iiinova — where your digital self keeps evolving

Kasper Krunderup Jakobsen @kkj15.bsky.social · 07/10/2026

We're excited about how information-dense synthesis points to new ways of scaling lab-in-the-loop ML systems, not just by increasing the number of experiments, but by increasing the amount of information each measurement carries.

100

Kasper Krunderup Jakobsen @kkj15.bsky.social · 07/10/2026

Information-dense synthesis has not yet been validated in the wetlab, but the synthesis robots are already running. It is also yet to be scaled to bigger state-spaces, such as full-length proteins. Doing so will require further accelerating Bayesian experimental design.

100

Kasper Krunderup Jakobsen @kkj15.bsky.social · 07/10/2026

We further validated LIDS on protein language model (PLM) oracles in silico. It is very fast. Here, LIDS can find the optimum in 2 experimental rounds (median), while individual synthesis alternatives need an order of magnitude more experiments.

122

Kasper Krunderup Jakobsen @kkj15.bsky.social · 07/10/2026

Now the posterior concentrates further, and converges on the true optimum of the unknown molecule-activity map f.

100

Kasper Krunderup Jakobsen @kkj15.bsky.social · 07/10/2026

LIDS then designs new mixtures for the second round, which cover its remaining uncertainty about the optimum.

100

Kasper Krunderup Jakobsen @kkj15.bsky.social · 07/10/2026

This is used to update our beliefs about which DNA molecule has the highest activity.

100

Kasper Krunderup Jakobsen @kkj15.bsky.social · 07/10/2026

For each mixture, we observe a noisy measurement of activity.

100

Kasper Krunderup Jakobsen @kkj15.bsky.social · 07/10/2026

LIDS designs complex mixtures which cover this uncertainty about the optimum.

100

Kasper Krunderup Jakobsen @kkj15.bsky.social · 07/10/2026

On a synthetic DNA sequence-activity oracle, we simulated LIDS for sequential rounds of experimentation. After 2 rounds (median), LIDS converges on the true optimum, while individual synthesis methods require >100. We visualized LIDS using low-dimensional projections.

100

Kasper Krunderup Jakobsen @kkj15.bsky.social · 07/10/2026

Combinatorial synthesis makes it just as easy to synthesize a mixture as a pure compound. You can keep the number of reaction steps and the amount of reactants the same. The only difference is that the reactants are mixed in different ratios before and after each synthesis step.

100

Kasper Krunderup Jakobsen @kkj15.bsky.social · 07/10/2026

We optimize the stochastic synthesis, so that it carries as much information as possible, enabling accurate reconstruction of the optimum. We call this algorithm learning with information-dense synthesis (LIDS).

100

Kasper Krunderup Jakobsen @kkj15.bsky.social · 07/10/2026

To achieve this, we model the experiment as an encoder, which produces a low-dimensional representation of the unknown molecule-activity map.

110

Kasper Krunderup Jakobsen @kkj15.bsky.social · 07/10/2026

The conventional approach is individual synthesis, where all those molecules are identical. But you can also do combinatorial synthesis, such that they are different. It turns out you can gain exponentially more information by testing many different molecules at once, as a pool.

100

Kasper Krunderup Jakobsen @kkj15.bsky.social · 07/10/2026

Crucially, in molecular science, you never test just a single molecule. You might synthesize a nanomole of molecules: that is 6 × 10¹⁴ inputs to your experiment!

100

Kasper Krunderup Jakobsen @kkj15.bsky.social · 07/10/2026

How do lab-in-the-loop ML systems learn most efficiently? How can you pack as much information as possible into an experiment? We explored this in our new preprint “Information-Dense Synthesis for Molecular Discovery” (arxiv.org/abs/2610.08495).

arxiv.org

Information-Dense Synthesis for Molecular Discovery

Machine learning can accelerate molecular discovery by designing molecules and planning experiments. However, many scientific challenges demand molecules with very rare properties, and in this sparse ...

171