On the ML team at Manifold Bio, you'll lead and advance foundation model training efforts—pretraining, fine-tuning, and evaluating folding, docking, language, and generative design models on proprietary experimental data—while contributing deep expertise in training methodology, architecture selection, and optimization. You will develop and scale distributed multi-GPU/multi-node training pipelines, integrate model outputs into mBER to improve binder design and enable new capabilities, and design rigorous ML experiments with clear hypotheses, evaluation frameworks, and systematic analyses. You’ll also establish best practices for efficient large-scale training (e.g., mixed precision, gradient checkpointing) and produce clear documentation and analyses to support architecture and training decisions.
Responsibilities
Advance the team's ongoing foundation model training efforts—pretraining, fine-tuning, and evaluating folding, docking, language, and generative design models on Manifold's proprietary experimental data
Bring depth in training methodology, architecture selection, and optimization to complement the existing team's expertise
Develop and scale training pipelines for distributed, multi-GPU and multi-node training runs
Integrate foundation model outputs into mBER to improve binder design success rates and enable new design capabilities
Design and execute ML experiments with clear hypotheses, rigorous evaluation frameworks, and systematic analysis
Establish best practices for mixed-precision training, gradient checkpointing, and computational efficiency at scale
Produce clear documentation and analysis supporting architecture and training decisions
Not Met Priorities
What still needs stronger evidence
Requirements
Demonstrated experience pretraining and/or fine-tuning protein foundation models (folding, docking, language models, or generative design) with published or otherwise demonstrable results
Strong familiarity with AlphaFold architecture and training methodology
2+ years of hands-on experience with PyTorch and/or JAX for deep learning
Experience with large-scale model training: distributed training, multi-GPU/multi-node setups, mixed precision, gradient checkpointing
Solid understanding of deep learning architectures (transformers, attention mechanisms, diffusion/flow matching) and optimization techniques
Experience working with protein structure data (PDB, mmCIF) and/or protein sequence datasets
Strong statistical analysis and experimental design skills
Proficiency in Python scientific computing stack (NumPy, Pandas, scikit-learn)
Self-directed researcher who can balance guidance with independence
Excellent written and verbal communication skills for cross-functional collaboration
Preferred Skills
Excellent written and verbal communication skills for cross-functional collaboration
Experience with protein generative design methods (e.g., RFdiffusion, ProteinMPNN, flow matching approaches)
Experience with protein language models (e.g., ESM family)
Published research in computational biology, protein design, or structural biology
Experience training on proprietary or domain-specific biological datasets
Familiarity with Ray for distributed computing
Experience with Kubernetes (EKS) and cloud computing platforms (AWS)
Knowledge of protein engineering, directed evolution, or structural biology wet lab techniques
Experience working with agentic AI coding tools for fast, parallelized execution of modeling experiments
Previous biotech/pharma industry experience
You have deep experience training protein foundation models and want to apply that expertise to some of the richest proprietary experimental datasets in the field
Position Manifold's AI team is actively training protein foundation models on our proprietary experimental datasets. Our generative antibody design model, mBER, has already demonstrated controllable de novo binder design across multiple million-scale screening campaigns, and the team is now scaling foundation model capabilities to push well beyond current performance. We are looking for an AI/ML Scientist to join this effort. You will work alongside our existing model training team to accelerate the development of foundation models fine-tuned on Manifold's data, bringing additional depth in pre-training methodology, architecture development, and large-scale training. Your work will directly improve mBER's design capabilities and unlock new modeling paradigms for the broader team. You'll own foundation model projects end-to-end, from architecture selection and training infrastructure to evaluation against real experimental outcomes, while contributing to the team's shared research agenda. Responsibilities
Advance the team's ongoing foundation model training efforts—pretraining, fine-tuning, and evaluating folding, docking, language, and generative design models on Manifold's proprietary experimental data Bring depth in training methodology, architecture selection, and optimization to complement the existing team's expertise Develop and scale training pipelines for distributed, multi-GPU and multi-node training runs Integrate foundation model outputs into mBER to improve binder design success rates and enable new design capabilities Design and execute ML experiments with clear hypotheses, rigorous evaluation frameworks, and systematic analysis Establish best practices for mixed-precision training, gradient checkpointing, and computational efficiency at scale Produce clear documentation and analysis supporting architecture and training decisions Required Qualifications
Demonstrated experience pretraining and/or fine-tuning protein foundation models (folding, docking, language models, or generative design) with published or otherwise demonstrable results Strong familiarity with AlphaFold architecture and training methodology 2+ years of hands-on experience with PyTorch and/or JAX for deep learning Experience with large-scale model training: distributed training, multi-GPU/multi-node setups, mixed precision, gradient checkpointing Solid understanding of deep learning architectures (transformers, attention mechanisms, diffusion/flow matching) and optimization techniques Experience working with protein structure data (PDB, mmCIF) and/or protein sequence datasets Strong statistical analysis and experimental design skills Proficiency in Python scientific computing stack (NumPy, Pandas, scikit-learn) Self-directed researcher who can balance guidance with independence Excellent written and verbal communication skills for cross-functional collaboration Preferred Qualifications
Experience with protein generative design methods (e.g., RFdiffusion, ProteinMPNN, flow matching approaches) Experience with protein language models (e.g., ESM family) Published research in computational biology, protein design, or structural biology Experience training on proprietary or domain-specific biological datasets Familiarity with Ray for distributed computing Experience with Kubernetes (EKS) and cloud computing platforms (AWS) Knowledge of protein engineering, directed evolution, or structural biology wet lab techniques Experience working with agentic AI coding tools for fast, parallelized execution of modeling experiments Previous biotech/pharma industry experience This Role Might Be Perfect For You If
You have deep experience training protein foundation models and want to apply that expertise to some of the richest proprietary experimental datasets in the field You're excited about pushing beyond public model performance by leveraging unique, large-scale in vivo screening data You thrive in high-ownership roles where you can drive research direction while collaborating with a tight-knit, world-class team You want your models to directly impact real drug discovery programs If you're excited to train the next generation of protein foundation models on uniquely powerful experimental data, please reach out to careers@manifold.bio. We value different experiences and ways of thinking and believe the most talented teams are built by bringing together people of diverse cultures, genders, and backgrounds.
Manifold Bio
•
Boston, MA