Research planner for biology

The need for abstraction

What truly hinders progress in biology research? Oh, many things, but here’s one that I think deserves far more attention – protocols. Specifically, I’m referring to the challenge of adapting and building upon published work. In computer science, if you stumble upon an intriguing AI paper, you need not labor through its methodology to use it – simply clone their GitHub repository and you’ll be up and running within minutes.

Minutes! In biology, if I come across an interesting method and want to try it out, I’ll be up and running within… weeks, if not months.

This challenge extends far beyond the impoverished methods sections in papers. Published protocols merely provide static snapshots of procedures. For researchers interested in building on existing work, the true obstacle lies in the adaptation of the protocol. One must master the unwritten grammar of protocols – knowing which reagents can be substituted, which modifications are essential for your particular plasmid/protein/strain, and how to troubleshoot inevitable failures.

This is hard! Researchers spend countless hours troubleshooting failed experiments, fostering profound aversion to modifying established protocols – even when those protocols demand absurd rituals like tying knots on willow branches under full moons while chanting gene names.

In essence, we need generative models (assistants, agents) for biology protocols – systems that translate researcher’s requests into adaptable, step-by-step procedures that evolve as implementation unfolds. My central thesis: the primary bottleneck in contemporary biology research is the absence of abstraction layers above the technical details of how to develop and run assays. You don’t need to understand internal combustion to reach destinations. You don’t need to comprehend transistors to create software. And you don’t need to know how to cast gels to make groundbreaking discoveries.

Common ideas to 10X biology and their limitations

Idea 1: Generate more groundbreaking research questions. Selecting worthwhile questions is hard, so resources seem to be wasted on modest inquiries that consume as much time as great ones while yielding disproportionately smaller impacts.

But biological systems are complex and their application possibilities are vast. We are still in biology’s exploratory phase where most questions merit investigation as we do not know where that will lead us. We must accumulate substantially more data while entertaining multiple hypotheses before life’s fundamental patterns emerge clearly. Until then, no hypothesis proves too modest, no discovery too incremental. Rigorous lab testing of these hypotheses constitutes the true bottleneck limiting progress.

Idea 2: Expand grad student pool to test more ideas concurrently. While this approach benefits group leaders, individual students gain no acceleration. On the contrary, more students typically means less mentorship, fewer development opportunities, and consequently struggle throughout their doctoral journey. Many arrive with genuine curiosity and dedication, and we do them profound disservice by permitting endless cycles of failure. Grad students deserve better than serving as cannon fodder. Provide only shovels, and they’ll languish in stagnant trenches; equip them with fighter jets, and they’ll soar through research challenges. In biology, the right arms are the tools to plan experiments and discuss results.

Idea 3: Automate wet lab procedures to bypass human limitations. Established, rigid pipelines certainly benefit from connecting purpose-built machines. Yet research inherently demands flexibility – constant vigilance for anomalies, perpetual troubleshooting, and continuous adaptation. Ideally, we would systematically test numerous combinations (where automation excels), but prohibitive reagent costs make blind automation infeasible.

Furthermore, lab infrastructure remains hostile to automation. Instruments lack APIs and communication protocols, equipment is scattered across facilities, reagents are dispersed throughout refrigerators and storage containers, and techniques demand dexterity beyond current robotic capabilities.

The emerging concept of bimanual robots navigating lab spaces (teleoperated or autonomous) shows promise – perhaps within 5-10 years. Numerous routine tasks requiring minimal skill could liberate researcher’s attention: removing overnight cultures from incubators, performing centrifugation steps, or monitoring optical density during cultivation. However, grad student labor currently remains more economical and vastly more versatile than robotic alternatives, even if students need to eat, sleep, and constantly lack motivation.

No, experimental bottlenecks stem primarily from inadequate planning and troubleshooting capabilities (remember, grad students are still junior researchers) rather than manual execution. While liquid handling robots might theoretically offer superior speed and precision, what good does this speed do if the experiment itself contains critical flaws or employs suboptimal protocols?

Idea 4: Develop superior instrumentation, the ultimate form of automation. Remember when researchers had to manually transfer samples between water baths to conduct PCR? Now they simply program thermocyclers. Revolutionary devices would undoubtedly accelerate biology! Yet modern laboratories remain shocking enclaves of near-alchemical traditions – word-of-mouth gotchas (“she was struggling for three months with this plasmid before it turned out it was prepared incorrectly”), peculiar techniques that belong to the kids experiment’s section (slimy agarose gels, anyone?), and researchers scribbling cryptic notes on paper notebooks while laboriously calculating concentrations – all to conduct experiments yielding the most amazing, god-like advances human health. What a sight!

