AI Drug Discovery Scientist Job Description

An AI drug discovery scientist builds and runs the computational models that decide which biological targets a drug program goes after and which small molecules get made and tested against them. The job sits between machine learning and medicinal chemistry. On a typical project the scientist owns the model side of the design-make-test-analyze cycle: chemists propose and synthesize compounds, biologists run the assays, and the AI scientist turns each round of results into predictions that shape the next round.

The work splits into two halves. Target discovery comes first in a program's life: mining genetic association data, CRISPR knockout screens, single-cell and bulk transcriptomics, and the published literature to nominate proteins whose modulation should change the course of a disease, then ranking them on evidence strength, safety risk and whether a small molecule can plausibly bind them. Small-molecule discovery follows: virtual screening of very large purchasable or make-on-demand compound libraries, generative models that propose new scaffolds inside a property window, and predictive models for potency, selectivity, solubility, permeability, metabolic stability and toxicity (the ADMET panel).

The toolkit is a blend of classic cheminformatics and modern deep learning. Day to day that means RDKit for fingerprints and substructure work, graph neural networks and chemical language models for property prediction, docking engines, molecular dynamics, and structure prediction and cofolding models that place a ligand inside a protein pocket. The American Chemical Society describes the older half of this craft plainly: computational chemists "use high-performance computing to solve problems and create simulations," with examples including "using simulations to identify sites on protein molecules that are most likely to bind a new drug molecule."

The newer half needs a skeptic, and part of the job is knowing how far to trust it. MIT's Jameel Clinic released Boltz-2 with open weights and code and described it, to the developers' knowledge, as "the first AI model to approach the performance of free-energy perturbation (FEP) methods" for small molecule-protein binding affinity while being "at least 1000× more computationally efficient than FEP." An independent evaluation published in the Journal of Chemical Theory and Computation in August 2026 tested it on two large data sets, 16,780 compounds for one target and 21,702 for another, and concluded that it "lacks the energetic resolution required for lead identification." Holding those two statements together, and designing the benchmark that tells a project team which one applies to their target, is exactly the judgment this role is hired for.

Much of the unglamorous value is in data. Assay results arrive in mixed units, with duplicate measurements, missing stereochemistry and assay-protocol drift between campaigns. The scientist standardizes structures, reconciles units, flags activity cliffs and builds the training sets that every downstream model inherits. Models then have to reach the people who use them, so the role also ships tools: notebooks, internal web apps or design-tool plugins that let a medicinal chemist score a proposed compound in seconds.

The job is distinct from protein and antibody engineering. The molecules here are small, orally dosed chemical compounds made by synthesis, and the hardest questions are about binding affinity, selectivity across related proteins, synthesizability and drug-like properties rather than designing a new protein sequence.

As a rough guide, the job is staffed at the doctoral level. A PhD in computational chemistry, cheminformatics, physical or organic chemistry, chemical biology, computer science or a quantitative field is the usual entry credential, often followed by a postdoc that produced published or open-source work on molecular machine learning. For comparison, the American Chemical Society notes for medicinal chemistry that pharmaceutical companies generally prefer to hire people with research experience, advanced degrees and at least two years of postdoctoral experience. A master's degree plus several years building production models is an alternative route, typically starting at an associate or scientist level.

No license or certification is required. What matters is a portfolio, and it has two halves; candidates should be ready to show both:

• Machine learning depth: PyTorch or JAX, graph neural networks, transformers trained on SMILES or 3D structure, uncertainty estimation, active learning and Bayesian optimization, plus the discipline to split data by scaffold or time rather than at random so validation reflects real prospective use.
• Chemistry and biology judgment: reading a structure-activity table, knowing why a predicted binder is a likely assay artifact, understanding ADMET liabilities, and being able to argue with a medicinal chemist in their own vocabulary. For target discovery, comfort with human genetics, functional genomics screens and pathway biology.

Practical skills round out the profile: RDKit and other cheminformatics libraries, docking and molecular dynamics packages, cloud and GPU cluster computing, SQL against assay databases, Git, and containerized deployment so a model can run inside a chemist's design tool. Public benchmark work, open-source contributions and papers in journals such as the Journal of Chemical Theory and Computation or the Journal of Chemical Information and Modeling give interviewers evidence they can check.

