Four teams, one proof.

Mission

Own the model end to end: set the research direction, pick the architectures, and land a dose-conditioned liver-injury model that beats a free pharmacokinetic heuristic on chemistry it never saw. The headline result is yours, win or lose.

What you will do

• Set the modeling roadmap: one shared molecular encoder, an exposure head, a mechanism head, and an integration head that turns the two into a margin and a call.
• Make the small-data problem the priority it is. The mechanistic label sets run to a few hundred compounds, so multi-task and transfer learning are the core method, not a flourish.
• Own calibration and conformal prediction: every call ships with a coverage guarantee and an efficiency figure, and abstains outside its applicability domain.
• Own the benchmark harness and the negative controls. Shuffled labels, structure-blind baselines and ablations run alongside every result, so leakage shows up as leakage.
• Spend the compute on evaluation breadth rather than model size: nested cross-validation across multiple splits and seeds, with an equal hyperparameter budget for every baseline.
• Work with the DMPK hire so the exposure half and the potency half are built as one model, not two projects.
• Represent the science to investors, partners and advisors, including the parts that are not working.

Must have

• 5+ yrs applied deep learning, at least 2 leading a team or a research direction that shipped results.
• Hands-on PyTorch; graph neural networks or molecular representation learning in practice, not only on paper.
• Multi-task, transfer or low-data learning on small assay sets. This is the central modelling problem here.
• Rigorous experimental method: ablations, leakage-free evaluation, honest error bars, and negative controls designed to falsify your own result.
• RDKit and the standard molecular benchmarks at working level.

Experience

A track record of taking a model from idea to a result other people trusted: publications, a shipped product, or a benchmark win. Chemistry or drug-discovery exposure strongly preferred; toxicology a plus.

Nice to have

Conformal prediction or calibration depth; active learning; pharmacokinetics literacy; founding-team or early-startup experience.

Mission

Turn ideas into benchmarked results fast, reproducibly, and without breakage, and keep the infrastructure under them boring.

What you will do

• Implement model architectures and training loops across the encoder and the three heads.
• Build reproducible training and evaluation pipelines; run and track experiments; keep every result re-runnable in one command.
• Own the sweep infrastructure. The evaluation matrix, not the model, is what consumes the GPUs.
• Keep dataset builds versioned and traceable to source, so a number can always be walked back to the record that produced it.

Must have

• Python + PyTorch; solid software engineering (git, tests, clean modular code).
• Reproducible ML + experiment tracking (W&B / MLflow).
• Comfortable with RDKit basics and a GNN library (PyG / DGL).
• Containers, cloud GPU orchestration and cost control. We rent, so waste is visible.

Experience

2+ yrs building real ML systems (not just notebooks). Some exposure to molecules/graphs, or clear evidence you ramp fast on a new domain. Based in or around Warsaw: this seat is on site.

Nice to have

Cheminformatics; distributed / mixed-precision GPU training; data versioning (DVC / LakeFS).

Mission

Build the benchmark. It is the first deliverable of the programme and it survives even if the central hypothesis does not, so it is built to be published and rerun by people hoping we are wrong.

What you will do

• Harmonise the clinical liver labels, the mechanistic assay readouts and the measured exposure tables into one record per molecule, leaving every disagreement visible rather than averaged away.
• Standardise structures: salts stripped, tautomers and charges resolved, stereochemistry kept where the assay kept it, duplicates resolved.
• Audit for leakage across sources, not just within one. The public liver sets descend from a shared pool of regulatory labels and overlap more than their names suggest.
• Pair labels with the exposure that produced them, so the benchmark can ask a dose-conditioned question instead of a structure-only one.
• Design splits that can hurt: scaffold, time, and a harder physicochemical frontier split for chemistry genuinely unlike the training set.
• Document provenance and version every record. Pre-register the set, the splits and the metrics before any modelling starts.

