RSO.
  • Tools
  • Experience

Public research software

Tools built to make scientific reasoning inspectable.

My projects cover different biological questions, but share the same standard: clear scope, reproducible inputs, visible uncertainty and no inflated claims.

The eight tools at a glance

Grouped by the question they answer. Each one addresses a decision that comes up early, before anyone commits time or money to a candidate.

Field Tool The problem What it is for What it adds Key skills applied
Target discovery & safety TargetIntel-IO A gene linked to disease is not automatically a good drug target, yet most tools collapse that judgement into a single score. Sorting candidate genes into genuine targets, biomarkers, resistance mechanisms and poor candidates. Three separate rankings, one per therapeutic strategy, plus the evidence for and against each call.
  • Open Targets & DepMap
  • DuckDB
  • Benchmarking
  • Sensitivity analysis
  • Grounded LLM synthesis
  • GraphRAG
  • CI/CD
Activation Liability A target can look safely tumour-specific in healthy resting tissue and stop looking that way once the tissue is inflamed. Spotting targets whose apparent safety margin shrinks under immune activation. A direct resting-versus-activated comparison in public data, with abstention when the evidence is too thin.
  • Single-cell RNA-seq
  • Public GEO cohorts
  • Negative-binomial models
  • Donor-level robustness
  • Reproducible manifests
reach-gap A target can be highly expressed and still sit too far from a working blood vessel for an antibody to reach it. Asking whether target-positive tumour cells are physically accessible to an antibody. Mechanistic diffusion modelling paired with an evidence graph that returns NOT_COMPUTED rather than guessing.
  • Reaction–diffusion modelling
  • Spatial transcriptomics
  • Uncertainty propagation
  • Evidence provenance
  • Docker
Method reliability segmentation-fragility In spatial data, a small change in where a cell boundary is drawn can move a marker between cell types and flip the conclusion. Testing whether a spatial-transcriptomics claim survives a change in transcript-to-cell assignment. Each claim is labelled robust, segmentation-sensitive or not reportable, instead of trusting one segmentation as final.
  • Xenium / ProSeg / Baysor
  • Claim-level auditing
  • Planted-truth simulation
  • Abstention logic
FM Value Audit A foundation-model embedding can look biologically structured and still add nothing to a real decision. Testing whether single-cell foundation models beat cheaper baselines at target prioritisation. A leakage-aware temporal holdout designed to report a negative result when that is what the data shows.
  • Geneformer & scGPT
  • Leakage control
  • Temporal holdout
  • AUPRC benchmarking
  • Preregistration
Drug response prediction DrugMatch-Confidence Cancer models respond very differently to the same drug, and predictions rarely say how much to trust them. Predicting how a preclinical cancer model will respond to a given drug. Every prediction carries a calibrated probability, an uncertainty interval and a warning when the input is unfamiliar.
  • XGBoost
  • SHAP
  • Conformal prediction
  • Probability calibration
  • Grouped cross-validation
  • DepMap / PRISM / GDSC2
Protein deep learning MTF Prediction Whether a membrane protein carries transcription-factor-like features can change depending on where you draw its boundary. Screening human and mouse membrane-protein domains for those features under different boundary definitions. The same screen repeated across three boundary conditions, so a result that holds under only one is visible as such.
  • PyTorch
  • DeepTFactor
  • Captum integrated gradients
  • UniProtKB
Research engineering Dev Autopilot AI coding agents can move fast through a repository and leave no reliable trail of what changed or why. Running AI agents on real code with hard limits on what they may touch. Separate agents implement, audit and review; deterministic gates decide whether work advances, and a person keeps merge authority.
  • Agentic AI orchestration
  • Pydantic
  • SQLite
  • Strict typing (mypy)
  • State machines
  • CI/CD

01 v0.6.0 · Research preview

TargetIntel-IO

Therapeutic-intent-aware target triage

A deterministic platform for interpreting candidate genes in anti-PD-1-resistant melanoma. It avoids one universal score by separating antibody/IO, resistance-biomarker and small-molecule strategies.

  • Stable role classification and three intent-specific rankings
  • Typed, source-linked evidence with mandatory human-review boundaries
  • Real DepMap Public 26Q1 dependency layer that never silently changes the baseline
  • Graph-native evidence with a deterministic, GraphRAG-compatible research-preview export

Read the case study View repository ↗

Current scope

331genes in the discovery universe
56targets in the internal benchmark
0differing artefacts across two independent runs

The benchmark tests implementation consistency. It is not an independent clinical validation dataset.

