Dr. Lopez Aladid’s

Axiom Insilico

Predictive clinical models and generative AI to accelerate diagnosis, therapeutic decisions, and drug discovery

We transform clinical and omics data into interpretable, production-ready models, and apply generative AI to explore candidates and optimize hit-to-lead medical pipelines

Applied Artificial Intelligence for Medicine

Detection of brain tumors using artificial intelligence based on convolutional neural networks (CNNs).

01 Predictive Models for Diagnosis

Is your clinical data fragmented and underused, leading to an unreliable early detection and risk stratification?

Achieve earlier and more reliable diagnoses, consistent triage decisions, and models that hold up in real clinical settings, not just “on paper”.
To enable this, you can implement interpretable predictive models using EHR/EMR, laboratory data, imaging features, and omics, combining classical ML, deep learning, survival analysis, and multimodal approaches, validated with AUC/PR metrics, calibration, subgroup error analysis, and delivered with SHAP-based interpretability and reproducible documentation.

share your diagnostic case

02Precision Medicine Models for Treatment Efficacy

Are your treatment decisions made without knowing who will truly benefit?
Stop wasting time, budget, and patient opportunity

Simulate and anticipate treatment response at both patient and subgroup level, building decision-support tools and digital twins to explore therapeutic scenarios before acting.

We can implement response and survival models (PFS/OS) that predict treatment efficacy, control bias and confounding where needed, and are validated with ROC/PR metrics, clinically meaningful decision thresholds, and deployment-ready update strategies.

Which treatment do you want to predict?

Up: Radar plot showing pathway-level alignment between the digital twin’s gene expression profile and gemcitabine’s expected mechanisms. Down: Bar chart summarizing pathway modulation (green: high alignment; orange: partial; red: low).

03 Generative AI for Drug Discovery

Exploring chemical space manually is slow, expensive, and biased toward familiar scaffolds

You can accelerate hit discovery, reduce early-stage costs, and explore chemical space beyond familiar scaffolds without compromising scientific rigor.

We design generative AI pipelines for molecule generation, multi-objective scoring (affinity, ADMET, synthesizability), docking and rescoring, delivering ranked compound libraries, transparent selection logic, and fully reproducible, integration-ready workflows.

LET’S Define your target

Turn your data into decisions

Hospitals

risk, triage, readmission, sepsis

Biotech

hit discovery, lead optimization

Oncology

response prediction, relapse, survival

Clinical trials

cohort enrichment, endpoint optimization

MedTech

models embedded into products

Diagnostic labs

multimodal classifiers

Why work with Axiom Insilico?

Led by Rubén López Aladid, a biotechnologist with advanced training in clinical research and genetics, and extensive hands-on experience in data science and applied machine learning within medical and pharmaceutical environments. His background spans hospital-based and clinical research settings, as well as work within large pharmaceutical companies, giving him a deep understanding of how predictive models are evaluated, validated, and operationalized across the full clinical and drug development lifecycle.

His work focuses on translating real clinical and biomedical questions into accurate, production-grade AI systems, covering predictive modeling, survival analysis, multimodal data integration, and AI-driven drug discovery. This dual exposure to both clinical practice and industrial R&D allows him to design solutions that align with scientific rigor, regulatory expectations, and business constraints.

All projects delivered under Axiom Insilico adhere to strict standards of reproducibility, traceability, and maintainability, ensuring models can be audited, defended, and scaled over time. The emphasis is consistently on building decision-support systems that generate real-world value, prioritizing clinical utility and operational impact.

Rubén López Aladid

Data Science & Bioinformatics

Genetics & Genomics PhD.

How It Works

Building Decision-Ready AI for Medical Products

When working with clinical or biomedical data, acting without a clear decision framework is a risk. This process is designed to move you quickly from uncertainty to defensible action, saving time and enhancing credibility.

1 Define the scope

Before any modeling starts, the key question is clarified: What decision will this model change if it works?
This step prevents months of technically correct work that never translates into clinical, scientific, or business impact.

2 Audit your data

Data quality, leakage risks, bias, and feasibility are assessed upfront.
This critical step allows potential limitations to be identified early, before timelines, expectations, and decisions are locked in.

3 Build and validate

Models are developed and stress-tested using appropriate metrics, calibration, and subgroup analysis to reflect how they will behave outside the notebook, where consequences are real.

4 Make the model accurate and defensible

Interpretability and documentation are produced in parallel, so results can be explained to clinicians, reviewers, regulators, or leadership.

5 Delivery

Receive a deployment-ready output (API, dashboard), a clear hand-off, and a maintenance plan

BOOK A CALL

Frequently Asked Questions

How much data do you need to perform AI medical models?

It depends on the question, not just the volume. Many clinically useful models can be built with hundreds to a few thousand well-curated samples, especially when using structured clinical data or focused endpoints. During the initial data audit, feasibility is assessed and expectations are set upfront, before any heavy work begins.

How do you prevent data leakage and bias?

Leakage and bias are addressed by design, not after the fact. This includes proper temporal and group-aware splits, feature audits, strict separation of preprocessing steps, and subgroup error analysis.

Can this be used in regulated environments?

Yes. All projects are delivered with full traceability, reproducibility, and documentation, including model cards, clear assumptions, and validation reports.

What if external validation isn’t available?

When external datasets are not accessible, robust internal validation strategies are used (cross-validation, temporal splits, calibration analysis, and stress testing). Limitations are clearly documented, and a prospective or post-deployment validation plan is defined where needed.

How do you avoid biological bias caused by technical limitations?

Technical constraints are explicitly accounted for during study design and analysis. For example, the resolution limits of different sequencing technologies are considered to ensure conclusions are biologically and clinically valid, not artifacts of the method used.

What if the available data is not suitable for the intended analysis?

This is identified early in the process. When data limitations prevent meaningful conclusions, this is documented transparently, and alternative strategies, such as different sequencing techniques, revised endpoints, or complementary data sources are proposed.

Do you work with anonymized data?

Yes. Projects can be conducted entirely on anonymized or pseudonymized data, following data-minimization principles. No patient-identifiable information is required unless explicitly agreed and contractually defined.

Can models be integrated into our product?

Yes. Models can be delivered as APIs or containerized services, designed to integrate with existing clinical systems, products, or internal platforms. Handover, documentation, and update strategies are included to support long-term use.

Do you assess data quality before running any analysis?

Yes. All projects start with a data quality and integrity assessment, especially for sequencing data (e.g. FASTQ files). Coverage, enrichment, noise, and preprocessing artifacts are evaluated upfront to ensure the data is representative of the biological system and accurate by sequencing or sample-processing issues.

What happens if the sequencing data is low quality or poorly enriched?

Potential limitations are identified early and clearly communicated. Low-quality or poorly enriched data can introduce bias or leakage, leading to misleading conclusions. Detecting these issues early allows corrective actions—such as reprocessing, filtering, or redefining the analytical scope—before results are interpreted or acted upon.

Can all sequencing approaches support biomarker discovery?

No. The analytical resolution must match the biological question. For instance, 16S rRNA sequencing is limited to genus-level resolution, which is often insufficient for robust biomarker discovery. When species-level resolution is required, shotgun metagenomics or alternative approaches are recommended to avoid inconclusive or non-actionable results.

How does this improve interpretability and reliability of results?

By validating data quality and technical suitability upfront, downstream analyses remain interpretable, realistic, and biologically grounded. This ensures that any predictive model or biomarker reflects the disease biology rather than technical noise or methodological bias.

Ready to turn your data into clinical-grade decisions?

Let’s take your project further with a comprehensive plan:

Book discovery call