinoma), a classic mismatch-repair-deficient hypermutator with indels concentrated in mononucleotide-repeat slippage — a deliberately different mutational mechanism than KICH's own non-MSI indel skew.

80 MSI-High patients identified via two independent open sources (cBioPortal's MSIsensor score and TCGA's own published molecular subtype), cross-validated against each other. Real data confirmed via GDC's current mutation pipeline: 23,808 frameshift indels across all 80 patients (vs. KICH's 88 across 43/66) — every single patient has both mutation types, unlike KICH's 35% with zero frameshift mutations, making a per-patient paired comparison possible for the first time.

Where this stands: the question has split in two — the headline half is resolved, a narrower half is under active test. An equal-N subsampling check (2026-09-02) removed the window-count confound by construction: for every mutation, sample a fixed number of candidate windows before taking the best score, so both arms feed the same number of draws into the best-of-N. Across every configuration the naive effect collapses — rank-biserial r falls from 0.14 to |r| ≤ 0.05 — and the cleanest apples-to-apples comparison (frameshift mutations whose window count genuinely overlaps missense's, both subsampled to a common N) is a flat null: p=0.66. So at the level of a whole mutation event, a frameshift does not make its best candidate neoantigen meaningfully more or less immunogenic than a missense one, once frameshift stops getting more attempts at the target. That confirms the KICH project's corrected null at ~240× the frameshift sample size; the p=7×10⁻¹⁴¹ was almost entirely the multiplicity artifact.

What remains is a per-peptide question: holding the number of attempts exactly equal, is a frameshift-derived peptide itself scored differently? Compared strictly within bins of identical candidate-window count — and deduplicated against the MSI-High habit of producing the same neo-ORF peptide at recurrent slippage hotspots — this is not a flat null: it changes sign with neo-ORF length. Frameshift scores worse for short novel reading frames and better for long, well-expressed ones; in the bin matching a normal full-length missense mutation, frameshift is favored at p=9×10⁻¹². Whether that is real per-peptide biology or an artifact of PRIME's %Rank scale at its extremes was checked with a structure-aware second opinion: ColabFold structures for the 302 contested peptide-MHC complexes, re-scored by ImmunoStruct (a 2025 model built on predicted structure and biochemistry, not just sequence, trained independently of PRIME). The result does not corroborate PRIME's signal — flat null (p=0.91, rank-biserial r=−0.008; a feature-sensitivity variant agrees). It isn't simply noise: ImmunoStruct's own scores still correlate with PRIME's %Rank within this set (Spearman=−0.19, p=7×10⁻⁴) — it specifically doesn't see a class-level frameshift effect where PRIME sees a large one. Honest read: the per-peptide question is now unresolved rather than confirmed reversed-to-null; PRIME's signal alone looks meaningfully weaker under an independent, structure-aware model. The available ImmunoStruct checkpoint is infectious-disease-trained rather than cancer-trained, so a stronger cancer-specific comparative check (folding the wild-type peptides too) is the natural next step, not yet run.

🧬Genome Language Models

Foundation models trained directly on raw DNA sequence, and whether what they learn transfers to predicting the real-world consequences of a mutation — without ever looking at a protein sequence, a structure, or an alignment.

Zero-shot variant-effect prediction in bacteriophage ΦX174 with a DNA language model

A 2026 preprint (Wei, Li & Lehner) measured the fitness effect of essentially every possible mutation across bacteriophage ΦX174's genome — every single-nucleotide substitution and every single-amino-acid substitution, ~44,500 variants in total — then benchmarked roughly 15 protein-sequence variant-effect predictors (the ESM family, EVE, GEMME, VESPA, Progen2, and more) against the real measurements. Every one of those predictors reads protein sequence, or a multiple-sequence alignment built from many related proteins; none of them look at DNA directly. This project fills that specific gap: Evo2, a genome-scale DNA language model pretrained across bacteria, archaea and phage, scored zero-shot on the exact same benchmark.

Method: for each of the 44,478 designed mutations, compare Evo2's own likelihood of the mutant genome against the wild-type genome — a bigger drop in likelihood means the model finds that specific DNA change more disruptive — and correlate that drop against the mutation's real, lab-measured effect on phage fitness, gene by gene, using the paper's own metric (per-gene Spearman correlation).

At the amino-acid level — the paper's own benchmark, identical rows, identical metric — Evo2 ranks 16th of 35 predictors (mean per-gene Spearman ρ=0.257, median 0.329, n=28,380 variants). A model that has never seen a protein sequence or a multiple-sequence alignment beats every single ESM-family model tested (ESM1v, ESM2 up to 15B parameters, ESM3, ESMC up to 6B), plus Tranception and PoET — while trailing the structure-aware and MSA-based specialists (ESCOTT, GEMME) and the best single-sequence protein language models (VESPA, Progen2).

Amino-acid level, excerpt of 35 predictors

At the nucleotide level — a comparison no prior study could make, since no protein model can be scored on raw DNA — Evo2's own predictions correlate strongly and consistently with measured fitness for every gene with an unambiguous reading frame: Spearman ρ=0.55–0.65 (all p<0.001). Genes B, E and K have no such unambiguous nucleotide positions at all — every base in them also sits inside another gene's overlapping reading frame, a real quirk of ΦX174's densely packed genome, not a gap in the analysis.

Nucleotide level, per gene

Caveat worth keeping attached to any of this: this is zero-shot in the sense that the fitness measurements were never a training target — the same situation every protein model in the comparison faces with ΦX174's own proteins — but ΦX174 (the first DNA genome ever sequenced) was almost certainly present somewhere in Evo2's training data, so it is not a test of generalization to a genuinely unseen genome.