02 v0.5.0 · Research preview

Activation Liability

Detecting hidden safety risk in surface targets

A target can look tumour-selective against resting healthy tissue and lose that selectivity once normal cells are inflamed or immune-activated. alia audits how much of that apparent selectivity survives biologically matched activation.

  • Paired resting and activated public single-cell cohorts
  • Donor-level robustness, negative controls and explicit abstention
  • Frozen manifests and checksums for every executed cohort

Read the case study View repository ↗

Verified release

52automated tests · 85.4% non-I/O coverage
2evidence types: RNA and paired tissue

A liability detector with abstention. It does not predict toxicity or establish a therapeutic window.

03 v0.8.0 · Research preview

reach-gap

Is the target physically reachable?

A tumour target can be strongly expressed and still sit too far from a perfused vessel for an antibody to engage it. reach-gap couples simplified reaction–diffusion modelling with an explicit evidence graph.

  • Separates same-tissue measurement from external validation and literature priors
  • Relative target ranking across six defensible vessel definitions
  • Returns NOT_COMPUTED instead of inventing an absolute answer

Read the case study View repository ↗

Verified release

117automated tests
20,000uncertainty draws per analysis

Research use only. It does not predict patient response or estimate clinical efficacy.

04 v0.5.0 · Research preview

segmentation-fragility

Does the conclusion survive the segmentation?

In spatial transcriptomics a small boundary change can move a marker between cell types and flip a biological claim. This tool makes the claim, not the segmented object, the unit of analysis.

  • Compares claims across Xenium, ProSeg and Baysor with fixed biological anchors
  • Labels each claim robust, segmentation-sensitive or not reportable
  • Abstains when claim-specific anchor coverage is insufficient

Read the case study View repository ↗

Verified release

62automated tests · 90.3% line coverage
77.1Mtranscript records in the Xenium perturbation

Robust and sensitive refer to a predeclared numerical threshold, not to clinical safety.

05 v0.5.0a1 · Pre-release

FM Value Audit

Do foundation models actually add value?

A leakage-aware benchmark testing whether representations from Geneformer and scGPT improve a real target-prioritisation decision over cheaper, easier-to-explain baselines.

  • Prespecified decision metric evaluated on unseen future outcomes
  • Explicit leakage control across feature selection and tuning
  • Designed to report a negative result when that is what the data shows

Read the case study View repository ↗

Current finding

NOT
SUPPORTED
static Geneformer geometry did not improve AUPRC
79,542rows in the temporal holdout

The contextual-embedding protocol is preregistered but not yet executed. This release is not evidence that foundation models help.

06 v1.0.1 · Released

DrugMatch-Confidence

Interpretable preclinical cancer drug-response prediction

Five real DepMap/PRISM XGBoost bundles produce continuous response, sensitive/resistant classification, calibrated probabilities, conformal intervals, out-of-distribution warnings and SHAP explanations.

  • Training-only feature selection and cell-line-grouped splits
  • Five-fold out-of-fold evaluation with locked GDSC2 validation
  • Honest release labels: validated, exploratory or insufficient evidence

Read the case study View repository ↗

Release evidence

0.84Trametinib strict GDSC2 AUROC · n=20
0.83Afatinib strict GDSC2 AUROC · n=22

Preclinical research only. These are small external subsets and not patient-response models.

07 v1.0.0 · Research analysis

MTF Prediction

Topology-aware deep learning for membrane proteins

A reproducible research archive screening human and mouse membrane-protein domains for transcription-factor-like sequence features under alternative transmembrane-boundary definitions.

  • Human and mouse UniProtKB membrane-protein domains
  • DeepTFactor inference across no-, half- and full-TMD conditions
  • Residue-level interpretation with Captum Integrated Gradients

Read the case study View repository ↗

Analysis matrix

2mammalian species
3 × 3domain classes × boundary conditions

Candidate-generating predictions only; experimental validation is required.

08 v0.1.1 · Alpha

Dev Autopilot

Persistent, auditable development orchestration

Coordinates implementation, adversarial audit and independent review while SQLite records state transitions, retries and decisions across restarts.

  • Fail-closed state machine and explicit path permissions
  • Mechanical scope, test, context and Git-metadata gates
  • Stops at human review; it never commits, pushes or merges product code

Read the case study View repository ↗

Design principle

Agents advise, implement and review. Deterministic gates decide whether work can advance. A person retains scientific and merge authority.

Collaboration

Interested in the science or the implementation?

Every repository is public. For research, product or computational biology conversations, feel free to get in touch.

Email me All GitHub repositories

© 2026 Rafael Soler Ortuño

 

Computational biology · Translational science · Research software