Senior/Staff Applied Scientist, Multimodal Representation Learning (Oncology)

Pathos•New York, NY

5d•Hybrid

About The Position

About the role: Where Frontier AI Meets Frontier Biology to Deliver Frontier Medicine We are hiring specialized scientists to accelerate development of our Oncology Foundation Model (OFM) stack. This is not a generic “model tinkering” role. The person in this seat will help define and build the modeling strategy that turns multimodal oncology data (clinical text/EHR, genomics, transcriptomics, pathology imaging, and derived features) into useful representations and predictive capabilities that directly support drug discovery and development. You’ll operate at the intersection of: Frontier AI (representation learning, multimodal learning, alignment, evaluation) Messy biomedical reality (clinical endpoints, censoring, confounding, missingness, batch effects) Mechanism + translation (models that can be interrogated, stress-tested, and connected to biology and outcomes) This role complements (not duplicates) the computational biology roles that focus on our program-facing biomarker analyses and trial decisions.

Requirements

Advanced degree (PhD strongly preferred) in ML/AI, CS, Statistics, Computational Biology, Bioinformatics, or a related field, or equivalent industry experience with a strong publication/impact record.
Deep hands-on experience with modern deep learning (PyTorch), including training large models and debugging optimization issues.
Demonstrated ability to design representation learning / foundation model approaches and evaluate them rigorously (not just “train and report AUCs”).
Comfort operating in ambiguous problem spaces with a bias toward execution and iteration.

Nice To Haves

Multimodal foundation model experience (any of: clinical + omics, imaging + text, multimodal retrieval, alignment, late fusion/mixture-of-experts).
Real experience with at least one of the following domains (enough to reason about the data-generating process and pitfalls): Clinical text / EHR (notes, longitudinal events, coding systems, leakage traps) Molecular/omics modeling (RNA/DNA/variant features, batch effects, multi-cohort generalization) Pathology imaging (WSI feature learning, weak supervision, MIL, slide-level endpoints)
Distributed training and systems experience (FSDP/DeepSpeed, multi-node performance profiling)
Experience with alignment methods (preference learning, instruction tuning, evaluation frameworks for reliability/robustness).
Publications in relevant venues (NeurIPS/ICML/ICLR/ACL/MLHC) and/or impactful open-source work.

Responsibilities

Foundation model development
Design and implement multimodal pretraining and fine-tuning strategies for oncology data (e.g., contrastive objectives, masked modeling, multitask learning, retrieval-augmented training, late/early fusion variants).
Build model components that improve cross-modality grounding (e.g., aligning clinical narratives with molecular state and pathology signals).
Develop robust approaches for missing-modality settings (train-time and inference-time), ensuring the OFM remains useful when only subsets of modalities exist.
Clinical + molecular fluency
Work with domain partners to define prediction targets and representation tests that matter: response, durability, toxicity, survival, progression, resistance, subtype stability, etc.
Incorporate oncology-specific realities into modeling and evaluation (censoring, treatment lines, temporal leakage, cohort shift, annotation noise).
Evaluation, benchmarking, and scientific rigor
Create evaluation harnesses that go beyond leaderboard metrics: ablations, cohort-shift tests, missingness stress tests, temporal generalization, calibration, and failure-mode analysis.
Define and maintain benchmark suites that reflect Pathos priorities and are reproducible across model iterations.
Partner with engineering to support scalable training/inference (multi-node GPU training, data pipelines, throughput optimization), while keeping scientific intent front-and-center
Translation enablement
Package model outputs so they can be consumed by internal science teams: embeddings, uncertainty estimates, interpretable signals, retrieval tools, and model cards that clearly state what’s reliable vs. not.
Collaborate with computational biologists, translational scientists, and clinicians to ensure the OFM supports mechanism discovery and patient stratification workflows