Documentation is part of the craft. When modeling evidence supports a regulatory filing, FDA's final guidance M15 General Principles for Model-Informed Drug Development, posted in June 2026 and prepared under the International Council for Harmonisation, sets out recommendations for "the planning, model evaluation, and documentation of evidence derived from model-informed drug development." Discovery models sit upstream of that stage, but keeping versioned training data, frozen evaluation sets and clear model cards makes the handoff to development teams simpler.

Typical career paths, as a rough guide: postdoc or associate scientist for the first 2 to 3 years, scientist and senior scientist through roughly year 8, then principal scientist, group lead or director of computational or AI-driven discovery. Some people move from medicinal or computational chemistry into the AI side by retraining in deep learning; others arrive from machine learning research and learn chemistry on the job alongside project teams.

BLS does not count AI drug discovery scientists separately, so the federal numbers are context rather than a forecast for this title. The nearest occupation is computer and information research scientists. O*NET rates projected growth for that occupation from 2025 to 2035 as much faster than average (7% or higher), with about 2,900 openings a year on average across all of computing research, not drug discovery alone.

For pay context beyond the band above, the May 2025 BLS median for that occupation across all industries was $140,300, and chemists working in scientific research and development services had a median of $105,930. Chemistry-trained candidates can check a second benchmark: the American Chemical Society's salary survey of its own members, reported by C&EN, found a median salary of $115,000 in 2024 for chemists who responded, from 3,097 members surveyed between April 14 and June 2, 2025.

BioSpace's tally counted about 42,700 employees affected by made or projected biopharma cuts in 2025, a 47.1% increase over the prior year, with the number of companies cutting up 27.6%. The two experts BioSpace interviewed named capital constraints as the main driver, and one described AI as playing "a secondary, more indirect role," one that has led leadership teams to reassess "how much incremental hiring is truly necessary."

Regulators have been building their own capacity in parallel. FDA's drug center reported experience with over 500 submissions with AI components from 2016 to 2023 and set up a CDER AI Council in 2024 to coordinate its AI work. FDA published a draft guidance in 2025, Considerations for the Use of Artificial Intelligence to Support Regulatory Decision Making for Drug and Biological Products, aimed at AI used to produce information on safety, effectiveness or quality. That guidance concerns information meant to support regulatory decisions, which sits downstream of early discovery work.

Employers fall into a few groups: large pharmaceutical companies with in-house AI and computational chemistry departments, venture-backed AI-first biotechs that pair a model platform with a pipeline of their own programs, contract research organizations that sell computational screening as a service, software vendors that build discovery platforms, and university or national laboratory groups.

The title itself is not standardized, so a job search should cast a wide net: machine learning scientist, cheminformatics scientist, computational chemist and computational biologist can all describe the same work. Federal classification splits the same way; O*NET's title file places machine learning scientist under computer and information research scientists and pharmaceutical chemist under chemists. Read the duties, not the title, and look for the words target, ADMET, virtual screening and design-make-test-analyze.

For job seekers, the most useful evidence to build is prospective: compounds a model nominated that later held up in the assay, alongside retrospective benchmark scores.

Dear Hiring Committee,

I am applying for the AI Drug Discovery Scientist position on your small-molecule platform team. For the past three years I have worked as a computational chemist at a mid-sized oncology biotech, where I built the property-prediction models our medicinal chemists use in every design meeting.

My most useful work there started with data rather than architecture. Our kinase assay history was spread across four contract labs with different units and inconsistent stereochemistry, so I rebuilt the training set from raw plate data, standardized every structure and split the validation set by scaffold and by date. The cleaned models were less impressive on paper and far more honest in practice: when we tested them prospectively on the next two design cycles, the chemists trusted the rankings enough to cut the number of compounds they sent for synthesis.

I also led a virtual screening campaign against a newly nominated target that our biology team identified from a CRISPR screen. I combined a cofolding model for pose generation with a physics-based rescoring step, and I set up the benchmark first, so the team knew in advance how we would judge success. Several of the purchased hits confirmed in the primary assay and one series is now in lead optimization.

My PhD in physical chemistry focused on graph neural networks for solvation energy, and I maintain a small open-source package for scaffold-aware data splitting. I write Python and PyTorch daily, work comfortably with RDKit and molecular dynamics tools, and enjoy the part of the job where I sit with a chemist and explain why the model disagrees with their intuition.

I would welcome the chance to bring that mix of modeling rigor and project focus to your team.

Sincerely, Jordan Ellison