Wet lab needs a revamp. Molecular cloning represents an outdated paradigm desperately needing replacement with affordable on-demand DNA synthesis at any scale. Protein expression and purification similarly demands streamlined services. Sequencing costs must plummet to finally leave agarose and ethidium bromide on the fingertips of the giants who kickstarted molecular biology. Experiment scales should shrink below microliter volumes, as picoliter reactions contain ample molecules for most protocols. And this list merely scratches the surface of necessary instrumentation advances.

The principal challenge with instrument development: substantial time and capital requirements. Even when innovations reach market (e.g., commercial e-gels), prohibitive costs prevent routine adoption, and technological democratization proceeds glacially. By all means, I want people to continue innovating new devices, but this is no immediate remedy I am looking for. Rather, I believe that substantial speedups can be achieved more rapidly and economically by making experiments more predictable and accessible through AI agents. This is, if you wish, a poor man’s hack to lab automation and new devices.

Idea 5: Construct virtual cells/organisms. Simulating biological processes would enable in silico outcome prediction, allowing strategic selection of only the most promising experiments for empirical validation.

Realistically, we remain distant from models sufficiently reliable to replace physical experiments. Some valuable simulation capabilities appear imminent (particularly protein dynamics predictions and accelerated molecular dynamics simulations). Other objectives, like whole-cell simulations, remain distant prospects due to massive data requirements. Acquiring this data presents formidable challenges – it is an expensive and thankless job. The Protein Data Bank architects don’t receive Nobel Prizes; those leveraging such resources do. As Stephen Malina observes, despite CryoEM accessibility, countless biomolecular structures await determination – work offering immense collective value, yet coordination remains daunting. Worse still, the most valuable biological data often faces regulatory constraints (clinical trials) or outright prohibition (human embryo developmental).

Moreover, even possessing virtual cellular models wouldn’t guarantee straightforward insights. AI research paradoxically resembles wet laboratories – complex systems misbehave inexplicably, prompting parameter adjustments and processing modifications, followed by days awaiting results while depleting AWS budgets hoping intuitions prove correct. But unlike biology, AI at least benefits from robust abstraction layers eliminating reproducibility concerns of a particular routine that used to work yesterday.

Our current reality: insufficient rapid, affordable data generation pipelines impede biological understanding, pushing truly useful virtual experimentation years into the future. While AI increasingly informs biomolecular selection for empirical testing, the fundamental bottleneck remains conducting actual experiments.

The anatomy of a research planner

The research planner essentially functions as an omnipresent Principal Investigator: approach them with your challenge, and together you’ll chart the optimal research path. For instance, this planner:

But unlike your PI, this planner remains perpetually available, continuously updated, and exceptionally insightful. Realizing such a system would genuinely democratize scientific inquiry – a vision particularly dear to me who has consistently challenged scientific institutional orthodoxy (as evidenced throughout my blog).

Fundamentally, this planner is just a contemporary large language model, accepting natural language research instructions alongside supporting materials – datasets, visualizations, publications – to produce a step-by-step reasoning and an action plan of what experiments need to be carried out in order to address the research question. This interactive process allows researchers to implement recommendations while the planner continuously refines subsequent steps based on emerging results.

Crucially, the planner must embody meta-scientific reasoning patterns. I conceptualize scientific research as a decision tree with select branches leading to significant discoveries. Researchers must first map this conceptual territory, then systematically explore promising paths. Some routes terminate decisively, requiring backtracking; others branch endlessly, demanding judicious termination decisions; still others require creative connections between seemingly disparate concepts, occasionally necessitating fundamental research question reformulation. Throughout this process, the planner must provide precise factual information from contemporary literature, explicitly state assumptions and confidence levels, thoroughly explore methodological alternatives, and employ various reasoning frameworks including counterarguments (actor-critic), hypothesis refutation, counterfactual thinking (what if?), and additional techniques xAI researchers will undoubtedly develop.

While such sophisticated systems remain aspirational, these capabilities represent foundational requirements for successful language models broadly – not exclusively research-oriented ones. I anticipate remarkable progress in these domains as LLMs evolve. Encouragingly, biology research represents a relatively bounded universe; there is a finite amount of approaches for molecular cloning, DNA quantification, or plasmid design. Humans require years developing comprehensive biological understanding and creatively operate with the available building blocks, yet language models could conceivably assimilate all established techniques and accumulated evidence, then innovate by recombining methodological building blocks into novel protocols.

As data accumulates and models refine their reasoning capabilities, the subsequent frontier – proposing transformative research directions – becomes attainable. It has happened in Go, it starts happening in protein design, and will happen for the whole field of biology. How long will human researchers remain essential once we develop such AI scientists?

Just like there was a shift from free AI research to AI engineering, similar transformations await all research domains. In an era of abundant intelligence, knowledge itself ceases to be an exclusive asset. Our collective focus must shift toward determining what wisdom we, as a society, wish to harvest from the expanding tree of knowledge.