Must have

• RDKit (expert) + Python / pandas.
• Deep understanding of molecular-data pitfalls: activity cliffs, duplicates, assay variability, label noise, and labels that are not reconcilable between sources.
• Familiarity with the public liver-injury and ADMET sets, and with what is wrong with each of them.
• Reproducible, tested dataset builds with lineage on every record.

Experience

3+ yrs cheminformatics / computational-chemistry data work. Has personally built a clean dataset others trusted.

Nice to have

QSAR modeling; medicinal-chemistry intuition; some ML; toxicology familiarity.

Mission

Own the pharmacology. The engine divides predicted exposure by predicted mechanistic potency, and both halves need someone who knows what the numbers mean in a body. This is the gap that left every structure-only competitor stuck at the same score for a decade, and closing it is the company.

What you will do

• Own the exposure half: physiologically based pharmacokinetic prediction of unbound concentration from structure, and the free-fraction correction that decides whether a readout ranks compounds correctly or misranks them.
• Define the mechanisms the potency half is predicted for: mitochondrial and electron-transport failure, bile salt export pump inhibition, reactive metabolite formation, oxidative stress.
• Hold the scope boundary. Dose-dependent injury is in. Idiosyncratic, immune-mediated injury is out, and the labels the field trains on are full of it.
• Design the prospective validation: compound selection, the mechanistic assay panel, blinding, and a statistical analysis plan written and filed before any assay runs.
• Run the contract-lab relationship that produces the measured data, and keep the comparison honest against predictions registered in advance.
• Draft the regulatory context of use and map our evidence to the published acceptance principles for computational safety tools.
• Define metrics that match real safety decisions rather than leaderboard habits, and tell us plainly when a result is a data artifact.

Must have

• Pharmacokinetics in practice: PBPK or physiologically based kinetic modelling, clearance and exposure prediction, plasma protein binding and unbound fraction.
• Toxicology or safety pharmacology: in vitro mechanistic assays, ADMET, and how a dose-dependent injury signal is actually generated.
• Judging label and data quality biologically, including when a published dataset should not be trusted.
• Comfort collaborating closely with an ML team, and the willingness to be the person who says the model is wrong.

Experience

PhD or industry background in DMPK, computational toxicology, pharmacology or safety pharmacology. EU remote is fine. Liver-injury domain depth is a strong plus.

Compensation

$100,000 a year.

Nice to have

Coding (Python / RDKit); regulatory experience (ICH, OECD, FDA drug development tool qualification); contract-lab management; translational or pharma experience.

Mission

Credibility and guidance at near-zero cost: validate the endpoints and the benchmark, pressure-test the regulatory position before a regulator does, open doors, and answer the question investors ask first, which is whether these software people understand biology.

What you will do

• Faculty or clinicians in toxicology, pharmacology and medicinal chemistry.
• Hepatology, for whether a liver-injury claim means what we think it means.
• Regulatory experience with computational tools: context of use, qualification pathways, and the evidence a tool needs for the weight it carries.
• Bonus: anyone with translational or pharma-industry ties.

Experience

A few hours a month. Small advisory equity (0.1–0.5%, 1–2 yr vesting) and/or honorarium. We aim for 2–4 engaged advisors, not a wall of logos.

Mission

Do a real piece of the programme. The evaluation breadth this benchmark needs is more work than five people can run, and an intern who builds one split, one baseline or one ablation properly has contributed to the published result.

What you will do

• Implement and run baselines, ablations and negative controls under the modeling lead.
• Help build and document pieces of the benchmark: curation checks, split construction, provenance records.
• Reproduce published results from the literature, and report honestly when they do not reproduce.

Must have

• Python, and enough statistics to know what a confidence interval claims.
• Chemistry, pharmacy, computer science or a related degree in progress.
• Willingness to work in the open and have your code reviewed.

Experience

Currently studying, in Warsaw or able to be. No industry experience expected.

Nice to have

RDKit; PyTorch; any prior contact with molecular data.